06-07 December, 2012, Danang, Vietnam

i

Table of Contents

The 2012 International Conference on BioSciences and BioElectronics


Table of contents _____________________________________________________________ i
Message from the Chairs ______________________________________________________ iii
Technical Program Committee _________________________________________________ v

Keynote Abstracts
Automated CCTV Surveillance Research at the University of Reading ________________ vi
Bioprospecting microbes to innovate new biotechnological uses for control of plant pests
and diseases _________________________________________________________________ vii


SESSION A: BIOELECTRONICS
A Survey on Advanced Video-Based Healthcare Monitoring Systems ____________________ 1
Le Thi My Hanh, Hoang Le Uyen Thuc, Tuan V. Pham.
Pulse Oximetry System based on FPGA ___________________________________________ 9
Kien Nguyen Trung, Ngoc Nguyen Van.
Towards Efficient Implementation of Neural Networks with Reduced Precision Floating Point
Parameters ___________________________________________________________________ 14
Huynh Viet Thang, Nguyen Hai Trieu Anh.
Human-Computer Interaction using ultrasound hand tracking via USB interface ____________ 19
Sau V. Tran, Tuan K. Tran, Tu T. Hoang, Trung X. Pham.
A Real-Time Scheduling Scheme for Distributed Control System ________________________ 24
Trong Cac Nguyen, Xuan Hung Nguyen, Van Khang Nguyen.
Fall Detection Based On Hidden Markov Model _____________________________________ 32
Viet Q. Truong, Hieu V. Nguyen, Tuan V. Pham.
Measurement of Biological Concentration Using Magnetic Agents_______________________ 37
Cao Xuan Huu, Tuan V. Pham, Dang Duc Long, Nguyen Thi Minh Xuan, Pham Thi
Kim Thao.
A Novel Approach to Protect Intellectual Property Core of FPGA-Based Partially
Reconfigurable Systems ________________________________________________________ 42
Tran Thanh, Tran Hoang Vu, Pham Ngoc Nam, Nguyen Van Cuong.
ii

A DAG - SVM based Method for Face Recognition using PCA. ________________________ 46
Tran Thi Minh Hanh.
A Speaker Recognition System using Combination Method between Vector Quantization and
Gaussian Mixture Model _______________________________________________________ 51
Ngo Quoc Hung, Tuan V. Pham.
Design of Television Remote Controller based on Vietnamese Speech Recognition _________ 55
Nguyen Tu Ha, Tuan V. Pham.
ECG Signal Transmission Using Wireless Technology in Patient Health-care and Monitoring
System ______________________________________________________________________ 58
Duong Trong Luong, Nguyen Duc Thuan, Nguyen Hung.
Combination of analog and digital solutions for Wireless ECG Monitor __________________ 63
Khue Tra, Phung T.Kim Lai, Khiem Duy Nguyen, Tuan V. Pham.


SESSION B: BIOMETRICS
Biodegradation of Phenol by Native Bacteria Isolated from Dioxin Contaminated Soils ______ 68
Bui Ba Han, Nguyen Thi Lan, Dang Duc Long.
Biopolymer Film from Chitosan for Shelf-life Extension of Fruits _______________________ 74
NGUYEN, Thi Xuan Lam; NGUYEN, Thi Minh Xuan; DANG, Duc Long
Effect of CO
2
Utilization on the Growth of Chlorella Vulgaris for Food Technology _________ 79
Nguyen Hoang Minh, Nguyen Thi Thanh Xuan, Dang Kim Hoang, Nguyen Dinh Phu
Effect of Carbon Sources on Proliferation of Zedoary (Curcuma Zedoaria Roscoe) Cell
Suspension Cultures ___________________________________________________________ 83
Vo Chau Tuan, Tran Quang Dan
Use of BMWP
VIET
and ASPT Indices as Bioindicators for Testing Water Quality of Rivers in
Danang city __________________________________________________________________ 88
Nguyen Van Khanh, Vo Van Minh, Kieu Thi Kinh, Tran Duy Vinh, Phan Thi Hien
Whole Cell Immobilisation of Bacillus Subtilis on Cellulose Carriers and Waste Water
Treatment Application __________________________________________________________ 93
TRAN, Thi Xo, NGUYEN, Thi Diem Quynh.
Lactic Acid Fermentation from Jackfruit Seed Flour __________________________________ 96
Trng Thi Minh Hanh, Ho Thi Hao.
Gold Nanoparticles based Localized Surface Plasmon Resonance in Combination with
MicroFluidic System for Biomolecule Determination. _________________________________ 100
Nguyen Ba Trung, Le Tu Hai, Yuzuru Takamura.

iii

Preface
Dear readers and colleagues,

It is a great honour to introduce the proceedings of the BioSciences and
BioElectronics Conference (ICBSBE2012). At the conference, many scholars and experts
from industrial organizations and government agencies have participated in various
sessions to share research ideas, potential collaboration opportunities and state-of-the-art
achievements in the field of emerging BioSciences and BioElectronics and BioMetrics
technologies throughout the world. In the coming decades, these technologies will certainly
play a decisive role in the sustainable development of Vietnam and many other countries.
We acknowledge that BioSciences and BioElectronics are critical to sustainable
development and poverty reduction efforts. It affects all aspects of development such as
social, economic, and environment including livelihoods, access to water, agricultural
productivity, health, population levels, education, and gender-related issues. None of the
Millennium Development Goals can be met without major improvement in the quality and
quantity of such BioSciences and BioElectronics services in developing countries.
Therefore, this Conference is an excellent opportunity for researchers, industry
professionals, academic and administrative organizations to exchange the latest innovations
and developments in technologies between United Kingdom and Vietnam. Accordingly,
conference outcomes will contribute to the environmental improvement as what we desire
so that Vietnam can develop sustainably.
The 2012 ICBSBE is co-organized by Danang University of Technology The
University of Danang, British Council Vietnam, and the University of Reading. The goal of
the conference is two-fold: First, to exchange the latest innovations and developments in
the fields of BioSciences, BioElectronics, BioMetrics and Applications in related
technologies; Secondly, to gather academics, researchers, industry professionals, and higher
educational-related organizations from Vietnam and foreign countries, especially from
United Kingdom. ICBSBE2012 will include keynote speeches, paper presentations,
tutorials, panel discussions and field trips.
In recent years, we have set up some Teaching Research teams (TRTs) to conduct a
series of research projects in the field of technology and engineering with our international
partners. We have also focused on the collaborative programs such as student and staff
exchange programs, advanced programs, co-organisation of international symposiums,
especially joint research and co-publication with university partners.
The staff and students at Danang University of Technology - The University of
Danang are very pleased to work as the host university for the two thousand and twelve

iv

(2012) BioSciences and BioElectronics Conference ICBSBE2012. This is a meaningful
event for the University of Danang and Danang City, a potential and dynamic city in socio-
economic development as well as in environmental protection. I am proud to say that
Danang will become a smarter city to improve living and working environtments. It is
evident that our staff and students and people in Danang will be of the beneficiaries from
the outcomes of this important conference.

Thank you again for choosing Danang University of Technology, The University of
Danang as the location for the conference.
Thank you for coming and attending the conference and I hope you enjoy the time in
Danang. I wish all of you good health and the conference the best success.

Conference Chair
Prof. Dr. Le Kim Hung
v

Technical Program Committee
Honorary chairs:
Tran Van Nam (President, The University of Danang)
Le Kim Hung (Rector, Danang University of Technology)
Steven Mithen (Pro-Vice-Chancellor, University of Reading)
General chairs:
Vo Trung Hung (Director of Science Technology, The University of Danang)
Hoang Hai (Vice-Director of International Collaboration, The University of Danang)
Robin Rickard (Director, British Council Vietnam)
Vu Kieu Dung (Assistant Director, British Council Vietnam)
Truong Hoai Chinh (Vice-Rector, Danang University of Technology)
Le Thi Kim Oanh (Vice-Rector, Danang University of Technology)
Nguyen Dinh Lam (R&D, Post-graduate & Int. Collaboration, Danang University of
Technology)
Technical program
David Hogg (University of Leeds)
Jenq-Neng Hwang (University of Washington)
James Ferryman (University of Reading)
Rob Jackson (University of Reading)
Tang Tan Chien (The University of Danang)
Nguyen Van Tuan (Danang University of Technology)
Vo Van Minh (Danang University of Education)
Dang Duc Long (Danang University of Technology)
Huynh Huu Hung (Danang University of Technology)
Pham Van Tuan (Danang University of Technology)
Publicity
Duong Mong Ha (University of Danang)
Nguyen Thu Giang (British Council Vietnam)
Bui Thanh Nga (British Council Vietnam)
Finance
Lam Thi Hong Tam (Danang University of Technology)
Vu Ngoc Ha (Danang University of Technology)
Publication
Ngo Thai Bich Van (Danang University of Technology)
Huynh Viet Thang (Danang University of Technology)
Huynh Tan Tien (Danang University of Technology)
Local arrange
Le Minh Duc (Danang University of Technology)
Nguyen Le Hung (Danang University of Technology)
Tran Thi Huong (Danang University of Technology)
Nguyen Thi Minh Xuan (Danang University of Technology)
Contact:
Pham Van Tuan, Email: pvtuan@dut.udn.vn

vi

Keynote Talk 1
Automated CCTV Surveillance Research at the University of Reading
Abstract:
The Computational Vision Group (CVG), within the School of Systems Engineering at the
University of Reading, are very active in security research, with a focus on automated
surveillance including CCTV analysis for safety, security, and threat assessment. This talk will
present an overview of security research activities with particular attention paid to recent and
ongoing FP7 security projects SUBITO (detection of unattended baggage and the identification
and tracking of the owner), Co-Friend (cognitive understanding of airport aprons), and EFFISEC
(more efficient integrated security checkpoints), in which CVG play a significant role. The talk will
also address the overarching themes of performance evaluation of detection systems, in particular
details of the PETS (Performance Evaluation of Tracking and Surveillance) workshop series
which has most recently addressed multi camera crowd image analysis.

Speakers Biography:
Dr. Ferryman leads both the Computational Vision Group (CVG) and the wider Computing
Research Group in the School of Systems Engineering, University of Reading. Dr. Ferryman's
personal research interests include automated CCTV surveillance of public spaces (including
improving the effectiveness of CCTV), behavioural analysis, cognitive systems, multimodal
interaction, robotics and autonomous systems, novel imaging modalities, and performance
evaluation. Dr.Ferryman has received extensive funding from Research Councils UK, the EU, and
Industry. Dr. Ferryman was Co-investigator on the EPSRC network ViTAB (Video based Threat
Assessment and Biometrics) and was Principal Investigator on the EPSRC REASON project
(EP/C533402) examining robust methods for monitoring and understanding people in public
spaces. Dr. Ferryman has been Principal Investigator on a number of EU projects including the
PASR project ISCAPS (013800) on integrated surveillance of crowded areas for public security, the
EU Framework 6 Aeronautics project SAFEE, which addressed aircraft onboard threat detection,
and Framework 7 Security/Cognitive Systems projects Co-Friend (airport apron monitoring) and
EFFISEC (efficient integrated security checkpoints.) Dr. Ferryman is a member of the British
Machine Vision Association (BMVA) and the Security Information Technology Consortium
(SITC).

vii

Keynote Talk 2
Bioprospecting microbes to innovate new biotechnological uses for
control of plant pests and diseases
Abstract:
With an increasing number of restrictions on the use of chemicals to control pathogens and
pests of plants, there is an urgent need to identify alternative means of control. One approach that is
receiving a resurgence of interest is the use of biological agents, which may include whole
organisms, viruses, or components of them. In our laboratory, we employ bioprospecting to identify
novel biological agents that may be utilised for pest and disease management. Our main targets are
bacteria and bacteriophages, with a view to controlling fungal and oomycete pathogens, insect pests
such as aphids and thrips, and bacterial pathogens. We also employ genetic screens to identify the
systems employed by bacteria to survive in different niches and under different climactic conditions.
As well as providing key insights to bacterial function, it can also help to reveal novel bioactive
molecules such as toxins and surfactants. Taken together, these approaches can innovate the
discovery of new biological controls for plant pests and diseases.

Speakers Biography:
Robert W. Jackson is a Lecturer in Microbiology within the School of Biological Sciences,
University of Reading. He is also Admissions tutor for the BSc (Hons) Microbiology degree and
Director of Enterprise for the School.
Having always had an interest in science from a young age, his appetite for research started
during a 6 month Erasmus placement at the University of the Algarve in Portugal working with
Prof. J ose Leitao after a one month period of dossing on the beach and drinking too much wine
from the local garrafao, he finally got down to work examining tissue culture methods to propagate
local plant species. A further 6 month spell working with Dr David Royle on Septoria disease on
winter wheat at Long Ashton Research Station was the project that sparked an interest in plant
pathology and led to searches for a PhD in the discipline. His PhD (1994-1997) research was done
in the laboratory of Alan Vivian at the University of the West of England examining the role of
plasmids in Pseudomonas syringae pathogenicity on plants. A major breakthrough led to 3 years
BBSRC postdoc work with Alan and J ohn Mansfield (Imperial College). In 2001, he moved to the lab
of Paul Rainey in the Department of Plant Sciences, University of Oxford, where a whole new world
of [non-molecular] ecology and evolution did its best to befuddle him. After this highly inspiring
and insightful spell, Rob was awarded a British Society for Plant Pathology fellowship to do a 2 month
research project at the University of Auckland in 2004. The following 5 months were spent back at
UWE working for Dawn Arnold and in late 2004, he joined the lab of Richard Cooper in the University
of Bath studying bacterial extracellular polysaccharides. In 2006, he finally achieved his ambition

viii

by landing a lectureship at the University of Reading.
In his personal time, he enjoys cooking, wine, reading, gardening, scuba diving, hiking and
dreaming that Liverpool might once again lift the league cup!

1

A Survey on Advanced Video-Based Healthcare
Monitoring Systems
Le Thi My Hanh
Department of Information Technology
Danang University of Technology
Danang, VIETNAM
ltmhanh@dut.udn.vn
Hoang Le Uyen Thuc, PhamVan Tuan
Department of Electronic and Communication Engineering
Danang University of Technology
Danang, VIETNAM
{hluthuc, pvtuan}@dut.udn.vn



AbstractVideo-based healthcare monitoring technique has
become an important field in computer vision researches and
applications. Based on the intelligent video analysis, numerous
algorithms are proposed and implemented for healthcare
applications such as monitoring the daily activities of elderly,
detecting a fall, measuring the gait features to detect the early
symptom of some illness, and evaluating the recovery progress of
patients. This paper surveys recent video-based healthcare
monitoring systems, which are quite infant, challenging but
promising. Three main subjects are covered in this survey: first,
we describe a typical video-based healthcare monitoring system;
second, we select recent interested papers to summarize and
discuss in terms of their proposed algorithms, applications,
achievements and limitations; finally, we indicate the key
technical problems and future research directions of the field.
Keywords- video-based, healthcare, fall detection, gait analysis,
activity recognition, rehabilitation.
I. INTRODUCTION
The World Health Organization (WHO) estimates that
there is a need of 4.3 million additional trained health workers
worldwide to address basic health requirements [1]. Especially,
the scarcity of trained health workers has reached serious crisis
level in 57 countries. From among 57 above countries, 8
countries including Angola, Benin, Cameroon, Ethiopia,
Zambia, Haiti, Sudan and Vietnam with acute shortages of
human resources for health (HRH) are selected as pathfinder
countries [2]. The shortage of health workforce causes serious
damage to the human public health worldwide. Simultaneous
combination of HRH development and relieving their workload
is really a globally urgent issue, in particular to 8 above
countries.
This motivates a promising research branch nowadays
designing, developing and applying the automatic health care
monitoring systems. Among a series of such researches all over
the world, video-based techniques are emerged as the
widespread proliferation thanks to their remarkable advantages
such as the easy installation, the unobtrusive operation and the
convenient maintenance. This problem is great challenging and
sophisticated due to very large variations of the context-
dependent motion appearance in different viewing angles,
different illumination and background, different clothes,
different action speed, camera movement, and occlusion in
terms of human-human occlusion, human-body part occlusion,
human-object occlusion, etc.
According to our careful observation, there is no survey on
video-based healthcare monitoring systems. Our survey, which
aims to provide the comprehensive view of the field,
concentrates on the applications of intelligent video analysis to
automatic healthcare monitoring systems, including monitoring
the daily activities of elderly, detecting a fall, measuring the
gait parameters, and evaluating the recovery progress of
patients. Most of our selected papers are published after 2000
in order to ensure the up-to-date knowledge in the survey.
Video-based healthcare monitoring systems concerns to
automatic recognition of human actions/activities from video
sequences. More specifically, the systems are to analyze and
understand patients actions and/or activities, so it can
facilitate health workers to diagnose, treat and care patients,
resulting in improving the reliability of diagnosis, decreasing
the working load for the medical personnel, shortening the
hospital stay for patients, and improving the quality of life for
patients, as well.
Common functional modules of a typical video-based
healthcare monitoring system include three main processing
steps as shown in Fig. 1.

Figure 1. Overview of the general video-based healthcare monitoring system
In the first step, feature representation, the moving humans
in video input are first extracted using human object detection
and segmentation algorithms. The humans are then
2

continuously tracked using object tracking algorithms. The
tracking results in every frame will be transformed into a
reduced representation set of features (also called feature
vector) simultaneously. In order to achieve the appropriate
performance, the extracted features have to deal with spatial-
temporal scaling variations associated with human actions, as
well as encapsulate the unique characteristics of an action
performed by different persons. A good feature descriptor is
expected to be able to extract the most suitable information
from the input video in order to perform the predefined tasks
with sufficient accuracy without using the full size input. Thus,
the first step is equivalent to condensing each input video
frame into a specified multi-dimension feature vector.
After feature vectors are appropriately obtained, we can
now implement the step 2 which is human action/activity
recognition. Activities can be simple actions such as walking,
waving; complex actions such as ballet dancing, doing
exercise; the interactions between persons-persons such as
hand shaking, hugging; or the interactions between humans -
objects such as preparing a dinner, punching a punch-bag.
Recognition of human activities is a challenging and
sophisticated task due to great diversity in the way an activity
is performed by different people or even by the same people
with varying viewing perspectives or time duration.
The goal of the last step applications is to analyze the
classified activities so that their semantic meaning can be
understood.
In this paper, we do a review for all the 3 stages of the
system. The rest of the paper is organized as follows. In
Section 2, we describe feature representation methods. Section
3 mentions the recognition algorithms. The applications are
introduced in Section 4. The discussion and summary are
described in Section 5, followed by the conclusion in Section 6.
II. FEATURE REPRESENTATION
Here we define the feature extraction as consisting of three
main processes: (1) data capturing, (2) object segmentation and
(3) feature description.
A. Data Capturing
In the video-based healthcare monitoring system, the input
data can be gathered from video cameras as a sequence of
image frames. Video signal can be captured using one camera
[3, 4] or multiple cameras. Normally, 2D analysis requires only
one camera, but the performance can be affected by many
factors such as the occlusion in terms of body - body part,
human object, human human, the viewing angle change as
well as the depth ambiguity. 3D analysis can overcome these
limitations; however it usually requires more than one camera
in order to measure the depth information of the subject to
reconstruct the 3D positions of interested human points [5, 6].
In order to extract the depth information, inexpensive depth
camera such as Microsoft Kinect is the new interest of many
researchers [7, 8]. In some cases, video combined with data
signals provided from optical markers attached to specific body
points or other kinds of wearable sensors are used but this is
out of scope of our paper.
B. Object Segmentation
This step in most of video-based systems is separating the
interested objects from the rest of the image frame called the
background. Based on the mobility of cameras, the object
segmentation can be categorized as two types of segmentation:
the static camera segmentation and moving camera
segmentation. For the healthcare applications, the most popular
method is static camera segmentation where the cameras are
located in the specific positions with fixed angles.
The extraction of moving humans from the input video
stream can be achieved by human object detection and
segmentation algorithms based on temporal difference of two
successive frames [9, 10], or background subtracting from each
image [11, 12].
Background subtraction is known as a powerful object
segmentation method and quite suitable for indoor static
camera environment in recent years thanks to the development
of sophisticated dynamic background estimation and updating
techniques. The basic scheme of background subtraction is to
subtract the image from a reference image that is the estimated
background. The method by Stauffer and Grimson [11] has
today become the most popular background subtraction
method. Each pixel is represented by a mixture of Gaussians
and then updated with new Gaussians during run-time. This
update was done recursively, so as to model the slow changes
in the scene such as illumination change and noises.
In addition, some recent researches implement the shadow
detection and removal, random noise filtering as well as
morphology operations to smooth the boundary and fill the
small holes to create well-defined silhouette images, such as [9,
10, 13]. An example of original frame and corresponding
segmented human silhouette is shown in Fig. 2.

Figure 2. An example of extracted human silhouette [13]
C. Feature Description
The step is to associate the detected humans in the current
frame with those in the previous frames, providing the
temporal trajectories of characteristics of the segmented objects
such as shape, color and motions in the form of features. As
mentioned before, the carefully chosen feature descriptor plays
an important role in the entire video-based system.
There are many feature description methods proposed in
literature, which we classify into two main types called
numeric [14-18] and Boolean (binary) features [19, 20]. The
numeric features are most popular and they are represented as
continuous-valued numbers. Boolean features take either
3

number 0 or 1 to express the binary geometric relations
between certain points of a human pose. Numeric features
include two main categories which are shape-based methods
and flow-based methods. Shape-based approaches attempt to
extract silhouettes of moving people such as [14-16]. Flow-
based methods estimate the optical flow field between adjacent
frames such as [17-18]. One inherent disadvantage of shape-
based features is that they can not capture the internal motion
of the object within the silhouette region. Moreover, the state-
of-the-art background subtraction techniques still cannot
segment precise silhouettes, especially in dynamic
environments. In contrast to shape-based methods, it is not
necessary to implement background subtraction in flow-based
techniques. However, the features are still dependent on the
camera view and therefore are not suitable for action
recognition from monocular videos.
Some of the effective feature descriptors can be listed as
follow:
Bobick and Davis represent an action by a motion
history image (MHI) [14]. In an MHI, pixel intensity is
a function of the recency of motion in a sequence at
that point, i.e., more recently moving pixel is brighter
than the past one.
Diaf and Benlamri represent an action by a single
image which is called motion intensity image (MII)
[15]. They get the MII by aligning the human
silhouette of each background-subtracted binary image
to a reference point and then form a single intensity
image of these silhouettes by taking into account the
difference between each subsequent silhouette. By this
way, the temporal occlusion and the imperfect
extraction are effectively removed.
In [16], Kim et al. adopt the accumulated motion
energy image (AMI) using image differences. For the
robustness of the features, the AMI is then resized to a
NxN sub-image by intensity averaging and a rank
matrix is generated by ordering the sample values in
the sub-image. Figure 3 shows an example of two quite
different sub-images due to different clothes and
backpack and two quite different sub-images but their
corresponding rank matrices are identical.
The method performed by Ahmad and Lee is for
human action recognition from any arbitrary view
image sequence that uses combination of the Cartesian
component of optical flow velocity and human body
silhouette feature vector information [18]. The action
region in an image frame is represented by Q-
dimensional optical flow feature vector combined with
R-dimensional silhouette feature vector.
Muller et al. introduce binary features which is a set of
relations to test whether the position of a specified
body point is in-front-of/behind or right/left a specified
plane, the existence of a bent/stretched pose of a body
part, the touched/untouched scenario of two specified
body points [19]. For example, to characterize whether
the leg is bent or stretched, the angle between thigh and
lower leg is used. The corresponding leg angle
feature is 1 (bent pose) if that angle is less than 120
0

and 0 (stretch pose) if that angle is over than 120
0
.

Figure 3. Example of two different images with identical features [16]
In general, the numeric features have been shown to
achieve relatively good results in human action recognition.
However, they are based on 2D information extracted from
image sequences; therefore, are sensitive to occlusions and
viewing dependent.
Binary features, which are mainly derived from 3D point
coordinates, can better handle the occlusion problems.
Moreover, the binary features are efficient to describe human
poses based on low dimensional feature vectors. Unfortunately,
due to the use of only binary numbers, 0s or 1s, to describe the
geometric relation in a pose, binary features cannot be so
discriminative in describing the sophisticated human body
motion. For example, the Boolean feature describing the hands
pose in a typical jogging action is not efficient because both
hands are repeating a series of forward-backward movement,
while always keeping in front of body plane, i.e., the Boolean
feature values are consistently equal to 1 in every frame
without revealing the actual motion of both hands.
III. ACTIVITY RECOGNITION
After obtaining a feature descriptor, the recognition
problem is performed as a higher level task. It may be simply
considered as a classification problem of temporal feature data,
i.e., statistically identifies the sequence of extracted features
into one of the categories of interested activities. We classify
the recognition algorithms into two main categories of
methods, namely, static and temporal classifications. Temporal
classification can be further classified into template matching
and state-space schemes.
A. Static classification
Static classification methods are not interested in the
temporal information of the image frames.
K-nearest neighbor (K-NN) is a popular closest pattern-
based method due to its simplicity and hence, is able to classify
real-time at a very low cost. The training phase of the
4

algorithm is only labeling and then storing the training
samples. In the testing phase, an unknown pattern is assigned
to the group which has most samples among the K training
samples closest to that testing point. Kumari and Mitra propose
an action recognition method based on the average discrete
Fourier transform (DFT) feature extracted from the small
image blocks then apply K-NN algorithm for recognition [21].
Besides KNN, support vector machine (SVM) provides a
powerful approach to the recognition problems. The main idea
of the standard SVM algorithm is as follows: suppose there are
two 2D data sets separated by some hyperplanes. In the
training phase, it is necessary to choose the optimal hyperplane
so that the margin is maximum in order to minimize the error
rate. Once we have trained an SVM, in the testing phase, we
simply determine on which side of the decision hyperplane a
testing point lies and then assign the corresponding class label.
The standard SVM algorithm is formulated to solve only linear
two-class categorization problems. Therefore, extending
standard SVM algorithm to multi-class and nonlinear has great
significance. Schuldt et al. [22] use local space-time features in
a video and apply SVM to recognize human actions. Laptev et
al. [23] apply non-linear SVM with a multi-channel Gaussian
kernel for recognition of various natural human actions,
including AnswerPhone, GetOutCar, HandShake, HugPerson,
Kiss, SitDown, SitUp and StandUp by building spatial-
temporal bag-of-features.
B. Temporal classification
Unlike static classification method, temporal classification
methods pay attention to the temporal information of the data.
Temporal classification can be categorized as template
matching and state-space schemes.
1) Template matching
Template matching methods represent a human action as a
set of template feature sequences of executed action. When a
video is coming, the sequence of feature vectors extracted from
that unknown video will be compared with the template
patterns to determine the similarity. The highest similarity (or
the smallest distance) is chosen as the criterion for
classification of the actions.
Humans may perform an action in different styles and/or
different rates, and the similarity must be measured considering
such variations. The dynamic time warping (DTW) has been
widely used for matching two human movement patterns.
DTW deals with differences between sequences by operations
such as deletion-insertion, compression-expansion, and
substitution of subsequences [24]. By defining a metric of how
much the sequences differ before and after these operations,
DTW classifies the sequences. Even if the time scale between a
test pattern and a reference pattern may be inconsistent, DTW
can still successfully establish the matching as long as time
ordering constraints hold. Figure 4 shows an example of
matching between two sequences with different execution rates
[25]. Sempena et al. [26] use depth camera Microsoft Kinect to
recover 3D human joints, and then build feature vector using
joint orientation so that the orientation is invariant to human
body size. For action recognition, the sequence of feature
vectors extracted from video input is compared to the list of
defined feature vectors using DTW method. Upper part related
actions such as clap, punch, smash and wave can be recognized
quite well.

Figure 4. Example of two sequences with different execute rates. images
with identical features [25]
The most significant advantage of template matching is
simple implementation. The computational cost is proportional
to the number of sequences in database. However, the template
matching is sensitive to noise and spatial-temporal scale
variances.
2) State-space methods
State-space methods represent a human action as a model
composed of a set of states. The model is statistically trained so
that it corresponds to sequence of feature vectors belonging to
its action class. Generally, one model is constructed for each
action. For each model, the probability of the model generating
an observed sequence of feature vectors is calculated to
measure the likelihood between the action model and the input
image sequence. The maximum likelihood is selected as the
criterion for recognition of the actions. In an action model, a
state can be repeated generated itself with the arbitrary number
of times, so it is superior to template matching in temporal data
recognition problems; and therefore, are able to be widely
applied in the human action recognition.
One very popular model used in the state-space methods is
hidden Markov model (HMM) [27]. HMM structure includes a
hidden Markov chain and a finite set of output probability
distribution. More specifically, an HMM is completely
characterized by a set of three matrices = {A, B, },
where A = transition matrix = {a
ij
}, with a
ij
being the
transition probability from state q
i
to q
j
, (i, j) [1: N); B =
observation matrix ={b
j
(k)}, with b
j
(k) being the probability of
observed output (discrete) symbol v
k
at state q
j
, k [1: M); =
{
i
}, with
i
being the initial state probability.
In order to recognize the actions based on HMM, the use of
HMM covers two stages: training and classification. In the
training stage, each HMM is statistically trained for each
action, using the feature vectors extracted from training data
belonging to its action class. The number of states of an HMM
must be specified, and the corresponding state transition and
observed symbol probabilities are optimized in order that the
generated symbols can correspond to the observed vector
features. In the classification stage, for each model, the
probability of the model generating an observed sequence of
feature vectors is calculated to measure the likelihood between
the action model and the input image sequence. If the
calculated probability of a particular model is highest, it is able
to decide that the action corresponding to that model occurs in
the given input. In Figure 5, a human is assumed to be in one
5

state at each time, and each state generates an observation
which is a feature vector. Each human image in the figure
represents a pose with the highest observation probability b
j
(k)
for each state q
j
.

Figure 5. Example of a HMM to model the waving action
Yamato et al. are the first authors applying HMM into
human action recognition [28]. Every set of time-sequential
images is transformed into a mesh feature vector sequence, and
converted into a symbol sequence by vector quantization. They
then apply HMM to recognize six tennis strokes and achieve
the acceptable recognition rates. Ahmad and Lee present each
action by a set of hidden Markov models [29]; each model is
for each action at any viewing direction. The training data set
contains eight specified views. Experimental results of
different actions from any viewing direction are correctly
classified, which indicates the robustness of this view-
independent method.
IV. APPLICATIONS
The goal of the last step applications is to analyze the
classified activities so that their semantic meaning can be
understood. Activity understanding requires expert knowledge
to characterize the uniqueness accurately and to build the
scenario suitable to each specified application. This makes the
activity recognition techniques to be more valuable and widely
used in our daily lives in forms of numerous numbers of
diversified applications including healthcare environment. In
this section, we are interested in four healthcare applications:
daily life activity monitoring, gait analysis, fall detection, and
rehabilitation applications.
A. Daily Life Activity Monitoring
Daily life activity monitoring mainly focuses on learning
and recognizing the daily life activities of seniors at home. The
proposed systems are to provide seniors an opportunity to live
safely, independently and comfortably. In order to accomplish
this, most of proposed systems are to continuously capture the
movements of individual senior/ multiple seniors at home,
automatically recognize their activities and detect the gradual
changes in baseline activities such as mobility functional
disabilities, mental problems, as well as the urgent warning
signs of abnormal activities such as falling down or stroke.
Some of these applications can be listed as follows:
Respiration behavior can be critical in diagnosing
patient illness or recognizing distress during sleep.
Many diseasessuch as obstructive sleep-apnea
syndrome, cardiovascular disease, and strokeinduce
abnormal respiration. Automated respiration
monitoring is performed by Lee et al. [30]. They use
near-IR images, captured by a near-IR camera to
measure the sleeper's respiration based on the periodic
rising and falling motion of their chest or abdomen.
J. Gao et al. measure feeding difficulties in nursing
home residents with severe dementia, by automatically
measure number of hand movements to the mouth
using motion feature vectors and HMM to identify the
start and ending of individual dining events [31].
Huynh et al. present a method for detecting and
tracking of face, mouth, hands and medication bottles
in the context of medication intake monitored by a
camera [32]. This aims to monitor medicine intake
behavior of elderly at home to avoid the inappropriate
use.
In order to recognize the activities as the higher semantic
level, the activity duration, the position of human, the
interaction between people and person-object are the essential
elements received lots of interests.
For the activity duration, Luhr et al. use the explicit
state duration HMM (ESD-HMM), in which a duration
variable is introduced into the standard HMM [33]. In
[34], similar to [33], Duong et al. also add the duration
information to the standard HMM to result in hidden
semi-Markov model (HSMM); however, they model
the state duration by using the generic exponential
family. In [35], Duong et al. introduce the switching
hidden semi-Markov model (S-HSMM) which exploits
the benefit of both the inherent hierarchical
organization of the activities and their typical duration.
For the human position, in [33, 34], the door, the stove,
the fridge, the sink, the cupboard, and the table areas
are used to define the activities in kitchen room. For
example, the meal preparation and consumption
include twelve steps: take-food-from-fridge bring-
food-to-stove wash-vegetable come-back-to-
stove-for-cooking take-plates/cup-from-cupboard
return-to-stove-for-food bring-food-to-table
take-drink-from-fridge have-meal-at-table clean-
stove wash-dishes-at-sink leave-the-kitchen. In
[36], the kitchen is quantized into 28 square cells of 1
m
2
each and the position of the human is captured by
four cameras mounted at the ceiling corners, and the
tracking system returns the list of cells visited by
person.
For the interaction, Liu et al. propose an interaction-
embedded hidden Markov model (IE-HMM)
framework for detecting and classifying individual
human behaviors and group interactions in a nursing
home environment [37]. The framework comprises a
switch control (SC) module, an individual duration
HMM (IDHMM) module, and an interaction-coupled
duration HMM (ICDHMM) module. The SC module is
to extract the atomic behavior units as an individual
behavior unit (comprising a single participant) or an
6

interaction behavior unit (comprising two or more
participants) based on the distances between the
various participants in each scene, and monitoring the
duration for which these distances are maintained. The
individual behavior units are passed to the IDHMM
module to be classified as the corresponding human
behavior. Similarly, the interaction behavior units are
passed to the ICDHMM module, where the
corresponding interaction mode is classified. The IE-
HMM is applied to analyze the human actions and
interactions in a nursing home environment and get the
overall recognition performance of 100% when applied
for individual human actions and 95% when applied
for group interactions.
B. Gait Analysis
Human gait analysis can be simply described as the study
of human locomotion by measuring body movements, body
mechanics, and the activities of muscles. Most gait analysis
researches focus on the walking gait although people can move
by running, hopping, skipping, or even crawling. Intuitively,
walking gait is just a simple action people do to move whole
body from one place to the other. However, walking is actually
a very complex act and it requires many human organs, such as
brain, ears, eyes, and muscles, to coordinate closely to make it
right.
Lots of researches conclude that loss of the ability to walk
correctly can be a result from a significant health problem, due
to the fact that pain, injury, paralysis or tissue damage can alter
normal gait [38]. In addition, the mental problems also can
cause the gait problems such as the gait slowing is a predictor
of incident dementia [39]. Thus, the main goal of gait analysis
for healthcare is to detect the gait abnormalities, early
symptoms of some diseases, which may affect the gait.
Liao et al. presents a video-based system for analyzing four
posture features of walking human including body line, neck
line, center of gravity (COG) and gait width based on the
extracted silhouettes from front view and side view [40].
Similarly, Leu et al also use two cameras and the feedback
control structure at the segmentation level to extract the gait
features such as torso angle, left and right thigh angles, and left
and right shank angles [41]. With the aim of removing the
inconvenience of many methods that require the captured
images of the walking human from either front or side view, Li
et al. propose a system with much less restriction on walking
direction [42]. The system successfully extracts gait features
such as COG and pace length from images obtained from two
cameras with orthogonal view.
C. Fall Detection
Fall in the elderly is the major health risk as it caused
serious injuries and it is known that fall is the leading cause of
injury deaths among seniors. Foroughi et al. conduct some
methods to real-time detect the fall [43-46]. For example, fall is
detected based on human shape variation. Extracted features
which are combination of best-fit approximated ellipse around
the human body, projection histograms of the segmented
silhouette and temporal changes of head pose, are finally fed to
a multi-class SVM [43] or MLP Neural Network [44] for
precise classification of motions and determination of a fall
event. The other features are based on the combination of
integrated time motion images (ITMI) and Eigen space
technique [45, 46].
D. Rehabilitation Applications
Traditional rehabilitation systems often require patients lots
of clinical visits for the physical therapy exercises and the
scheduled evaluation until full recovering of mobility function
for daily activities. Such clinical visits can be decreased by
innovative rehabilitation systems, which are home-centered
and self-health care with the help of video-based activity
recognition techniques. Moreover, by continuously monitoring
the daily activities and gaits, the early symptoms of some
diseases can be timely detected so that the diagnosis and the
intervention are more useful. Some of these applications can be
listed as follows:
Stroke is a major cause of disability and health care
expenditure around the world. Ghali et al. design a
system to provide real time feedback to stroke patients
performing the everyday kitchen activities necessary
for independent living e.g. making a cup of coffee [47].
They envisage a situation in which a stroke patient
stands at a standard kitchen and makes a cup of coffee
in the usual way. The position and movement of the
patients hands and the objects he/she manipulates are
captured by overhead cameras and monitored using
histogram-based recognition methods. The key events
(e.g. picking up a cup) are recognized and interpreted
in the context of a model of the coffee-making task.
In order to objectively evaluate the improvement of
motor functions of the elders at home, as well as
reduces the burdens on fitness instructors, Ryuichi et
al. propose the multimedia fitness exercise progress
notes system [48]. They use a popular video camera
to capture the movements of the elders doing exercises,
then send videos to an analysis center: on a screen,
video files can replay to extract still images, so that
many kinds of measurements, such as lap time,
distances and angles can be performed using a mouse.
Goffredo et al. propose the Gauss-Laguerre transform-
based (GLT-based) motion estimation method in order
to analyze the sit-to-stand (STS) motion from
monocular videos [49]. STS movement mainly
involves hip and knee flexion-extension, and ankle
plantarflexion-dorsiflexion is analyzed by utilizing a
2D human body model that considers the projections of
body segments on the sagittal plane.
V. KEY PROBLEMS AND POTENTIAL RESEARCH
DIRECTIONS
Although the recent video-based healthcare monitoring
approaches have achieved encouraging performance, there are
apparent problems that make it extremely difficult for real-
7

world applications. The performance of the system is adversely
influenced by several factors as follow:
Viewpoint issue may be the main challenge for human
activity recognition. In the real healthcare monitoring
system, the activity sequences are usually observed
from arbitrary camera viewpoints; and therefore, the
applications require the view-independent methods.
This means that the performance of system needs to be
invariant from different camera viewpoints. However,
most recent algorithms are based on the constrained
viewpoints, such as person has to be in front-view (i.e.,
face a camera) or side-view. Some effective ways to
solve this problem have been proposed, such as using
multiple cameras to capture different view sequences
then combining them as training data.
Most of moving human detection algorithms are based
on background subtraction, which requires a reliable
background model. In practice, some humans often
walk in a complex and dynamic cluttered background,
and/or varied light condition (e.g., day, night).
The natural human appearance can be changed due to
many factors such as the walking surface conditions
(e.g., hard/soft, level/stairs, etc.), clothing (e.g., long
dress, short skirt, coat, etc.), footgear (e.g., stockings,
sandals, slippers), object carrying (e.g., backpack,
briefcase).
Promising directions for research in future are outlined
below as potential solutions to those challenging problems:
The view-invariant method from monocular videos,
although getting some preliminary results, is still an
open and challenging research direction.
The human detection is still a problematic for real-time
video-based systems. A large number of key problems
still need to be solved, such as how to decrease the
computational requirement, how to deal with the
practical cluttered background, how to distinguish the
static background with the motionless human, how to
track multiple people, how to handle the occlusion in
terms of body-body part, human-human, human-
objects, etc.
The change of human action appearance leads
researchers to a promising research branch: how to
describe the movement that is less sensitive to
appearance but still capture the most useful and unique
characteristics of each action.
In addition, the patient behavior understanding has become
a dynamic and challenging research branch. It attracts more
and more researchers all over the world. Behavior
understanding requires the additional contextual information
such as W5+(who, where, what, when, why, how) [50]. The
same action may have several different meanings depending on
the context in which it is performed; therefore it can be
recognized as different behaviors. Place context can provide
the location information to be used to detect abnormal
behaviors. For example, lying on the bed or a sofa is taking
relax or sleeping, but at the irrelevant places such as floor in
bathroom or kitchen, it can be a falling or a stroke. Time
context is also a popular contextual description for behavior
understanding. For example, a person usually watching TV
after 2 am can be regarded as insomnia. Another example is
that a person will be detected as picking up stuffs if he squats
and stands up soon. But if he squats for a while, he can have a
motion difficulty due to osteoarthritis or senility. Moreover, the
number of repetitions of an action is also a good hint. For
example, eating too many times or too little a day can be an
early symptom of depression. The interaction between people
or between person and objects is also a good context to identify
the action meaning. For example, if a person is punching a
punch-bag, he might be doing exercise. But if he is punching
himself, it will be a mental disorder.
VI. CONCLUSIONS
There have been a series of research projects on video-
based healthcare monitoring system all over the world in recent
years. The aim of the survey is to review the related works on
video-based healthcare monitoring techniques, including all
modules of a typical system. Four video-based scenarios which
are monitoring of daily life behaviors, analysis of gait,
detecting of fall and evaluating of therapy exercises of patients
were discussed. Although video-based healthcare monitoring
approaches have achieved preliminary achievements, at this
stage, there are still lots of technical problems need to be
solved for real-world applications.
REFERENCES
[1] J . J . Divino, New Financing Tool to Help Reserve Global Health
WorkforceShortage,from
http://www.who.int/workforcealliance/media/news/2009/rrtlaunchpr/en/,
2009.
[2] Global Health Workforce Alliance, WHO, GHWA Collaboration with
eight Pathfinder Countries, Status Report, Geneva, Switzerland, Nov.
2008.
[3] Xin Zhanh and Guoliang, Dual Gait Generative Models for Human
Motion Estimation From a Single Camera, IEEE Trans. On System,
Man and Cybernetics Part B: Cybernetics, vol. 40(4), 2010, pp. 1034-
1049.
[4] Shian-Ru Ke, J enq-Neng Hwang, Kung-Ming Lan, Shen-Zheng Wang,
View-invariant 3D Human Body Pose Reconstruction using a
Monocular Video Camera, 5
th
ACM/IEEE Int. Conf. on Distributed
Smart Cameras, 2011, pp. 1-6.
[5] Yu-Ren Li, Shaou-Gang Miaou, Charles Kater Hung, and J ulius Tube
Sese, A Gait Analysis System Using two Cameras with Orthogonal
View, Int. Conf. on Multimedia technology, 2011, pp. 2841-2844.
[6] Gunawan Ariyanto and Mark S. Nixon, Model-based 3D Gait
Biometrics, Int. Joint Conf. on Biometrics, 2011, pp. 1-7.
[7] Erik E. Stone and Marjorie Skubic, Passive In-home Measurement of
Stride-to-Stride Gait Variability Comparing Vision and Kinect Sensing,
33
rd
Annual Int. Conf. of the IEEE EMBS, 2011, pp. 6491-6494.
[8] Erik E. Stone and Marjorie Skubic, Passive, In-home Gait
Measurement Using an Inexpensive Depth Camera: Initial Results, 6
th

Int. Conf. on Pervasive Computing Techologies for Healthcare and
Workshops, 2012, pp. 183-186.
[9] R. Cucchiara, C. Grana, M. Piccardi, and A. Prati, Detecting moving
objects, ghosts, and shadows in video streams, IEEE Trans. Pattern
Anal. Mach. Intell., vol. 25, no. 10, pp. 1337-1342, Oct. 2003.
[10] J . S. Hu, T. M. Su, and S. C. J eng, Robust background subtraction with
shadow and highlight removal for indoor surveillance, 2006 IEEE/RSJ
Int. Conf. Intell. Robots and Syst., pp. 4545 4550, Oct. 2006.
8

[11] C. Stauffer and W. E. L. Grimson, Adaptive Background Mixture
Models for Real-time Tracking, Proc. IEEE ICCV 1999, pp. 246-252,
Jun. 1999.
[12] A. Elgammal, D. Harwood, and L. S. Davis, Non-parametric Model for
Background Subtraction, Proc. ECCV 2000, pp. 751-767, Jun. 2000.
[13] Shian-Ru Ke, J enq-Neng Hwang, Kung-Ming Lan, Shen-Zheng Wang,
View-invariant 3D Human Body Pose Reconstruction using a
Monocular Video Camera, 5
th
ACM/IEEE Int. Conf. on Distributed
Smart Cameras, 2011, pp. 1-6.
[14] Aaron F. Bobick and J ames W. Davis, The Recognition of Human
Movement Using Temporal Templates, IEEE Transaction on Pattern
Analysis and Machine Intelligence, vol. 23(3), pp. 257-267, Mar. 2001.
[15] Abdunnaser Diaf and Rachid Benlamri, An Effective View-based
Motion Representation for Human Motion Recognition, Proc. Of Int.
Symposium on Modeling and Implementation of Complex Systems, May
2010.
[16] Wonjun Kim, J aeho Lee, Minjin Kim, Daeyoung Oh, and Changick
Kim, Human action recognition using ordinal measure of accumulated
motion, EURASIP journal on Advances in Signal Processing, vol.
2010, Article ID 219109, 11 pages, J un. 2010.
[17] S. Danafar and N. Gheissari, Action recognition for surveillance
application using optic ow and SVM, Proc. ACCV, Tokyo, J apan,
Nov. 2007.
[18] Mohiuddin Ahmad and Seong-Whan Lee, HMM-based Human Action
Recognition Using Multiview Image Sequences, Proc. IEEE ICPR,
2006.
[19] Meinard Muller, Tido Roder, Michael Clausen, Efficient content-based
retrieval of motion capture data, Association for Computing Machinery
Transactions on Graphics, vol. 24, Issue 3, pp. 677-685, J ul. 2005.
[20] Tido Roder, Similarity, Retrieval, and Classification of Motion Capture
Data, PhD Dissertation, Mathematic Natural Sciences Department,
University of Bonn, pp. 50-53, Sep. 2006.
[21] S. Kumari, S. and S.K. Mitra, Human Action Recognition Using
DFT, Third National Conference on Pattern Recognition, Image
Processing and Graphics, pp.239-242, Dec. 2011.
[22] C. Schuldt, I. Laptev, B. Caputo, Recognizing human actions: a local
SVM approach, 17
th
Int. Conf. on Pattern Recognition, vol.3, pp. 32-
36, Aug. 2004.
[23] I. Laptev, M. Marszalek, C. Schmid, B. Rozenfeld, Learning realistic
human actions from movies, IEEE Conf. on Computer Vision and
Pattern Recognition, pp.1-8, J un. 2008.
[24] J . K. Aggarwal and Q. Cai, Human Motion Analysis: A Review,
Computer Vision and Image Understanding, 73(3), pp.428-440, 1999.
[25] J . K. Aggarwal and M. S. Ryoo, Human Activity Analysis: A Review,
Journal ACM Computing Surveys, vol. 43 (3), Apr. 2011.
[26] Samsu Sempena, Nur Ulfa Maulidevi, Peb Ruswono Aryan, Human
Action Recognition Using Dynamic Time Warping, IEEE ICEEI, pp.
1-5, J ul. 2011.
[27] Lawrence R. Rabiner, A tutorial on hidden Markov models and
selected applications in speech recognition, IEEE, vol. 77(2), pp. 257-
286, 1989.
[28] J . Yamato, J . Ohya, and K.Ishii, Recognizing human action in time-
sequential images using hidden Markov model, Proc. IEEE CVPR, pp.
379-385, J un. 1992.
[29] Mohiuddin Ahmad and Seong-Whan Lee, HMM-based Human Action
Recognition Using Multiview Image Sequences, Proc. IEEE ICPR,
2006.
[30] J iann-Shu Lee, Yung-Ming Kuo, and Pau-Choo Chung, A Visual
Context-Awareness- Based Sleeping-Respiration Measurement System,
IEEE Trans. Inf. Technol. Biomed., vol. 14, no. 2, pp. 255-265, 2010.
[31] J iang Gao, Alexander G. Hauptmann, Ashok Bharucha, and Howard D.
Wactlar, Dining Activity Analysis Using a Hidden Markov Model,
Proc. 17
th
Int. Conf. on Pattern Recognition, UK, Aug. 2004.
[32] Huynh H. H., et al., Real-time Detection, Tracking and Recognition of
Medication Intake, World Academy of Science, Engineering and
Technology, vol. 60, pp. 280-287, 2009.
[33] S. Luhr, S. Venkatesh, G. W. West, and H. H. Bui, Explicit State
Duration HMM for Abnormality Detection in Sequences of Human
Activity, Proc. 8
th
Pacific Rim Itll Conf. Artificial Intelligence, pp.
983-984, Aug. 2004.
[34] Thi V. Duong, Hung H. Bui, Dinh Q. Phung, and Svetha Venkatesh,
Human Behavior Recognition with Generic Exponential Family
Duration Modeling in the Hidden Semi-Markov Model, Proc. IEEE
ICPR, pp. 1-6, 2006
[35] Thi V. Duong, Hung H. Bui, Dinh Q. Phung, and Svetha Venkatesh,
Activity Recognition and Abnormality Detection with the Switching
Hidden Semi-Markov Model, Proc. IEEE CVPR, vol. 1, pp. 838-845,
J un. 2005.
[36] Tsung-Yen Liao, Shaou-Gang Miaou, and Yu-Ren Li, A Vision-based
Walking Posture Analysis System without Markers, 2
nd
Int. Conf. on
Signal Processing Systems, pp. 254-258, 2010,
[37] Chin-De Liu, Pau-Choo Chung, Yi-Nung Chung, and Monique Thonnat,
Understanding of Human Behaviors from Videos in Nursing Care
Monitoring Systems, Journal of High Speed Network, vol. 16(1), pp.
91-103, 2007.
[38] http://www.patient.co.uk/doctor/Abnormal-Gait.htm
[39] L. M. Waite, D. A. Grayson, O. Piguet, H. Creasey, H. P. Bennett, G. A.
Broe, Gait Slowing as a Predictor of Incident Dementia: 6-year
longitudinal data fromthe Sydney Older Persons Study, Journal of the
Neurological Sciences, pp. 89-93, 2005.
[40] Tsung-Yen Liao, Shaou-Gang Miaou, and Yu-Ren Li, A Vision-based
Walking Posture Analysis System without Markers, 2
nd
Int. Conf. on
Signal Processing Systems, 2010, pp. 254-258.
[41] A. Leu, D. Ristic-Durrant and A. Graser, A Robust Markerless Vision-
based Human Gait Analysis System, 6
th
IEEE Int. Symposium on
Applied Computational Intelligence and Informatics, 2011, pp. 415-420.
[42] Yu-Ren Li, Shaou-Gang Miaou, Charles Kater Hung, and J ulius Tube
Sese, A Gait Analysis System Using two Cameras with Orthogonal
View, Int. Conf. on Multimedia technology, 2011, pp. 2841-2844.
[43] Homa Foroughi, Alizera Rezvanian, and Amirhossien Paziraee, Robust
Fall Detection using Human Shape and Multi-class Support Vector
Machine, IEEE 6
th
Indian Conf. on Computer Vision, Graphics and
Image Processing, pp. 413-420, 2008.
[44] Homa Foroughi, Baharak Shakeri Aski, and Hamidreza Pourreza,
Intelligent Video Surveillance for Monitoring Fall Detection of Elderly
in Home Environments, 11
th
IEEE Int. Conf. on Computer and
Information Technology, pp. 219-224, 2008.
[45] Homa Foroughi, Hadi Sadoghi Yazdi, Hamidreza Pourreza, and Malihe
J adivi, An Engenspace-Based Approach for Human Fall Detection
Using Integrated Time Motion Image and Multi-class Support Vector
Machine, 4
th
IEEE Int. Conf. on Intelligent Computer Communication
and Processing, pp. 83-90, August 2008.
[46] Homa Foroughi, Aabed Naseri, Alizera Saberi, and Hadi Sadoghi Yazdi,
An Eigenspace-Based Approach for Human Fall Detection Using
Integrated Time Motion Image and Neural Network, IEEE Int. Conf. on
Signal Processing, pp. 1499-1503, 2008.
[47] Ahmed Ghali, Andrew S. Cunningham, and Tony P. Pridmore, Object
and Event Recognition for Stroke Rehabilitation, Visual
Communication and Image processing Conf., 2003.
[48] Ayase Ryuichi, Hagashi Terumasa, Takayama Satory, Sagawa Setsuko,
and Ashida Nobuyuki, A Method for Supporting At-home Fitness
Exercise Guidance and At-home Nursing Care for the Elders, Video-
based Simple Measurement System, IEEE 10
th
Int. Conf. on e-Health
Networking, Applications and Service, pp. 182-186, 2008.
[49] Michela Goffredo, Maurizio Schmid, Silvia Conforto, Marco Carli,
Alessandro Neri, and Tommaso DAlessio, Markless Human Motion
Analysis in Gauss-Laguerre TransformDomain: An Application to Sit-
To-Stand in Young and Elderly People, IEEE Trans. on Information
Technology in Biomedicine, vol. 13, no. 2, pp. 207-216, Mar. 2009.
[50] Maja Pantic, Alex Pentland, Anton Nijholt and Thomas S. Huang,
Human Computing and Machine Understanding of Human Behavior: A
Survey, Artificial Intelligence for Human Computing, 2007.


9

Pulse Oximetry System based on FPGA
Kien Nguyen Trung, Ngoc Nguyen Van
Department of Electronic and Telecommunication
Danang University of Technology
Danang,Vietnam
ntkien@dut.udn.vn, ngvng@gmail.com


AbstractWith Soft Processor, embedded DSP blocks,
configurable logic blocks and high speed IOs, Field
Programmable Gate Array (FPGA) are potential for a lot of high
performance system on chip (SoC) applications. The paper
proposes a completed method using System Generator to
developing a low cost portable Pulse Oximeter System in FPGA
for monitoring patient or old people where many commercial
systems are working on. In FPGA hardware, many implemented
preprocessing algorithms such as Led switching controller, noise
filter and DC tracking, data computing run simultaneously and
so that it is potential to give us very high performance, accuracy,
real-time and low cost Pulse Oximeter System. Moreover, this
Pulse Oximeter IP core can integrate into other IPs such as
Power Monitor IPs, Security IPs to build the completed Smart
House system.
Keywords- Pulse Oximeter, Blood Oxigen Saturation, FPGA,
System Generator, PicoBlaze, SpO
2
.
I. INTRODUCTION
One of biggest objectives in the development of Danang
city to year 2016 is to make the wifi network available
everywhere in the city. This wifi network not only plays an
important role in studying and accessing information from
internet, but also constructs basic facility for many projects in
information technology which applies into the citizen life such
as remotely monitoring the traffic, automatically supplying
tourist information, etc To align with the network
development and to take its advantages, this research develops
a portable medical device to monitor online the old peoples
oxygenation status. This device, with wireless module inside,
can connect to the wifi network to transmit information to
doctors then they can monitor the health of their patients over
the internet. It can also connect to the mobile phone through
bluetooth standard that helps the host software in the mobile
phone monitor the customers health status and do the alarm
action such as sending urgent messages or calling to relatives
when the accident happens.
Pulse Oximetry is a rapid noninvasive measurement of
arterial oxygen saturation and cardiac pulse using simple light
emitter diodes (LED) and sensor application. The measurement
is done directly without pre-adjustment or calibration. There
are two types of pulse oximetry: the transmission type and the
reflection type. Now, most of commercial products are in
transmission type while reflection pulse oximetry is still in
research and evaluation. The name transmission pulse oximetry
means light sources are transmitted through a body part to a
light receiver in opposite side to determine blood oxygen
saturation and the heart beat. With the development of modern
technology, higher integrated circuit, better and smaller LED
and optoelectronic sensor, the early large, heavy and expensive
pulse oximeter with price around 5500 in the pass becomes
much cheaper and more powerful personal portable devices
nowadays as illustrated in figure 1.

Figure 1. Examples of portable pulse oximetry.
In the following sections, we first review the principle of
the transmission pulse oximetry to understand the mechanism
of calculating the blood oxygen saturation and cardiac pulse
from sensor with LED, infrared LED (IR) and an
optoelectronic sensor only. Then third section is our proposal
about a pulse oximetry system based on FPGA technology. We
will try to explain why FPGA was selected and how to build
the system into FPGA. The last section discusses experimental
results, conclusion and improvement solutions.
II. PRINCPLE OF REFLECTANCE PULSE OXIMETER
The typical pulse oximetry configuration on a finger shown
in the figure 2 has two different wavelength light sources: red
LED with 660 nm wavelength and IR LED with 910 nm
wavelength, with a photo-detector in opposite positions. When
only one light source is active, the incident light is the
summary of transmitted light, absorbed light, scattered light
and reflection light on the finger. Fortunately, commercial
probes have very good mechanical structure which rejects
almost scattered light and reflection light so that incident light
is simplified to the summarize of transmitted light and
absorbed light. The principle of pulse oximetry bases on the
absorbance of lights when they are emitted through a part of
body. Each wavelength light source has different light
absorbance on different thickness of skin, color, tissue, bone,
blood, and other material of the body part.


10


Figure 2. Typical finger pulse oximeter with two wavelengths light source is
emitted, diffusely scattered through the finger and detected on the opposite
side by a photodetector. [4]
According to the Beer-Lambert law, the absorbance (A) of
the wavelength light source with a molar absorptivity () is
directly proportional to both the concentration (c) and
pathlength (l) of the absorbing material: A = cl. Actually,
there are four difference components: oxyhemoglobin in the
blood (concentration c
o
, molar absorptivity
o
, and effective
pathlength l
o
), reduced deoxyhemoglobin in the blood
(concentration c
r
, molar absorptivity
r
, and effective
pathlength l
r
), specific variable absorbances that are not from
the arterial blood (concentration c
x
, molar absorptivity
x
, and
effective pathlength l
x
), and all other non-specific sources of
optical attenuation, combined as A
y
, which can include light
scattering, geometric factors, and characteristics of the emitter
and detector elements. The total absorbance at the two
wavelengths can then be written:

During a cardiac pulse, the absorbances that are not from
the arterial blood and other non-specific sources of optical
attenuation have constant values. If we assume two blood
pathlength changes are equivalent, the ratio R of time rate
change of the absorbance at wavelength 1 to that at wavelength
2 can be reduced to below equation.

where I
1,
I
2
are light intensity of received light from photo
detector and I
0
is light intensity of emitted light. In this case I
o

can be removed because it does not change in time rate.
Because the oxygen saturation is calculated by equation S =
c
o
/(c
o
+c
r
) so that we can rewrite it in term of the ratio of R in
following equation
,
by which the saturation was defined as shown in figure 3 (on
the right). The absorptivity () of the red LED and of the
infrared LED was illustrated in Figure 3 (on the left).

Figure 3. Relationship of red R/infrared (IR) numeric ratio value to arterial
oxygen saturation (SaO2). [1]
III. LOW COST PULSE OXIMETER BASED ON FPGA
This section proposes a design of a low cost pulse oximeter
based on FPGA. Normally, the low cost FPGA family does not
have any analog devices such as op-amp, Analog Digital
Converters (ADCs), Digital to Analog Converters (DACs),
discrete electric elements such as capacitors, resistors,
inductors. Therefore, system has to use external devices as
presented in figure 4. All external circuits are used to support
preprocessing signal, amplifier, noise filter, ADC, DAC and
the LED gain controller circuit. With rich logic cell resource,
FPGA is suitable for implementing digital filters, finite state
machine controllers, arithmetic computations and control
algorithms. Moreover, a soft micro controller named PicoBlaze
of Xilinx FPGA is proposed to writing the embedded program
to control the operation of system. Although PicoBlaze is only
an 8 bit RISC micro controller, it has many advantages
included free, easy to assembler, powerful performance, and
small logic size. PicoBlaze Micro-controller embedded in
FPGA provides the optimal balance between FPGA and Micro-
controller solution for the project that requires both higher
performance and controlled state machine application.
In this proposed pulse oximetry system, the DC tracking
using IIR filter and the automatic LED gain controller need to
response immediately to the input signal, hence they should be
implemented by FPGA logic element for parallel computing.
DC signal after the IIR filter then is subtracted by the input
signal to get the AC signal for computing the oxygen saturation
value and heart beat. This work can run with algorithm code in
PicoBlaze micro-controller but in this case, this micro-
controller is assumed to interface all time with the
communication modules such as Bluetooth or Wifi through
common control protocol UART or SPI.

Figure 4. Systemdesign of a finger pulse oximetry bases on FPGA


11

IV. EXPERIMENT
Initial experimental results showed that the maximum
amplitude of signal from pulse oximetry probe is small around
40 - 60 mV and contains a lot of noises. Therefore, integrated
circuit Op-amp AD704 is considered to be used in the pre-
amplifier and noise filter because it is a pico ampere input
current quad bipolar Op-amp specialized for ECG/EKG
instruments. Some great advantages of the Op-amp are active
low pass filter inside with 0.5 Vpp 0.1Hz to 10Hz noises and
very low offset voltage 15 V. This helps the system removes
low noise at the beginning stage but it protects the DC value
from drifting offset for accuracy oxygen saturation computing
in later phase.

Figure 5. a) A simple 3.55Hz lowpass filter and pre-amplifier. b) The 50 Hz
notch filter
A simple circuit low pass filter in figure 5a) was used to
remove non-information signal from photo-detector in pulse
oximetry probe. There are always existence of high frequency
noises from fluorescent lights or radio waves. This low pass
filter was designed to have cut-off frequency f = 1/(2RC) =
3.55 Hz and its gain is H =100K/5K =20. In the next stage of
the preprocessing phase, circuit notch filter is very important to
remove noises with frequency 50 Hz which comes from
electrical grid. This type of noise exists in any stage of the
circuit because of the electromagnetic interference from AC
sources. Therefore, notch filter was considered to be put into
the end of the pre-processing stage of the system. In figure 5b),
the cut off frequency is calculated in equation f
c
= 1/(2RC) =
50 Hz, of which R = R1 =R2 =R3 =R4 =33 k and C =C1 =
C2 =C3 =C4 =100nF.


Figure 6. a) Model of DC tracking filter. b) DC tracking filter designed in
SystemGenerator.
An DC filter tracking is shown in figure 6a). In fact, it is an
simple IIR filter of which the value of K found by experiment
has a value of about 1/f sampling rate [6]. Figure 6b) shows us
the IIR filter design in development tool in Simulink, the
Xilinx System Generator. This tool has all necessary block set
for implement DSP system in FPGA quickly. Moreover, it has
some special tools such as Resource Estimation, Timing and
Power Analysis, and Matlab EDA Simulator that allows us to
evaluate the product in the system design stage. It is known
that Matlab EDA Simulator is a very useful tool to co-simulate
between Simulink and real FPGA hardware (as seen in figure
7).

Figure 7. Co-simulation with Matlab EDA Simulator Link.
V. RESULTS AND DISCUSSIONS
In the first stage of the research, the pre-processing based
on external electronic device was implemented and tested
carefully. Simulation of feature DC extracting of IIR filter is
illustrated in figure 8 shows that designing digital filters in
FPGA using System Generation gives us not only a very fast
tool but also a low cost and low power consumption method.
The filter uses a few hardware resources and draws only 81
mW power supply for operating. System Generator works in
the visualization environment Simulink where many powerful
tools of Matlab can be used to speed up the system design and
verification system.


Figure 8. IIR simulation in System Generator: Channel 1 is a input sample
signal and the Channel 2 is the DC level of input signal

Figure 9a) shows us that the simple low pass filter works
well in removing most of high frequency noises and pre-
(a)
(b)
(a) (b)
12

amplifies original signal with the gain 20. However, signal
after the low pass filter is still contain the 50 Hz interference
noise and this noise was removed completely after the notch
filter as illustrated in figure 9b). Power consumption of two
filters which is about 20 mA is quite good for the mobile pulse
oximetry system for monitoring health status of old people. In
near future, this pre-processing stage will be used to collect
many data from different people whose skin color, skin
thickness, and finger size are different. These data will be used
to calculate range of the gain in the programmable amplifier to
adapt signal to the ADC circuit. It makes sure that the variation
of input signal is still in the range of the analog input to the
ADC. This information decides the resolution and the
frequency sampling of the ADC.
In second stage, as seen in the figure 4, DC value after the
IIR has two parts DC value of Red LED and DC value of the
IR LED. The difference between two values will be used to
control the LED Gain Controller automatically to receive the
same DC value (or the light intensity) in the receiver. In FPGA,
designing a finite state machine for this work is easy and it will
work independent with PicoBlaze. Accordingly, it frees the
micro-controller for focusing only on communication tasks.
The more convenient is System Generator also supports block
PicoBlazed MicroController that means we dont need any
more FPGA development tools to develop the completed SpO
2

system into FPGA.
In the last stage of the project, analyzing the hardware
resource usage and power consumption of the system is the
most important work that makes the system can be
commercialized. The product should be more accuracy than
95%, compared to the common commercial product. The first
selected communication using in this system is Bluetooth
because it is popular in most of mobile phones or laptops. That
helps building software in Android mobiles to transmit old
peoples oxygen status to doctor or relatives through
GSM/GPRS. It is potentially a realistic solution because the
system can also be converted to Very Large Scale Integrated
(VLSI) circuit to be the Pulse Oximetry on Chip with very low
cost in high volume.



Figure 9. a) Noise in signal from the pulse oximetry photodetector (aboved figure) and the effect of pre-amplifer and noise filter on signal.
b) 50 Hz noise signal was rejected by the 50 Hz notch filter.
VI. CONCLUSIONS
This paper proposed a completed method to design and
implement a Pulse Oximetry system on chip based on FPGA
technology. In this method, Simulink and System Generator
blocks are used in all stages of the design from design system
to implementing in to the hardware. Moreover, Matlab EDA
Simulator which can co-simulate between Matlab and FPGA
board, helps the designing faster, easier and more visualization.
First results of this method show us pulse oximetry based on
FPGA is potential compact, low cost and low power
consumption that DSP equation can be implemented and
(a) (b)
(IR LED)
(IR LED)
13

evaluated quickly. In the future, the obtained results can be
improved to a commercialization product of which the higher
compatibility will be achieved when working with Wifi and
Bluetooth available in all mobile phones, computers, and
laptops.
ACKNOWLEDGMENT
The authors would like to thank the Center of Excellence of
Danang University of Technology in Vietnam for their support
of this research project.
REFERENCES
[1] Wukitsch MW, Petterson MT, Tobler DR, Pologe J A. Pulse oximetry:
analysis of theory, technology, and practice, J Clin Monit, Vol 4, pp.
290-301, 1988.
[2] Yitzhak Mendelson, Pulse Oximetry: Theory and Applications for
Noninvasive Monitoring, Clinical Chemistry, Vol. 38, No. 9, pp. 1601-
1607, April 1992
[3] K. Ashoka Reddy, Novel Methods for Performance Enhancement of
Pulse Oximeters, Phd Thesis, Indian Institute of Technology Madras,
April 2008
[4] Ed. J oshep D.Brozino, The Biomedical Engineering Handbook,
Second Edition, CRC Press LLC, 2000
[5] Serhiy Matvieyenko,Pulse Oximeter, Application Note, Cypress,
11/09/2005.
[6] Vincent Chance, Steve Underwood, A Single-Chip Pulsoximeter
Design Using the MSP430, Texas Instrument, November 2005
[7] Vishal Markandey, Pulse Oximeter Implementation on the
TMS320C5515 DSP Medical Development Kit (MDK), Texas
Instrument, November 2010
[8] PicoBlaze 8-bit Embedded Microcontroller User Guide for Spartan-3,
Virtex-II, and Virtex-II Pro FPGAs, Xilinx, November, 2005
[9] Xilinx, System Generator, http://www.xilinx.com/ise/optional_prod/
system_generator.htm.
[10] MathWorks, Simulink, http://www.mathworks.com/products/simulink/.
[11] A Hands-on Guide to Effective Embedded System Design, Xilinx
white paper, 2007
[12] SystemGenerator for DSP User Guide, Xilinx white paper, 2008



14

Towards Efficient Implementation of Neural Networks
with Reduced Precision Floating Point Parameters
Huynh Viet Thang, Nguyen Hai Trieu Anh
Danang University of Technology (DUT), Danang, Vietnam
Emails: thanghv@dut.udn.vn; nhtanh@dut.udn.vn

Abstract Multilayer perceptron artificial neural networks have
widely been implemented on reconfigurable hardware (FPGAs)
to perform a variety of applications including classification and
pattern recognition. These networks are often trained with
double-precision floating-point representations for input data
and parameters using software like MATLAB before they are
implemented on reconfigurable hardware with the optimal
parameters represented in reduced-precision formats. This paper
investigates the effect of reduced precision floating-point formats,
used for the representation of the optimal parameters, on the
recognition rate of the neural network at hand. By gradually
reducing the precision of the floating-point weighting coefficients,
we observed that a reduced-precision floating-point format of 4
mantissa bits is able to provide the same recognition rate as
provided by the highly accurate double-precision floating-point
format. Our work allows for an efficient investigation of tradeoffs
in operand word-length, recognition rate and hardware resources
of floating-point neural network implementations on
reconfigurable hardware.
Keywords-neural network; reconfigurable hardware; floating-
point; bit width allocation; reduced precision; MPFR; design space
exploration.
I. INTRODUCTION
Artificial neural networks have widely been used in
biometrics for the identification of humans. Typical
applications of neural networks in biometrics include
fingerprint and/or face classification/recognition. As of today,
efficient implementations of neural networks on hardware are
very desirable.
Over the last two decades, neural networks have been
implemented on field programmable gate arrays (FPGAs) with
fixed-point arithmetic to perform a variety of machine learning
applications [1-4]. The main reason for using fixed-point
arithmetic for hardware implementations of neural networks in
existing work is that fixed-point numbers require less hardware
resources to implement on FPGAs and provide faster learning
time than floating-point numbers. With respect to numerical
accuracy, floating-point arithmetic has, however, been proven
as the best number format as it can guarantee a more accurate
computational result. Additionally, designing with floating-
point arithmetic is much more convenient for designers as
overflow and underflow are automatically handled.
Despite the superior numerical benefit of floating-point
arithmetic, the use of standard floating-point number formats
(i.e., single precision or double precision) in implementing
neural networks has not attracted many designers because of
two main reasons: i) standard floating-point formats are not
area-efficient as fixed-point formats, and ii) neural networks
seem to be tolerant of precision loss, i.e., a neural network
implementation with 16-bit fixed-point weights could provide
the same recognition performance as provided by another
neural network implementation with 53-bit (double precision)
floating-point weights. In the following, we briefly give a
summary of related work.
A. Related work
Dating back in 1991, Holt and Baker [1] conducted
extensive simulations over a wide range of data sets to compare
the accuracy of limited precision fixed-point neural network
implementations with the accuracy of floating-point ones. A
16-bit fixed-point format was used in their simulations. The
experiments showed that the limited precision fixed-point
simulation is able to perform as well as the floating-point
simulation, suggesting that the 16-bit fixed-point format can
potentially be the typically chosen number format for the
weights of neural network implementations.
Nichols et al. [2] examine the feasibility of using floating-
point arithmetic for FPGA based neural network
implementations. A classic logic-XOR problem is used to
benchmark the learning ability of neural networks. A 16-bit
fixed-point implementation is compared with a 32-bit (single
precision) floating-point implementation. It is shown in [2] that
the 16-bit fixed-point implementation uses less area than and
outperforms the single precision floating-point implementation
in terms of training speed, suggesting that single precision
floating-point format is not feasible for FPGA based neural
network implementation.
In [3], a three-layer fully connected feed-forward neural
network used for character recognition problem has been
developed and implemented on FPGA. Each neuron uses 12-bit
fixed-point weights. Experimental results reveal that the
character recognition system with 12-bit fixed-point weights
provides a comparable recognition rate with another character
recognition system using full precision floating-point weights.
Focusing on a similar direction, Savic et al. [4] has
investigated the impact of number representation on
implementing multilayer perceptron neural networks on
FPGAs. Again, a neural network for solving the classic logic-
XOR problem is implemented. Both floating-point and fixed-
point formats are studied and the effect of precision of
representation and FPGA area requirements are considered. It
is shown in [4] that a multilayer perceptron network with back-
propagation learning uses less clock cycles and consumes less
15


Figure 1. IEEE-754 floating-point number formats.

Figure 2. A simple neural network with one single neuron.
w
1

w
2

w
3

b
hardware resources when compiled in a fixed-point format,
compared with a larger and slower functioning compilation in a
floating-point format with similar data representation bit width
or a similar precision and range.
Even though many authors have made comparison between
using fixed-point format and using floating-point format for
implementing neural networks, it seems that the use of floating
point number formats for neural networks implementations and
machine learning applications has not fully been studied, as
another way of employing floating-point number format is not
completely explored.
B. Scientific contributions
Recently, some research groups, including ourselves, have
found custom precision floating-point number format very
potential for achieving efficient implementation of floating
point algorithms on reconfigurable hardware. More
specifically, we are interested in using floating point formats
with customizable bit width configurations, i.e., using the most
suitable bit width for the implementation of a given algorithm,
retaining the results required accuracy while minimizing a cost
function including hardware resources, execution time and
power consumption. Custom precision floating-point
applications considered include dot products [5-6], support
vector machines [7] and Bayesian network classifiers [8].
Motivated by recent optimistic results in using custom
precision numbers, we aim to examine the feasibility of
floating point arithmetic for implementing multilayer
perceptron neural networks on reconfigurable hardware. Neural
networks operate in two phases: back propagation for obtaining
the optimal weights and forward computation for classification
(i.e., we will consider a character recognition problem in this
paper). In the back propagation phase, neural networks are
often trained with very high numerical precision, i.e., double-
precision floating-point representations for input data and
weights, before they are implemented on reconfigurable
hardware with the optimal parameters represented in reduced
precision formats in the forward computation phase. In this
paper, we only focus on the forward computation phase.
This paper investigates the effect of reduced precision
floating-point formats, used for the representation of the
optimal weights, on the recognition rate of the neural network
at hand. By gradually reducing the precision of the floating
point weighting coefficients, the optimal precision
guaranteeing a required recognition rate can be identified.
To the best of our knowledge, we are the first to investigate
the impact of reduced precision floating-point arithmetic on the
recognition rate of neural networks. Investigation of the impact
of reduced-precision floating-point formats on the back
propagation phase and hardware implementation of multilayer
perceptron neural networks with reduced precision floating-
point parameters will be reserved for future work.
The rest of this paper is organized as follows. Section II
briefly gives some background on floating-point arithmetic and
multilayer perceptron neural networks followed by a
motivating example showing the potential of reduced precision
formats in neural network implementation in Section III. Next,
Section IV presents our design of a multilayer neural network
for the character recognition problem and experimental results,
which investigate the impact of reduced precision formats onto
recognition rate. Optimal floating-point bit width used for
efficient hardware implementation of neural networks is also
identified. Finally, Section V concludes the paper.
II. BACKGROUND
In this section, we briefly present the background for
further studies of custom precision floating-point neural
networks in the next sections.
A. Floating-Point Arithmetic and MPFR
Floating-point arithmetic is the standard approach for
approximating real number arithmetic in modern computers.
The floating-point number format as specified by the IEEE-754
standard [9] consists of three fields: a sign bit (s), a biased
exponent (e) and a mantissa (f). The precision p is equal the
mantissa bit width plus one bit for the implicit leading one.
Figure 1 shows the fields positions as defined for the IEEE-
754 single (p =24) and double precision (p =53) formats.
Hardware cost is typically dominated by the size of the
mantissa field. It is known that reducing the precision
(mantissa bit width) of floating-point operands directly
translates into increased parallelism, and if exploited correctly,
to increased performance [5, 10]. Evaluation of custom
precision floating-point computations can be carried out quite
effectively in Matlab environment via the GNU MPFR
Library [11] (version 3.0.0), in which MPFR C functions are
compiled into MEX files and called from Matlab.
B. Multilayer Perceptron Neural Networks
Neural networks are among state-of-the-art algorithms that
mimic how the human brain works. Neural networks were
widely used throughout the 1980's and 1990's, but their
popularity dropped in the late 90's due to the fact that they are
computationally expensive algorithms. As of today, since
computers are fast enough to run large-scale and compute-
intensive applications, neural networks have a major
resurgence.
16



Figure 3. Multilayer perceptron neural network
Figure 4. Effect of reduced-precision on error of sigmoid function
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
0
0.05
0.1
0.15
0.2
0.25
Precision
A
b
s
o
l
u
t
e

e
r
r
o
r
Absolute error versus precision of the function 1/(1+exp(-z))
The simplest possible neural network consists of a single
neuron, shown in Figure 2. This neuron receives three inputs
3 2 1
, , x x x and produces one output
) ( ) ( ) (
3
1
,

i
i i
T
b W
b x w f b x W f x h , where ) (z f is
the activation function, b is the bias,
T
w w w W ] , , [
3 2 1
is a
column vector representing the weights of the neuron,
T
x x x x ] , , [
3 2 1
is the input column vector, and
b x W z
T
. Different activation function can be used
depending on the inputs/outputs of the network and the
application requirement as well. In this paper, we use the
sigmoid activation function defined by Equ. (1).

) exp( 1
1
) (
z
z f

. (1)
Multilayer perceptron neural networks (Figure 3) consist
of many layers; each layer in turn has multiple simple neurons
that are distributed and operating in parallel. The outputs of
the preceding layer become the inputs of the next layer. In a
fully connected neural network, the output of each neuron in
the preceding layer is connected to inputs of all neurons in the
next layer. In Figure 3, the leftmost layer is the input layer and
the rightmost layer is the output layer.
In the forward computation phase, the computations
performed at each layer include evaluating z and applying the
activation function ) (z f . While the activation function ) (z f
can be a linear or non-linear transform, the evaluation of z is a
linear transform. The evaluation of z is basically a matrix-
vector multiplication. This matrix-vector multiplication is
based on the dot-product computations conducted for every
neurons, in which the quantity z
i
of each neuron is the dot-
product of the input vector and the weight vector.
III. A MOTIVATING EXAMPLE
In multilayer perceptron neural network implementations, it
has been shown that precision loss (due to using reduced
precision number formats) does not significantly degrade the
recognition rate [1, 3, 4]. In the forward computation phase,
two numerical operations are performed at the hidden and
output layers: the matrix-vector multiplication of the weight
matrix and the input vector to the respective layer, and the
sigmoid activation function. As the rounding error of the
matrix-vector multiplication scales with the precision p
following an exponential function
p
2 (see [6] for more detail
about relation between rounding error of floating-point dot-
products with working precision; and the dot product is the
basic operator for matrix-vector multiplications), reducing one
mantissa bit approximately increases the error of the matrix-
vector multiplication by a factor of 2. This relation suggests
that the less degradation of recognition rate of the neural
network due to using reduced precision floating-point format in
the forward computation phase could mainly be due to the
numerical behavior of the sigmoid activation function. In the
following, we present an example to verify this assumption.
We set up a simple example in Matlab to investigate the
effect of reduced precision onto the resulting rounding error of
the sigmoid function. We compute the absolute error of the
reduced precision sigmoid function )) exp( 1 /( 1 z when
gradually reducing the working precision in unit step from
single precision 24 p to the minimum precision 2 p .
MPFR is employed for performing custom precision
computations. The exponent width of reduced precision
numbers is fixed at 11 bits as it is a constraint of MPFR. The
double precision computed version of the sigmoid function is
chosen as the reference value. The chosen input range is
] 10 , 10 [ and 1000 random numbers are used to obtain each
data point corresponding to each precision in the experiment.
The experimental result is presented in Figure 4. As can be
seen, there is no accuracy loss due to reduced precision formats
for high precisions from 24 p to 10 p . From 9 p , the
accuracy loss starts increasing and rapidly becomes even worse
with small precisions.
It is evident from Figure 4 that using a reduced precision
floating-point format with 10 mantissa bits can provide the
same numerical accuracy as using a double precision format
for the sigmoid activation function. As the activation function
mainly defines the output of the network, this motivating
example suggests that the use of reduced precision floating-
point formats is very potential for achieving efficient hardware
implementation of neural networks. An open question is how a
floating-point neural network implementation behaves when
17



Figure 5. Samples of training set (left) and testing set (right)
Figure 6. Impact of reduced precision floating-point formats on the F1
score of the character recognition problem using neural network.
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Precision
F
1

s
c
o
r
e
F1 score versus precision for the character recognition problem


Average F1 score
Min. F1 score
Respective F1 score in double precision (p=53)
we reduce the precision of the weight coefficients, which will
be studied in more detail in the next section.
IV. MULTILAYER PERCEPTRON NEURAL NETWORK WITH
REDUCED PRECISION FLOATING-POINT PARAMETERS
In this section, we study the impact of reduced precision
format onto the performance of a multilayer perceptron
network used for character recognition problem.
A. Neural networks for character recognition problem
There is a large number of character recognition techniques
reported. Most of them use feature extraction technique, which
extracts some features from individual characters, to
distinguish the characters. The extracted features form a
feature vector which is compared with the pre-stored feature
vectors to measure the similarity. After feature extraction,
many different classifiers can be used to recognize characters,
such as neural networks, support vector machines and hidden
Markov model.
In this paper, we directly use the pixels intensities as the
features to a single hidden layer neural network. We build and
train a 3-layer perceptron neural network with 576 neurons in
the input layer, 88 neurons in the hidden layer and 35 neurons
in the output layer. The back propagation technique is used for
training the network in this work.
B. Database and Measurement metrics
To evaluation the effect of using custom precision floating
point number to the classification performance of neural
networks, we use a database consisting of 6651 characters of
24 by 24 dimensions for designing and testing the neural
network. The database is built by our colleagues from Danang
University of Technology. The database is then divided into a
training set of 80% samples and a testing set of 20% samples.
In the training phase we use hold-out cross validation for
tuning the parameters of the neural network; thus, only 60% of
samples were actually used for training, the remaining 20% is
for validating purpose. Samples for training and testing sets
are presented in Figure 5.
In this paper, we use the F1 score for evaluating the test
accuracy of the neural network based recognition system. The
F1 score considers both recognition precision (RP), i.e., the
number of correct results divided by the number of all
returned results (note that in this paper the precision p denotes
the precision of a floating-point number format); and recall
(RC), i.e., the number of correct results divided by the number
of results that should be returned. The F1 score can be seen as
a harmonic mean of RP and RC and is defined as follows:

RC RP
RC RP
F1

*
. 2 . (2)
C. Results
Figure 6 presents the F1 scores versus precision obtained
by simulations with custom precision floating-point arithmetic
in Matlab via MPFR library. The average and minimum
(worst case) F1 scores are both reported. The F1 score
corresponding to using double precision format is used for
comparison. When full precision floating numbers are used,
our experimental results show that the average and minimum
F1 scores on the test set are 99.37% and 92.06 %,
respectively.
For both average and worst cases, the recognition rate of
the neural network degrades significantly when the used
precision is less than or equal to 3 bits.
With a reduced precision format with 4 mantissa bits, the
neural network for character recognition problem can perform
as well as the highly accurate double precision format version.
V. CONCLUSION
In this work, we have investigated the impact of reduced
precision floating-point numbers on the recognition rate of
multilayer neural networks. By gradually reducing the
precision of the floating point weights, we observed that a
floating point format of 4 mantissa-bits is able to provide the
same recognition rate as provided by the highly accurate
double-precision format.
Our work allows for an efficient investigation of tradeoffs
in the operand word-length, recognition rate and hardware
resources of floating-point neural network implementations on
18

reconfigurable hardware. In fact, with a very compact
floating-point format with 4 mantissa bits for character
recognition problem, it could also be possible to try
implementing neural network with reduced precision floating-
point number format on low cost, low power embedded
systems.
The approach presented in this paper can be repeated with a
very little effort in other recognition and classification
problems using neural network as well, for example, face
recognition or fingerprint classification.
Future work will i) investigate the impact of reduced
precision floating-point parameters on the learning speed of
the back propagation technique and ii) focus on the
implementations of artificial neural networks on hardware.
ACKNOWLEDGMENT
We would like to thank Dr. Tuan V. Pham for his valuable
advice and for allowing us to use the database of characters for
training and testing the neural network.

REFERENCES
[1] J . L. Holt and T. E. Baker, "Back propagation simulations using limited
precision calculations," in Neural Networks, 1991., IJCNN-91-Seattle
International Joint Conference on, vol. ii. IEEE, J ul. 1991..
[2] K. R. Nichols, M. A. Moussa, and S. M. Areibi, "Feasibility of Floating-
Point arithmetic in FPGA based artificial neural networks," in In CAINE,
2002, pp. 8-13.
[3] M. Hoffman, P. Bauer, B. Hemrnelman, and A. Hasan, "Hardware
synthesis of artificial neural networks using field programmable gate
arrays and fixed-point numbers," in Region 5 Conference, 2006 IEEE.
[4] A. W. Savich, M. Moussa, and S. Areibi, "The impact of arithmetic
representation on implementing MLP-BP on FPGAs: A study," Neural
Networks, IEEE Transactions on, vol. 18, no. 1, pp. 240-252, J an. 2007.
[5] M. Mcke, B. Lesser, and W. N. Gansterer, Peak performance model
for a custom precision floating-point dot product on FPGAs, in 3
rd

Workshop on UnConventional High Performance Computing, 2010.
[6] T. V. Huynh, M. Mcke, "Error analysis and precision estimation for
floating-point dot-products using affine arithmetic," in The 2011
International Conference on Advanced Technology for Communications
(ATC2011). IEEE, Aug. 2011.
[7] B. Lesser, M. Mcke, and W. N. Gansterer, "Effects of reduced
precision on floating-point SVM classification accuracy," in
International Conference on Computational Science (ICCS
2011). Elsevier, J un. 2011.
[8] S. Tschiatschek, P. Reinprecht, M. Mcke, and F. Pernkopf, "Bayesian
network classifiers with reduced precision parameters," in European
Conference on Machine Learning and Principles and Practice of
Knowledge Discovery in Databases (ECML PKDD), 2012.
[9] IEEE standard for floating-point arithmetic, IEEE Std 754-2008, pp. 1
58, 29 2008.
[10] T. V. Huynh, M. Mcke, and W. N. Gansterer, "Native double-precision
LINPACK implementation on a hybrid reconfigurable CPU," in 18th
Reconfigurable Architectures Workshop (RAW 2011). IEEE, May
2011.
[11] The GNU MPFR Library: http://www.mpfr.org/ (last accessed: Nov 09,
2012).

19

HUMAN-COMPUTER INTERACTION
USING ULTRASOUND HAND TRACKING VIA USB INTERFACE
Sau V. Tran, Tuan K. Tran, Tu T. Hoang
Electronics and Telecommunication Department
Danang University of Technology
The University of Danang
{tranvan678,boydtvtbk,thanhtu08dt3}@gmail.com
Trung X. Pham
Electronics and Telecommunication Department
Danang University of Technology
The University of Danang
trung_ete@dut.edu.vn


AbstractThis paper presents a hand tracking method by
adopting ultra-sound sensors. Besides, the paper demonstrates a
new device that can be connected to computer via USB interface
in order to recognize complex one-hand gestures. The device uses
the Doppler effect to characterize complex movements of one-
hand through the spectrum of an ultra-sound signal. The
parameter of time-delay, amplitude and spectrum at the receiver
changes according to the movement of the hand such as speed,
direction and location. The outstanding character of this method
andgadget is useful and user-friendly because it could potentially
control and drive any application on computer; furthermore, any
untrained person could use it easily. (Abstract)
Keywords-Ultra-sound; Ultra-sound sensor; hand tracking;
hand gestures; human-computer interaction(key words)
I. INTRODUCTION
The act of gesturing is an integral part of human
communication because it may be used to express a variety of
feelings and thoughts such as fury, despair, pleasure, etc. In
fact, gestures can be the most natural way for humans to
communicate. In the digital era, there are more and more
communication technology achievements in recognizing hand
gestures as a major mode of interaction between the users and
the system. Such gesture-based interfaces would operate
effectively when they understand these gestures accurately.
However, this is a difficult task and remains an area of active
research. In order to simplify the task, gesture-recognizing
interfaces in our research typically use a variety of simple
assumptions.
In order to create interaction between users and the devices
of the DiamondTouch, Microsoft Surface and iPhone, people
have to touch their surface. This means that they require the
users to hold a device and make the simplest inferences that
might be deduced from the acceleration of a hand-held device.
To other mechanisms, they have to make more generic
inferences to classify into [1] mouse or pen-based input, [2]
methods that use data-gloves, and [3] video-based techniques.
Mouse/pen-based methods get highly accurate at identifying
gestures; however, these methods require the user to be in
physical contact with a mouse or pen (in fact, the
DiamondTouch, Surface and iPhone may all arguably be
claimed to be instances of pen-based methods with a special
glove). In addition, such required actions could be just applied
to a few applications as well as not suitable to the disable
people.
In this paper, we propose a dissimilar method based on the
Doppler Effect for the recognition of one-hand gestures. Our
device is structured by a single ultrasonic transmitter and a
receiver. The transmitter emits an ultrasonic tone that is used to
reflect hand movement. This tone will continue to be processed
through a Doppler frequency shift phase that is dependent on
the current velocity of the hand. Then, it forms the general
characteristic of the hand velocity in multiple directions as
function varying according to time. At the end, the signals are
captured by the receiver to recognize the specific gesture.
Another strength of this device is its cost effectiveness, it
costs under $5.33. Indeed, the ultrasound device based on
gesture recognizer is really cheap because its processing signal
is simple. Besides, experiments show the exact recognition of
the number of vocabulary of one-handed gestures in the office
environment of classification schemes and the ultrasound
device is also a factor to reduce cost.
II. THE DEVICE STRUCTURE AND OPERATION PRINCIPLE
The ultrasound Doppler based device used for gesture
recognition is an extension of the device suggested by
Kalgaonkar and Raj [8]. The device uses the Doppler Effect to
characterize complex movements of objects through the
spectrum of an ultrasound signal and time among signal
transmit and receiver. The reflected signal is captured by
receiver sensor in order to characterize the motion of object.
A. The device structure









Figure 1: Ultrasound device.
20

Figure 1 shows a picture of our Doppler Effect-based
gesture recognition device. It consists of a MSP430F5510
microcontroller and a LED indicator. The microcontroller is
the main component putted at the central processing circuit to
connect with the computer via USB port - 12mHz external
crystal. With the aim of simplifying the design and reducing
the size of the circuit, a Bootstrap loader button is added in
order to download firmware from computer via USB instead of
the JTAG interface. Another advantage of MSP430F5510
microcontroller is an integrated 3.3V low drop-out regulator
with sufficient output to power entire MSP430 and system
circuitry from 5V of VBUS.
The 40kHz-signal which is generated by the algorithm, will
be taken to ultrasound transmitter through a transistor power
amplifier. The power amplifier is used to control the range of
the device. Long-range sensor can be used by people with
disabilities efficiently.
The characteristic of the hand movement are presented by
the signals and frequency shifts of the 40 kHz center receiver.
This bandwidth of the signal receiver which is less than 40 kHz
would be not recognized. The received signals are digitized by
sampling. When the receiver is turned highly, the principle of
band pass sample may be applied; then, the received signals
are sampled at conversion rate of 200ksps.
All gestures to be recognized will be performed in front of
the device. The sensitive range of the receiver sensor depends
on the power of the transmitter and the gain of receiver circuit.
A noise-remover algorithm will be added to avoid capturing
random movements in the field of the sensor.
B. Principle of operation








The ultrasound device operates basing on the principle of
the Doppler effect and the distance from the hand to the device.
The phenomenon capturing movement in the shift of
frequency of ultrasound in response to a moving object is
called the Doppler Effect. This frequency shift is proportional
to source regularly and to the velocity of the moving object. In
our approach, the original transmitter and listener are
stationary, thus in absence of any motion, there is no change of
frequency. Hand movement would be reflected by waves
causing a shift in frequency. The reflected signals are then
measured by the receiver sensor (fr) and can be described by
the following equation which is used for Doppler radar as well
as for estimating frequency changes in reflection of light by a
mobile mirror.
) .(
v c
v c
t
f
r
f

(1)
Where,
f
r
: perceived frequency at receiver.
f
t
: original frequency from transmitter.
c: speed of sound in air.
v: velocity of target (hand).
The distance is calculated on the basis of propagation time
by the following formulae:
2
34600 10 _
6

time prop
dis
(2)
Where,
dis : distance from object to device (cm).
prop_time: propagation time (us).
III. HAND MOVEMENT RECOGNITION ALGORITHM
The signal which reflects on this device would be
processed to recognize the gestures of hand. Before a
command is sent to computer from microcontroller, we build a
program execution consists of two main components: an
application program running on the computer and a processing
data running on the MSP430 microcontroller.
A. Application program











Figure 3 shows the tasks of this overall block receiving
data from the computer's USB port, then processing data and
finally sending dummy commands to control applications.
USB Driver: Exchange data between computer and the
device. Here we use the driver developed by Texas Instruments
called "MSP430 USB Debugging Interface".
User Interface: The program interface and algorithm
are designed by using C#and Microsoft Visual Studio SDK
that aim to reduce time to develop an application.
Figure 2: Transmitted and received
Figure 3: Computer application diagram
21

Command Mapping: Translate data that is received
from the device to correspond key code.
Get Active Windows: Get handle of current active
window.
Send Command: Send a dummy command to control
application which is active window; for example, mouse click
event, mouse move event.
MSP430 sends a key code to computer by using USB
communication. According to operation of the computer, USB
drivers will receive and transfer data to our control application.
Here, the key code will be mapped onto corresponding
command. Rich set of command is really simple such as: next -
back, up - down, enter, back, exit, page up, page down... Then,
this command will be sent to the current active window that we
focus on.
For example, when users are suffering the internet and
viewing a nice picture on Chrome browser, if they wave their
hand from right-to-left, the program will send a command
"back" or if they wave their hand from left-to-right, the
program will send a command "next" to the browser to view
next image without using mouse or keyboard.
B. Data processing program
USB Interface: Exchanging data between microcontroller
and computer by using USB communication is very difficult to
implement. To save time for developing applications, we use
the USB framework which was provided by Texas
Instruments. This is a powerful tool and fully compatible with
the MSP430.
Ultrasound Generator: MSP430 drives the transmitter to
generate a 10-cycle burst of 40 kHz square-wave signal which
derives from the crystal oscillator with a time interval of 3ms.
Distance Measurement: By measuring travel time
between the first pulse at transmitter and the first received
pulse at receiver, we can determine the distance between the
object and the sensor.
Fast Fourier Transform: Transform data from the time
domain to the frequency domain.
Command Recognition: Result of FFT algorithm will be
analysed and mapped onto corresponding command such as
next, back, up, down, mouse move.








After a period of 3ms, MSP430 sends a 10-cycle burst of
40 kHz square-wave signal sequence to the transmitter
amplifier circuit. Ultrasound waves will spread over space with
propagation velocity 346m/s.
When encountering obstacle, the ultrasound waves will
react back to the receiver sensor. Here, MSP430 will measure
transmission time to determine distance from device to
obstacle. Then, microcontroller will get 128 samples from
analog-to-digital module for FFT module. Fast Fourier
Transform algorithm will transform all data from time domain
to frequency domain. Command recognition module will
compare the received signal spectrum with original spectrum
to determine the movement of obstacle by using the principle
of the Doppler effect. Finally, a key code will be sent to
computer to control current active application.
The following is flowchart the algorithm of program



















Two characteristic of distances and the frequency spectrum
are used to recognize hand gesture. These gestures are
performed within the range of the device. The orientation of
the fingers and palm has no bearing on recognition or the
meaning of the gesture. Figure 6 shows the sequence distance
from the hand to the device during time a gesture command is
executed.
Command enter-back: the undulating form according to
hand moving up and down.
Command left-right: the hand is horizontal movement so
that the distance is virtually unchanged.
Figure 5: Transmitter and processing
Figure 4: Data processing program
22

Command page up-page down: there is a strong variation
of distance.
























Figure 7 shows the spectrum of the received signal. This
spectrum has two peaks:
The first peak represents the negative artifacts caused by
signal frequency (<20kHz).
The second peak represents the contrast and the signal
spectrum when the arm does not move.
IV. EXPERIMENTS AND RESULT
Each command was tested 60 times, in four different
scenarios with condition of noise, temperature and humidity to
evaluate the stable of the device.
Office: This is the most common environment with 50dB
intensity of noise, temperature of 250C and air humidity of
70%.
Outdoor: 55dB intensity of noise, temperature range:
250C 350C and air humidity of 80%.
Class: 60dB intensity of noise, temperature range: 270C
and air humidity of 70%.
Coffee shop: 75dB intensity of noise, temperature of 300C
and air humidity of 70%.
Table 1 shows accuracy of algorithm, especially in low-
noise environment. Some other environments with high
intensity of noise such as coffee shop, the accuracy may be
decreased but it still meets the requirement to control
applications.

Command Office Outdoor Class Cafe Shop
Up 98.3% 98.3% 96.7% 96.7%
Down 98.3% 98.3% 96.7% 96.7%
Left 96.7% 96% 93.3% 93.3%
Right 95.0% 95.0% 93.3% 93.3%
Page up 96.7% 96.7% 95.0% 93.3%
Page down 96.7% 95.0% 93.3% 93.3%
Enter 95.0% 91.7% 91.7% 90.0%
Back 93.3% 90.0% 88.3% 83.9%

Enter and Back commands are often confused with each
other and with the Left, Right commands (Figure 9). The
confusion could exist because the users are not familiar with
usage, and the other part because of high-noise environment.










V. DISCUSSION AND FUTURE WORK
This paper shows that an ultrasound can identify the good
gesture of the objects. We can control the computer without
having contact with the device.
The high recognition of command performance averaged at
96.25% in the office environment. We believe that command
recognition performance will greatly increased if the next
version having the increasing number of sensors combining
with statistical methods such as HMM, neural network, fuzzy
control methods and confused phenomenon between enter and
back commands is removed. At the same time, it will help
typical gestures to clearly reflect on receiving device. This
makes control easier.

Figure 7: Fourier spectrum
(c) Page down waveform.
(a) Enter Exit waveform. (b) Left Right waveform.
(d) Page up waveform.
Figure 6: the type of waveform
TABLE 1: AVERAGE % OF CORRECTLY RECOGNIZED GESTURES
ACROSS THREE USERS IN FOUR PLACES.
Figure 9: Error rate and distribution for different
commands.
23

REFERENCE
[1] H. He, and J . Liu, The design of ultrasonic distance measurement
system based on S3C2410, Proceedings of the IEEE International
Conference on Intelligent Computation Technology and Automation, Oct.
2008, pp. 44-47.
[2] Y. J ang, S. Shin, J . W. Lee, and S. Kim, A preliminary study for
portable walking distance measurement system using ultrasonic sensors,
Proceedings of the 29th Annual IEEE International Conference of the
EMBS, France, Aug. 2007, pp. 5290-5293.
[3] Sidhant Gupta, Dan Morris, Shwetak N Patel, Desney Tan, SoundWave:
Using the Doppler Effect to Sense Gestures, In Proceedings of ACM CHI
2012, May 2012.
[4] Murugavel Raju, Ultrasonic Distance Measurement With the MSP430,
Application Report, October 2001.
[5] S. S. Huang, C. F. Huang, K. N. Huang and M. S. Young, A high
accuracy ultrasonic distance measurement system using binary frequency
shift-keyed signal and phase detection, Review of scientific instruments,
October 2002.
[6] Texas Instruments, Programmers Guide: MSP430 USB API Stack for
CDC/PHDC/ HID/MSC, in J anuary 2012.
[7] Silicon Labs, Using Microcontrollers in Digital Signal Processing
Applications, Rev. 0.2 8/08.
[8] Kaustubh Kalgaonkar, Rongquiang Hu, Bhiksha Raj, Ultrasonic Doppler
Sensor for Voice Activity Detection, Mitsubishi Electric Research
Laboratories, Inc., 2008
[9] Silicon Labs, Using Microcontrollers in Digital Signal Processing
Applications, Rev. 0.2 8/08.
[10] http://msdn.microsoft.com/en-us/library/k50ex0x 9.aspx
[11] Kaustubh Kalgaonkar, Rongquiang Hu, Bhiksha Raj, Ultrasonic
Doppler Sensor for Voice Activity Detection, Mitsubishi Electric Research
Laboratories, Inc., 2008
[12] Kaustubh Kalgaonkar, Bhiksha Raj, ONE-HANDED GESTURE
RECOGNITION USING ULTRASONIC DOPPLER SONAR 2009 IEEE.
24

A Real-Time Scheduling Scheme for Distributed
Control System

Trong Cac Nguyen, Xuan Hung Nguyen, Van Khang Nguyen
Hanoi University of Science and Technology
e-mail: cacdhsd@gmail.com

Abstract Real-time scheduling is an important mechanism for
Distributed Control Systems (DCS). The real-time scheduling is,
in several networks, mainly based on static priorities associated
to the messages (for example the network CAN). This static
priority scheme has intrinsic limitations (which have been
studied in [1]) and so, in order to overcome these limitations, we
propose in this paper a real-time scheduling strategies based on
three hybrid priority schemes.
Keywords - CAN bus; hybrid priority; real-time scheduling
I. INTRODUCTION
DCS are gaining increased popularity because they offer
several advantages such as modular and flexible system design
(e.g. distributed processing and interoperability), simple and
fast implementation (e.g., small volume of wiring and powerful
configuration tools), and powerful system diagnosis and
maintenance utilities (e.g., alarm handling and supervisory
packets). DCS contain a large number of interconnected
devices that exchange data through communication networks;
examples include industrial automation, building automation,
office and home automation, intelligent vehicle systems, and
advanced aircraft and spacecraft. The specific application
imposes different degrees of timing requirements to the DCS
implementation [2]. Therefore, the successful implementation
of DCS depends on many variables and on a good integration
of different bodies of knowledge. A distributed process control
application includes three remote tasks (the sensor task, the
controller task and the actuator task) which are on different
sites and which then require the periodic exchange of two
message flows through the network: the sensor flow which
concerns the transfer of the output samples, from the sensor of
the process to the controller which computes the control law;
the controller flow which concerns the transfer of the control
samples from the controller to the actuator of the process. The
scheduling through the network of the messages of these two
flows is an essential mechanism which strongly influences the
settling time and the stability [3] of a process control
application.
This paper is precisely concerned with this problem by
considering the network CAN (Controller Area Network) [4]
and the MAC layer which implements the scheduling of the
frames. The scheduling is done by means of priorities which
are represented in the IDentifier (ID) field of the frames.
Different types of priorities (static priority, hybrid priority) can
be considered. Here this paper focuses on hybrid priority
schemes.

Related Work
Branicky et al. [5] studied network issues such as
bandwidth, quantization, survivability, and reliability. They
applied the RMS algorithm to schedule a set of NCSs. Wu et
al. [6] proposed a distributed dynamic message scheduling
method based on deadline of message in order to satisfy
timeliness of messages and improve the systems flexibility
based on CAN. There are various resource allocation and
scheduling techniques now. Simple queuing methods sample
the output of the plant periodically and place the resulting data
record into a first-infirst-out queue. However, the sensor
sampling period must be larger than the average transmission
time interval; otherwise, the queue overflows. Another one is
try-one-discard, which does not implement a queue. Rather, it
discards the data whenever transmission is not possible, for
example, network is unavailable [7]. Mart et al. [8] showed
that the codesign of adaptive controllers and feedback
scheduling policies allows for the optimization of the overall
QoC. They described an approach for adaptive controllers for
the NCS to overcome some of the previous restrictions by
online adapting the control decisions according to the dynamics
of both the application and executing through message
scheduling. Li and Chow [9] proposed sampling rate
scheduling to solve the problem of signal fidelity and conserve
the available data transmission. Al-Hammouri et al. [10]
proposed an asynchronous, scalable, dynamic, and flexible
bandwidth allocation scheme for NCS, formulating the
bandwidth allocation as a convex optimization problem in a
fully distributed manner. Hong et al. [11] scheduled a life
science high-throughput platform using timed transition Petri
nets and heuristic searching. Mart and Velasco [12] reviewed
basic control models for flexible scheduling in real time and
built a new model which allowed irregular sampling while still
having better schedulability and robustness. Pereira et al. [13]
dealt with scheduling real-time transmission in a wireless
industrial network. They developed a new collision-free
wireless medium access protocol (MAC) with static-priority
scheduling. Grenier and Navet [14] highlighted a class of
online scheduling policies targeted at scheduling frames at the
MAC level.
The goal of this paper is mainly to evaluate the Quality of
Service (QoS ) provided by three hybrid priority schemes to a
process control application, to compare these QoS and to
explain their differences. This study is done by using the
simulator TrueTime [15] which allows to represent NCS (both
network and control aspects).
25

II. CONTEXT OF THE STUDY
A. The process control application considered
This application is represented on the Fig. 1: the process
control application has a transfer function
1000
( 1)
( )
s s
G s
+
= ; the
controller is a proportional derivative controller which
considers the output derivation (K is the proportional gain,
K=1.8; T
d
is the derivative time constant, T
d
=0.032s). We
have an overshoot of 5% and the rise time t
r
is about 40ms.

Figure 1. Model
The closed loop transfer function F(s) of this application is:
2
( ) 1000
( )
1 ( )(1 ) (1 1000 ) 1000
d d
KG s K
F s
KG s T s s KT s K
= =
+ + + + +
(1)
B. General view of the implementation through the network
CAN
1) The network operates both
The network operates both is presented in Fig. 2.

Figure 2. Implementation through the network CAN
- Between computer 1 (Cl) in association with the
numerical information provided by the AD conversion
(this computer includes a task that we call the sensor
task and which generates the sensor flow addressed to
the controller, we note fsc this flow) and computer 2
(C2) where we have the reference and the controller (in
C2 we have a task called controller task which
generates the controller flow addressed to the actuator,
we note fca this flow); fsc goes from Cl to C2 and fca
goes from C2 to the computer 3 (C3).
- And between C2 and C3 which provides numerical
information to the DA conversion in front of the Zero
Order Hold (ZOH). The task generating the flow fsc is
time-triggered whereas the task generating the flow fca
is event-triggered. Generally a network is not dedicated
to only one application but shared between different
applications. In order to make a general study of the
behavior of the process control application, when it is
implemented through a network, we have to consider,
in particular, the influence of the flows of the other
applications. It is why we have, in Fig. 2, what we call
the external flow, denoted as fex, which globally
represents an abstraction of the flows of all other
applications. We also consider that fex is periodic.
2) The choice of the sampling period (h) of the process
control application is a basic action.
The sampling period has, from the control point of view, an
upper bound. But from the network point of view, a value that
is too small gives load that is too great. So, the choice results in
a compromise. The relation
10 4
r r
t t
h s s , which has been given
in [16, page 66] is generally used. We consider here the bound
4
r
t
. As t
r
~ 40ms, we have h =10ms.
3) In a general way, the information transmission rate
requested by the applications to a network is the pertinent
parameter to compare the efficiency of the message
scheduling.
We call here this parameter the User Request Factor (URF).
Concerning the network CAN, the scheduling is done by the
MAC layer and concerns the frame scheduling. By calling:
- D
ca
, D
sc
, D
ex
the duration of the fca frame, the fsc frame
and the fex frame, respectively;
- h the sampling period of the process control
application (the period of fsc and consequently of fca)
and T
ex
the period of the external flow. We have:

ca sc ex
ex
D D D
URF
h h T
= + + (2)
In the context of this work, we will consider the following
numerical values:
- Bit rate in the physical layer of CAN: 125 Kbits/s;
- Length of 10 bytes for the fsc frames and fca frames
(thus a duration of D
sc
=D
ca
=640s);
- Length of 15 bytes for the fex frames (thus D
ex
=960s).
The component
ca sc
D D
h h
+ of the URF, which concerns the
process control application and which represents the network
capacity used by this application, has the value 12.8%. The use
by the external frame of the network capacity will depend on
its period T
ex
. It is this parameter that we will vary during our
study in order to analyze the robustness of the scheduling of the
process control application frames.
The frame scheduling in the MAC layer of CAN [4] is done
by comparing the field ID bit by bit (we start from the MSB).
In CAN the bit 0 is a dominant bit and the bit 1 is a recessive
bit. The lower the numerical value of the field ID, the higher
the priority. We consider here the standard length of 11 bits for
the field ID.

G(s)
Process
Output
y u
Controller Input
reference

K

1+sT
d
r _
+

Process
to
control
Output
y


N
E
T
W
O
R
K

C
A
N

Controller
Input
reference
r
u
C2
fca
fsc
fex
y
C1

D
A

Z
O
H
C3


Sensor

A
D
h
26

C. General idea on hybrid priorities
1) The field ID and the scheduling execution.
The identifier field of a frame is divided into two levels, see
in Fig. 3.

Figure 3. The structure of ID field (MSB: Most Significant Bit; LSB: Least
Significant Bit)
In Fig. 3: the first level represents the priority of a flow (it
is a static priority specified off-line); the second level
represents the priority of the transmission urgency (the urgency
can be either constant or variable). The idea of structuring the
ID is present in the Mixed Traffic Scheduler [17] which
combines EDF (dynamic field) and FP (static field). In [3] the
authors propose encoding the weighted absolute value of the
error in the dynamic field (this idea is also presented in [18]) to
resolve the collisions with the static field. A constant
transmission urgency is characterized by a static priority (one
m bit combination) specified off-line. A variable transmission
urgency is characterized by a dynamic priority (which can take,
generally speaking, m-bit combination among a subset of the
m-bit combinations). The frames of the flows fsc and fca of a
process control application have variable needs (strong urgency
in a transient behavior after an input reference change, weak
urgency in a permanent behavior). That is why, in this study,
we consider that the dynamic priority of the frames of the flows
fsc and fca of a process control application can take any m-bit
combination of the set of the m-bit combinations. The
scheduling is executed by, first, comparing the second level
(needs predominance), and, secondly, if the needs are identical,
by comparing the first level (flow predominance).
Remark: for the first level of the field ID we will consider
here: Priority fex >Priority fca >Priority fsc.
2) Cohabitation of flows with constant needs and flows of
process control applications (variable needs)
We have the objective of good performances for the
process control applications in transient behavior. This means
the urgent needs of these flows must be satisfied very quickly.
For that, we impose a maximum with constant needs for the
priority of these needs (concept of Pr_th for the constant
needs). In this way, a strong transmission urgency of a process
control application flow (dynamic priority with a very high
value i.e. higher than Pr_th) will be scheduled first.
Remark: the external flow fex will have in this study
constant needs (characterized by Pr_th).
D. Three hybrid priority schemes
We have defined three schemes. The first is what we call
the strict hybrid priority (hp) scheme (computation of the
dynamic priority directly from a function of the control signal
u; re-evaluation after each sampling instant). The second is the
hp scheme extended with a static time strategy (sts) for the re-
evaluation of the dynamic priority (re-evaluation not always
after each sampling time). This scheme is noted hp+sts. The
third is a scheme which does not compute the dynamic priority
directly from the control signal u (definition of a timed
dynamic priority reference profile and trip in this profile by
means of an on-line temporal supervision based on a function
of the control signal u). The dynamic priority is re-evaluated
after each sampling instant. This third scheme, which
implements a dynamic time strategy for the trip in the timed
dynamic reference profile, is noted hp+dts. We now detail
these three schemes.
1) hp scheme
The needs are translated into a dynamic priority by
considering an increasing function of u (call it ( ) f u
characterized by a saturation for a value of u less than the
maximum of u (noted
max
u ). We do not want the dynamic
priority to take its highest value only when u is maximum but
already for values before the maximum, in order to react
quickly as soon as the needs begin to become urgent. So we
decide (it is an arbitrary choice) to take
2
3 max
u as the value of
u where the dynamic priority reaches its highest value P
max
.
Several functions ( ) f u have been studied, for this work
we consider the function ( ) f u represented in Fig. 4. This
function is defined by:
( )
2
max 3 max
2
3 max
2
max 3 max
,0
,
u
P u u
f u u
P u u

s s

>

(3)

Figure 4. The considered nonlinear function
The computation of the dynamic priority is done by the
controller each time it receives a frame that the sensor sends
after each sampling instant (dynamic priority re-evaluated after
each sampling instant). Then, after the reception of a frame
from the sensor, the controller sends a frame with the value of
the new dynamic priority. This frame reaches all the sites and
as the sensor site knows the first level of the ID of fca (it is a
constraint for our implementation), it will learn the dynamic
priority that it will put in the next frame that it will send (the
dynamic priority is then used by the two flows of a process
control application) .
Taking into account the task implementation (sensor task is
time-triggered, controller task is event-triggered), it is the
sensor task which transmits the first frame at the start of the
second level first level
MSB LSB
m bits (n-m) bits
27

application. For this first frame, the sensor site has no
information about the dynamic priority and thus we consider
that it uses the maximum priority. This way, the first fsc
reaches the controller site as quickly as possible.
2) (hp+sts) scheme
A criticism of the hp scheme is that we can have oscillatory
behavior of the dynamic priority values (resulting from a
damped sinusoidal transient behavior of u). We can have, for
example, this scenario for the dynamic priority values at three
successive re-evaluation instants [19]: the highest value at the
first re-evaluation instant, then an intermediary value at the
second, and again the highest value at the third re-evaluation
instant. Such an oscillatory behavior shows that the control of a
situation requiring a big value of the dynamic priority is
inadequate in terms of the maintenance of this big value, since
after leaving this value for an intermediary one, at the second
re-evaluation instant, we come back to this big value at the
third re-evaluation instant. The observation of this phenomenon
suggests increasing the duration of the dynamic priority with a
big value in order to improve transient behavior.
The (hp+sts) scheme is then the following. Contrary to the
scheme hp, where the dynamic priority is re-evaluated in the
controller site, after each reception of a fsc frame, the instant of
the re-evaluation is no longer so closely related to the sampling
instants. Here the duration of the time interval between two
successive re-evaluations depends on the value of the dynamic
priority at the beginning of the time interval. This duration
must be relevant, in particular, from the point of view of the
transfer function of the process control application and more
precisely, of its transient behavior (defined before its
implementation through the network). We considered the
following algorithm:
- If the dynamic priority has a value between the highest
priority (P
max
) and half the highest priority (
1
2
P
max
), we
keep this value for 4 sampling intervals and we re-
evaluate the dynamic priority afterwards; this duration
is equal to the rise time t
r
which represents a good
characteristic of the transient behavior.
- If the dynamic priority has a value inferior to half the
highest priority, we re-evaluate it after each sampling
instant as in the previous algorithm.
3) (hp+dts) scheme
Main ideas: We define, at first, a reference profile of
dynamic priorities for apprehending with efficiency one
transient behavior (i.e. an input change or a disturbance). It
consists in a continuous decreasing time function from a
priority P
max
(start of the transient behavior) to a priority P
min

(end of the transient behavior and then the beginning of the
permanent behavior), which gives the values of the dynamic
priorities at all the sampling times (these values are
decreasing).
However the only consideration of the reference profile is
not enough to handle the actual behavior. In the actual
behavior, we have to take into account for the influence of the
network and also the possibility of successive input changes
and/or disturbances which lengthen the transient behavior with
respect to the one considered in the reference profile. Then,
actually, the temporal evolution of the dynamic priorities,
cannot be always decreasing, i.e., at a sampling instant, we can,
by considering the reference profile curve, move back to values
higher than the value of the previous sampling instant.
So, in order to take into account for an actual behavior, we
add a component, called on-line temporal supervision based on
a function g(u) which will allow to do, with respect to the
reference profile, the temporal repositioning of the values of
the dynamic priorities. We consider the reference profile
represented in Fig. 5.
The function P(t) is defined by:
( )
2
max max min max
max
( ) ,0
t
P t P P P t t
t
| |
= s s
|
\ .
(4)
P
min
is the priority used in the permanent behavior. The
dynamic priority decreases slowly at the beginning of the
transient behavior (we need several successive sampling
instants with high priority in order to be reactive) and more
quickly towards the end. The reference profile expresses that
the priority, related to the sampling instants, t
k
, is lower than
the priority related to the previous sampling time, t
k-l
.
Concerning the time t
max
: as our objective is to tend towards
a transient behavior guided by the transient behavior of the
process control application without the network, we take t
max

equal to the settling time at 5% of the process control
application without the network.

Figure 5. Reference profile
On line temporal supervision: We have defined several
functions g(u) which allow, at the sampling instant t
k
, to move
back in the reference profile with respect to the previous
sampling instants t
k-l
. These functions g(u) give the time values
which must be subtracted to the value t
k-l
+h to come back more
or less towards the beginning of the reference profile (then
using, at the instant t
k
, a priority higher than at the instant t
k-1
).
Note that the maximum of this time value can be t
max
. Here we
use the function g(u) represented in Fig. 6 and defined by:

2
max 3 max
2
3 max
2
max 3 max
, 0
( )
,
u
t u u
g u u
t u u

s s

>

(5)
Then | |
max
( ) 0, g u t e
28



Figure 6. Example of g(u)
III. STUDY OF THE THREE SCHEMES BASED ON HYBRID
PRIORITIES
A. Study conditions
We consider the process control application which was
presented in the subsections II-A and II-B. The input is a
position step which starts at time 0 and we study the transient
behavior until it reaches permanent behavior. The QoS
parameters, which need to be taken into consideration, are the
mean delayD of the control loop and its standard deviation o.
The QoC parameter is the settling time at 5% (t
s
) which is
obtained directly from the tool TrueTime. In order to evaluate
the QoS parameters, we use, on the one hand, the message
exchange temporal diagrams which are also provided by
TrueTime, and the value of t
s
. From the message exchange
temporal diagrams, we can get the delay in the control loop
(delay of the message of the flow fsc+delay of the message of
the flow fca+D
sc
+D
ca
) for each sampling period (call D
i
this
delay for the sampling period i). Counting the number n of
sampling periods in the settling time t
s
, we deduce the value
of D and o by these formulas:
1
n
i
i
D
n
D
=

= and
( )
2
1
n
i
i
D D
n

=
In order to make a quantitative analysis, we cause a
variation in the network load (URF) by varying the period T
ex

of the external flow: we consider an external flow, the
frequency of which (noted 1/T
ex
) is a multiple of the sampling
frequency (1/h). The different URFs being considered are
given in Table I.
TABLE I. DIFFERENT URFS
URF (%) Multiple of 1/h Tex (ms)
99.2 9 1.1111
89.6 8 1.25
80 7 1.4286
70.4 6 1.6667
60.8 5 2.0
51.2 4 2.5
41.6 3 3.3333
32 2 5.0
22.4 1 10.0
The following important points must still be emphasized:
- The flows fsc (which are generated at the sampling
times) and fex are synchronous (starting at the same
time) and as we consider the cases where the frequency
of fex is a multiple of the sampling frequency, then
their medium access attempts coincide at every
sampling time.
- Up to the value 70.4% of the URF (value of 1.6667ms
for T
ex
), we can see that during T
ex
, one frame of each
flow can access the medium:
0.96ms+0.64ms=1.6ms<1.6667ms (the third flow can
begin to be transferred and then cannot be interrupted).
- A last point must be still noted: at the beginning of a
transient behavior, as the control signal is at a
maximum, the dynamic priority of the flows of the
process control application is P
max
.
B. hp scheme
Concerning the process control application, we giveD and
o in Table II and t
s
in Table III. The values depend on the
network load URF, and on the Pr_th.
TABLE II. HP SCHEME (QOS): D AND o (MS)
Pr_th
URF 0.9Pmax 0.5Pmax 0.25Pmax
(%)
D

D

D

99.2 5.333 1.680 3.743 2.262 1.804 1.380
89.6 3.264 1.286 2.240 1.228 1.629 0.846
80 2.48 0.887 1.987 0.828 1.5418 0.592
70.4 1.891 0.462 1.716 0.478 1.472 0.384
51.2 1.891 0.462 1.716 0.478 1.472 0.384
22.4 1.891 0.462 1.716 0.478 1.472 0.384
TABLE III. HP SCHEME (QOC): TS (MS)
URF Pr_th
(%) 0.9Pmax 0.5Pmax 0.25Pmax
99.2 359 228 105
89.6 148 110 103
80 111 108 101
70.4 107 105 99
51.2 107 105 99
22.4 107 105 99

Concerning the values of D , we observe the following
main points:
1) For each value of Pr_th:
For URFs70.4%, we note that we have the same values
of D and o whatever the value of URF is. This is a
consequence of the fact that the two frames of fsc and fca,
during each sampling period, can be sent during the period of
fex, which is not the case with URF>70.4% whereD and o
increase with the value of URF. We explain the difference by
means of two exchange temporal diagrams provided by
TrueTime (Fig. 7 and 8). On the Fig. 7, we see that the frames
fsc or fca can be delayed, during a sampling period, for the
duration of one frame of fex (0.96ms). On the Fig. 8, we see
that the two frames of fsc and fca can be delayed and the delays
for the frame of fca can be more than the duration of one frame
of fex. Note then, when URF>70.4% and for increasing values
of URF, D increases because the network load increases (thus
more chances to delay the frames of fsc and fca).
29


Figure 7. hp scheme (URF=70.4%, Pr_th=0.9Pmax): time exchanges

Figure 8. hp scheme (URF=89.6%, Pr_th=0.9Pmax): time exchanges
2) For increasing values of Pr_th:
D also increases because the dynamic priorities of the
frames of fsc and fca have fewer chances of being higher than
the threshold.
3) Concerning the values of , we have the following
comments:
For each value of URF, the variation of o, when Pr_th
increases, presents a maximum. The explanation is given by
means of the Fig. 9, 10 and 11 (which represent the dynamic
priority variation for Pr_th=0.25P
max
, Pr_th=0.5P
max
and
Pr_th=0.9P
max
). These figures allow us to evaluate the number
of times where, during the t
s
, the frames of fca have a higher or
lower priority than the threshold (a higher priority means a
lower delay; a lower priority means a bigger delay). Then we
can see that we have for Pr_th=0.5P
max
, the maximum value of
o. For Pr_th=0.25P
max
(Pr_th=0.9P
max
), the number of times
where the dynamic priorities are higher (lower) than the
threshold is much greater than the number of times where the
dynamic priorities are lower (higher) than the threshold. Thus,
we have values of o smaller than with Pr_th=0.5P
max
(in the
case of Pr_th=0.25P
max
with a small value of D ; in the case of
Pr_th=0.9P
max
with a higher value of D ). Obviously, for each
value of Pr_th, o increases with URF (the reason is still the
increase of the network load). Important remark: for
Pr_ths0.15P
max
i.e. low threshold (we have not represented the
results for reasons of limited space), we have the minimal value
for D (1.28ms i.e. a frame of fsc (0.64ms) and then a frame of
fca (0.64ms) that always use the medium before the frames of
fex because the dynamic priority is always higher than Pr_th
during the settling time). Then, of course, o=0.

Figure 9. hp scheme, URF=99.2%, Pr_th=0.25Pmax

Figure 10. hp scheme, URF=99.2%, Pr_th=0.5Pmax

Figure 11. hp scheme, URF=99.2%, Pr_th=0.9Pmax
C. hp+sts scheme
For the hp scheme, we give D and o in table IV and t
s
in
table V. The values are obviously function of URF and Pr_th.
TABLE IV. (HP+STS) SCHEME (QOS): D AND o(MS)
Pr_th
URF 0.9Pmax 0.5Pmax 0.25Pmax
(%)
D

D

D

99.2 2.589 2.138 2.589 2.138 1.28 0.0
89.6 1.856 1.152 1.856 1.152 1.28 0.0
80 1.664 0.768 1.664 0.768 1.28 0.0
70.4 1.28 0.0 1.28 0.0 1.28 0.0
51.2 1.28 0.0 1.28 0.0 1.28 0.0
22.4 1.28 0.0 1.28 0.0 1.28 0.0
30

TABLE V. (HP+STS) SCHEME (QOC): TS (MS)
URF Pr_th
(%) 0.9Pmax 0.5Pmax 0.25Pmax
99.2 103 103 50
89.6 100 100 50
80 98 98 50
70.4 50 50 50
51.2 50 50 50
22.4 50 50 50
We can see important differences with the hp scheme:
1) For URF70.4%
D is now always constant, whatever the Pr_th is (this is for
two reasons: the first one is because of the consequence of the
property URFs70.4%) indicated in the subsection III-A; the
second is the fact that now, at the beginning of the transient
behavior, the dynamic priority is used by the flows fsc and fca
for a duration, at least, equal to 4h. Obviously as D is
constant, o =0.
2) For Pr_th=0.25P
max

We haveD which is constant for all URF values (this
means that, on all the network load conditions, the dynamic
priority is higher than the threshold). The explanation is given
by the exchange temporal diagram in Fig. 12.

Figure 12. (hp+sts) scheme (URF=99.2%, Pr_th=0.25Pmax): time exchanges
3) For Pr_th>0.25P
max

We have the same values of D and o whatever the value of
Pr_th. The explanation is given by the exchange temporal
diagrams of the Fig. 13 and 15 where we consider
URF=99.2%. These diagrams are identical.
4) For URF>70.4%
We note an increase of D and o with URF (the explanation
is given by the Fig. 14 and 15); the delay of the frame fca in the
Fig. 15 is higher than in Fig. 14.
With respect to the hp scheme, all the improvements (which
give best settling time for the process control application) result
from the fact that the dynamic priority P
max
is used a longer
time. In Fig. 16, we have an example of the evolution of the
dynamic priority (we have P
max
during 8h), compare the Fig. 16
with the Fig. 10.

Figure 13. (hp+sts) scheme (URF=99.2%, Pr_th=0.5Pmax): time exchanges

Figure 14. (hp+sts) scheme (URF=80%, Pr_th=0.9Pmax): time exchanges

Figure 15. (hp+sts) scheme (URF=99.2%, Pr_th=0.9Pmax): time exchanges

Figure 16. (hp+sts) scheme (URF=99.2%, Pr_th=0.5Pmax): time exchanges
31

D. hp+dts scheme
We give, as for the previous schemesD and o in Table VI
and t
s
in Table VII. We can see now that we always have the
minimum constant value D (duration of the fsc frame (0.64
ms)+duration of the fca frame (0.64 ms)), then o=0, and the
best settling time (50 ms). This is a consequence of the fact that
the dynamic priority is continuously controlled (by the control
signal u) and that it is higher than the threshold for a time
longer than the t
s
(Fig. 17).
TABLE VI. (HP+DTS) SCHEME (QOS): D AND o(MS)
Pr_th
URF 0.9Pmax 0.5Pmax 0.25Pmax
(%)
D

D

D

99.2 1.28 0.0 1.28 0.0 1.28 0.0
89.6 1.28 0.0 1.28 0.0 1.28 0.0
80 1.28 0.0 1.28 0.0 1.28 0.0
70.4 1.28 0.0 1.28 0.0 1.28 0.0
51.2 1.28 0.0 1.28 0.0 1.28 0.0
22.4 1.28 0.0 1.28 0.0 1.28 0.0

TABLE VII. (HP+DTS) SCHEME (QOC): TS (MS)
URF Pr_th
(%) 0.9Pmax 0.5Pmax 0.25Pmax
99.2 50 50 50
89.6 50 50 50
80 50 50 50
70.4 50 50 50
51.2 50 50 50
22.4 50 50 50


Figure 17. (hp+dts) scheme, URF=99.2%, Pr_th=0.9Pmax
IV. CONCLUSION
This study has presented the interest of a hybrid priority
strategy for the real-time scheduling on the network CAN
where we have two distributed applications with different
needs in terms of transmission urgency in their messages flows
(variable needs for the process control application). In
particular, an important characteristic in an DCS context is the
capacity to implement distributed process control applications
with good performances whatever the network load is. We
have precisely shown that real-time scheduling strategies,
based on hybrid priority schemes, allow the implementation of
a distributed process control application even if the network
load is important. We have considered three hybrid priority
schemes and we have demonstrated the particular interest of a
scheme, call (hp+dts), with a double aspect: dynamic priority
based on a temporal supervision function of the control signal
of the process control application and a reference profile.
REFERENCES
[1] G. J uanole, G. Mouney, and C. Calmettes, "On different priority
schemes for the message scheduling in Networked Control Systems:
Definition and Analysis of a Hybrid Priority Scheme for the Message
Scheduling", in Proc. 16th Mediterranean Conference on Control and
Automation, France, J un. 2008, pp. 1106-1111.
[2] J . Ypez, P. Mart, and J . M. Fuertes, Control loop scheduling
paradigm in distributed control systems, in Proc. IEEE IECON03,
Roanoke, VA, Nov. 2003, pp. 14411446.
[3] G. Walsh and H. Ye, "Scheduling of Networked Control Systems",
Control Systems Magazine, IEEE, Vol. 21, No. 1, Feb. 2001, pp. 57-65.
[4] Bosch GmbH, CAN specification 2.0 (A),
www.semiconductors.bosch.de/pdf/can2spec. pdf, 1991.
[5] M. S. Branicky, V. Liberatore, and S. M. Phillips, Networked control
system co-simulation for co-design, in Proc. Amer. Control Conf.,
2003, pp. 33413346.
[6] W. Qinmu, L. Yesong, and Y. Qin, A scheduling method based on
deadline for CAN-based networked control systems, in Proc. IEEE Int.
Conf. Mechatronics Autom., J un. 2006, pp. 345350.
[7] G. C. Walsh, Y. Hong, and L. G. Bushnell, Stability analysis of
networked control systems, IEEE Trans. Control Syst. Technol., Vol.
10, No. 3, May 2002, pp. 438446.
[8] P. Mart, J . Yez, M. Velasco, R. Vill, and J . M. Fuertes, Managing
quality-of-control in network-based control systems by controller and
message scheduling co-design, IEEE Trans. Ind. Electron., Vol. 51, No.
6, Dec. 2004, pp. 11591167.
[9] Z. Li and M.-Y. Chow, Sampling rate scheduling and digital filter
codesign of networked supervisory control system, in Proc. ISIE, 2007,
pp. 28932898.
[10] A. T. Al-Hammouri, M. S. Branicky, V. Liberatore, and S. M. Phillips,
Decentralized and dynamic bandwidth allocation in networked control
systems, in Proc. 20th Int. Parallel Distrib. Process. Symp., Apr. 2006,
8 pp.
[11] T. Hong, M.-Y. Chow, P. Haaland, D.Wilson, and R.Walker,
Scheduling a life science high-throughput platformunder starvation
constraints using timed transition Petri nets and heuristic search, in
Proc. IEEE Int. Symp. Ind. Electron., J un. 2007, pp. 18931898.
[12] P. Mart and M. Velasco, Toward flexible scheduling of real-time
control tasks: Reviewing basic control models, in Proc. 10th Int. Conf.
Hybrid Syst.: Comput. Control, Vol. 4416, LNCS, 2007, pp. 710713.
[13] N. Pereira, B. Andersson, and E. Tovar, WiDom: A dominance
protocol for wireless mediumaccess, IEEE Trans. Ind. Informat., Vol.
3, No. 2, May 2007, pp. 120130.
[14] M. Grenier and N. Navet, Fine-tuning MAC-level protocols for
optimized real-time QoS, IEEE Trans. Ind. Informat., Vol. 4, No. 1,
Feb. 2008, pp. 615.
[15] M. Ohlin, D. Henriksson, and A. Cervin, TrueTime 1.5-Reference
Manual, Lund Institute of Technology, Sweden, J an. 2007, pp. 1-107.
[16] K. J . Astromand B. Wittenmark, Computer-Controlled Systems-Theory
and design, Third Edition-Prentice Hall, 1997.
[17] K. Zuberi and K. Shin, "Scheduling messages on Controller Area
Network for real-time CIM applications", IEEE Transactions On
Robotics And Automation, Vol. 13, No. 2, 1997, pp. 310-314.
[18] P. Marti, J . Yepez, M. Velasco, R. Villa, and J . Fuertes, "Managing
Quality-of-Control in network-based control systems by controller and
message scheduling co-design", Industrial Electronics, IEEE
Transactions on, Vol. 51, No. 6, Dec. 2004, pp. 1159-1167.
[19] H. Nguyen Xuan, Rseaux de communication et applications de
contrle-commande, Ph.D. dissertation, Univ de Toulouse, INSA,
LAAS, Toulouse, France, Dec 2011.

32

Fall Detection Based On Hidden Markov Model
Viet Q. Truong, Hieu V. Nguyen, Tuan V. Pham
Electronic & Telecomm. Engineering. Dept. Danang University of Technology,
Danang, VietNam
Email: truongquocviet.genie@gmail.com, nvhieu@dut.udn.vn, pvtuan@dut.udn.vn

AbstractRisk of falling accident in the elderly is very high. If
this accident is not detected in time, it will become much more
serious and the victim cant be cured. In this paper, a novel fall
detector has been built based on Hidden Markov Model (HMM).
A feature set is extracted from an ellipse model which is
constructed from a focused object. Besides, the fall detector using
Artificial Neural Network (ANN) [5] has been rebuilt to compare
with the proposed detector. The experimental results obtained
from a self-built database show that: (i) The proposed detector
achieves very high accuracy for clean data set which is 95.96%,
(ii) For noise data sets, the accuracy is 72.41% for Test2 and
76.92% for Test3. These results are similar to the results of the
detector using ANN.
Keywords: fall detection, human behavior understanding,
recognition, neural network, hidden markov model.
I. INTRODUCTION
There is nearly 40-60% of falling accidents causing
dangerous injuries, especially the elderly [2]. Moreover, its
more dangerous if the elderly live in their own houses where it
is difficult to detect a fall. As reported in [3], there are
approximately 28-34% elderly in the community experience at
least one fall every year. Therefore, automatic falling detection
is really necessary to detect a fall in time. It has potential
applicants in medical care as well as health-care or behavior
recognition domain [4]. Many research on detecting falling
accidents, which use sensors, cameras or computer
technologies have been studied so far [6][7].
One of highly appropriated approaches to detect a fall is
based on intelligent video analysis. According to Andersons
report, motion history image and ellipse fitted human body are
efficiently used to detect falls [8]. In image processing
methods, the cameras positions are quite important because
these positions affect the detection accuracy [9]. Furthermore,
real time is always concerned in surveillance system as
proposed in [10] using Template Matching (TM) algorithm. In
order to improve recognition performance, more sophisticated
methods have been proposed based on Artificial Neural
Network [1][5] or Hidden Markov Model.







In this paper, a fall detector using HMM will be designed
and tested on self-built database. Figure 1 shows the proposed
detector include four steps which system processes in
sequence: object segmentation, feature extraction, index
selection and recognition. First step, background subtraction is
used to extract objects silhouette and Mathematical
Morphology method is used to clear this binary image. Next,
five features extracted from binary image sequence are Current
Angle, CMotion, CTheta, Eccentricity and CCentroid. Then,
these features will be clustered based on Euclidean distance to
every codeworks in the codebook has been built previously.
Finally, HMM algorithms are used to detect a fall.
II. OBJ ECT SEGMENTATION
A. Background subtraction
First, object segmentation block is responsible for getting
silhouette images using background subtraction method [10].
From [12], a certain pixel P
i
from current frame I
i
is marked
as foreground if
|P
i
B
i
| >
where B
i
is same position pixel of background frame B
i
and
is a predefined threshold. And, the updated background
B
i+1
is calculated by:
B
i+1
= I
i
+(1 )B
i

where is chosen 0.05 as in [10].
B. Pre-processing
Next, the pre-processing block removes redundant pixels
from images which result from previous block as well as noise.
Mathematical morphology method used for smoothing
silhouette images.
III. FEATURE EXTRACTION
A. Ellipse model
Ellipse model is usually used for tracking object because it
is quite easy to draw an ellipse around object [13]. To build an
ellipse, proposed system determines 3 parameters as follow:
1) Centroid of ellipse
For each binary frame, the centroid coordinate of ellipse
O(Ox, Oy) is determined: Abscissa (Ox) and ordinate (Oy) are
average of the all x coordinates and the all y coordinates of the
white pixels [9].
2) Current angle
This parameter is also called vertical angle which is the
angle between horizontal line and major axis.
Figure 1. Proposed system
Feature
extraction
Output
info.
Object
segmentation
Index
selection
Calculate likelihood
to recognize
Trained
HMMs
Codebook
Input
video
(1)
(2)
33

|
|
.
|

\
|


=
i j
j) .P(i,
2
y
i j
j) .P(i,
2
x
i j
j) x.y.P(i, 2
.arctan
2
1

(3)
where
i, j : position of pixel (i from 1 to width of frame, j from 1
to height of frame).
x =i Ox and y =j Oy (Ox, Oy: position of centroid).
P(i, j) : value of pixel (i, j).
The formula (3) derived tangent of double angle (2) is
used so that vertical angle of object is more stable [9].
3) Major and minor axes of ellipse
Major and minor axes are determined as [1].
B. Feature extraction
1) Current Angle
This feature is calculated as formula (3). Furthermore,
system would well-control if current angle is limited.
In figure 2, the non-cross area is the limited domain of
current angle. System modifies so that current angle belongs to
non-cross area. For example, current angle is 2250 which
system modifies to 450 by subtracting 1800. Figure 2 describes
Theta sequences in different activities.
2) Coefficient of motion (Cmotion)
The objects motion rate is determined by considering 15
successive frames; then, system creates an image that saves the
objects motion history [14].
pixel White pixel Gray
pixel Gray
motion
C
+
= (4)
Figure 3 shows a Motion History Image which is created
from synthesizing a sequence of 15 successive frames. Figure 4
presents the changes of Cmotion in different activities.
3) Deviation of vertical angles (Ctheta)
This parameter is standard deviation of a buffer of 15
successive frames (figure 5 describes the different Ctheta of
objects different activities) [1].
4) Eccentricity
As reported [9], this parameter is really efficient when
detecting a direct fall (figure 6 describes the different
eccentricity of objects different activities).
5) Deviation of centroid (Ccentroid)
Ccentroid is defined as standard deviation of 15 ordinates
of 15 successive centroids. Figure 7 show the different
Ccentroid sequences of objects different activities).
Figure 3. Motion history image
Figure 4. Cmotion in different condition.
Figure 2. Current angle in different condition.
Figure 5. Ctheta in different condition.
Figure 6. Eccentricity in different condition.
Figure 7. Ccentroid in different condition.
34

IV. RECOGNITION WITH HIDDEN MARKOV MODEL
In this proposed system, recognition block based on HMM
is used to detect a fall. A popular HMM is specified by a set of
state S ={s
1
,s
2
,s
3
.,s
N
} and a set of parameters = {,A,B},
where vector contains the prior probabilities, A is the
transition probability matrix and B is the density probabilities
[11]. The following subsections present HMM training and
testing of recognition block.
A. Training
1) Codebook generation
From three extracted features, its necessary to use vector
quantization in order to reduce feature-vectors dimensions.
This quantization generates a codebook which is used to
change a 5-feature vector into certain index number.
Therefore, there are 2 steps to create the codebook:
- Extracting five features from frames of all training
clips included fall and non-fall activities and putting all
training feature vectors into the concatenation of 5-
columns matrix.
- Then, K-means clustering algorithm is used to generate
a codebook which is used as index mapping table for
all built HMM models. K-means clustering
(MacQueen, 1967) is a method commonly used to
automatically partition a data set into k groups which
have k initial arbitrary centers of clusters. This
algorithm iteratively refines centers of clusters to
generate a codebook [16].
Figure 8 shows how to produce a codebook through K-
means clustering algorithm.








2) HMM models







In this work, the obtained symbol sequences are used to
train HMMs to learn the proper model for falls and other
activities. Five states left to right HMM (Figure 9) is used to
build fall model and non-fall model. Two HMMs are trained
based on Baum-Welch algorithm.
B. Testing
Testing progress (also recognition) in proposed system is
illustrated as following flow chart






















After training, proposed system uses two HMMs (fall
model and Non-fall model) in recognition block to detect a fall
by processing two main blocks: clustering and HMM decoding.
1) Clustering
Before HMM decoding, recognition block determines index
by clustering block. At first, clustering block calculates
Euclidean distances from all code-words to feature vector of
current frame. Then, index of current frame is assigned as
index of code-words which has minimum Euclidean distance to
feature vector (figure 11 shows flow chart of clustering block)
[17]. After that, recognition block checks if index-buffer is full
of 20 values. If not, objects action seems to be normal and
clustering block continues getting index of next frame.




If index-buffer is full of 20 indexes, this buffer will be
added new index at the end and removed first index. Then, it
will be decoded by 2 HMMs (fall and non-fall models) in
HMM decoding which is presented following section.
2) HMM decoding and detecting
Joining all
feature vector
K-means
clustering
Codebook
Feature
vector
sequence
Figure 8. Codebook generation.
From Feature
Extraction block
Clustering
Fall detection
Y
HMM decoding
Figure 10. Flowchart algorithm for testing.
Feature vector fromfeature
extraction block
Full index-buffer?
Index
N
Full label-buffer?
Label
No. Fall label>Th?
Normal action
N
N
Y
Y
Add new index
at the end
Add new index at
the end and
remove first index
Find minimum
distance
Index
Euclidean
distance to every
codeword
Figure 11. Clustering block and index selection.
Feature
vector
1 2 3 4 5
Figure 9. Five states left to right HMM.
35

After checking full index-buffer, 20-index buffers are
used for decoding both two built models to find which fall
model or non-fall model is more likelihood
As a result, the output of every 20-frame shift-buffer is 0
(if non-fall model is more likelihood) or 1 (if fall model is
more likelihood). Another buffer (called the label-buffer) is
responsible for storing 20 consecutive labels. When the total
number of 1 in this buffer is greater than predefined
threshold, the falling incident is detected.
V. RECOGNIZE WITH ARTIFICIAL NEURAL NETWORK
According to [5], a two-layer feed forward ANN is
initialized. The input layer consists of either 5 inputs (ANN1)
or 100 inputs (ANN2). And the results of the study in [5]
indicate that ANN2 is more reliable than ANN1 because the
inputs of ANN2 are sequence of feature which are extracted in
a period of 20 consecutive frames. Therefore, in this study,
ANN2 is rebuilt to compare with HMM. The optimization
criterion was chosen to be the minimization of the mean square
error (MSE) derived over the whole training set. The training
set is rst split up into a training subset and a validation subset.
The training of the ANN only stops if the MSE derived from
the validation subset could not be reduced within 5 consequent
training epochs. The Scale Conjugate Gradient (SCG)
algorithm is used for training the neural network with hidden
layers is fixed at 50 units and validation subsets is 20% [5].
VI. EVALUATION
A. Database
In this system, the DUT-HBU database used in our
previous work [20] has been reused. A data set containing 217
video clips as being depicted in Table I is constructed. There
are 107 video clips in which falling actions are recorded, and
the rest of 110 video clips are recorded non-falling ones. While
each falling video contains two actions including walking and
falling, each non-fall video can contain one of actions shown in
Table I.
TABLE I. CLASSIFYING VIDEOS ACCORDING TO ACTIVITIES












To facilitate the comparison and evaluation, the algorithms
are both trained with clean data (Scenario 1) and tested with all
three testing data subset.
- The Test1 subset: Contents and activities in the video
clips for training are very similar to the ones for
testing. In each clip, there is only one object with
stable background. This set has 12 falling videos and
11 non-fall videos.
- The Test2 subset: Includes videos that are similar to
training videos but the brightness of the Test2 videos is
not good or positions of the camera has changes. This
set consists of 15 falling videos and 14 non-fall videos.
- The Test3 subset: They have much changes compared
to training videos. It appears naturally. For example,
part of the object is obscured, background isnt stable,
or there are more than two motion objects in these
video.
B. Evaluation criteria
In this study, for comparing and evaluation, the following
statistical measures are used: Recall (RC), Precision (PR),
Accuracy (Acc).
Recall or Sensitivity is the proportion of Real Positive cases
that are correctly Predicted Positive.
Precision or Confidence denotes the proportion of predicted
Positive cases that are correctly Real Positives. This statistic
measures is calculated as follows:
The accuracy is the proportion of true results (both true
positives and true negatives) in the population. This statistic
measures is calculated as follows:



Where TP, TN, FN, FP are defined as follows:
- True positives (TP): amount of fall actions which are
correctly classied as fall.
- False positives (FP): amount of non-fall actions which
are wrongly considered to be fall.
- False negatives (FN): amount of fall actions which are
wrongly rejected and classied as non-fall actions.
- True negative (TN): amount of non-fall actions which
are correctly classied as non-fall.
C. Experimental results
1) Result of the proposed detector using HMM
Figure 12 shows the result of the proposed detector using
HMM. For clean data set, the result is very high. Recall,
precision and accuracy are 100%, 92.31% and 95.95%
respectively. Statistical results getting decreased dramatically
in the noise data. The accuracy is only 72.41% for Test2 and
76.92% for Test3.
Scenario
1
Scenario
2
Test
1
Test
2
Test
3
ALL
Cross (Fc) 4 18 4 4 10 18
Direct (Fd) 4 19 4 6 9 19
Side (Fs) 7 17 4 5 7 16
Cross (Ncb) 1 4 1 1 1 3
Direct (Ndb) 3 5 1 1 1 3
Side (Nsb) 1 3 1 2 2 5
Cross (Ncc) 1 3 1 2 1 4
Direct (Ndc) 2 4 1 1 1 3
Side (Nsc) 1 4 1 1 1 3
Cross (Ncl) 1 3 1 1 2 4
Direct (Ndl) 3 5 1 1 0 2
Side (Nsl) 1 4 1 1 2 4
Cross (Ncs) 0 2 0 1 2 3
Direct (Nds) 3 6 1 1 1 3
Side (Nss) 1 4 1 1 1 3
0 12 0 0 11 11
33 113 23 29 52 104
Sitting
DATABASE
train test
Other (No)
SUM
Fall
Non
Fall
Bend-
ing
Creep-
ing
Lying
(7)
(5) (6)
36

2) HMM in comparison with ANN for fall detection
Figure 13 shows that: (i) For clean data in Test1, the results
of HMM and ANN are the same, (ii) For noise data, ANN has
higher accuracy than HMM in Test2 but HMM is better in
Test3. In general, the results of ANN and HMM are equivalent.
VII. CONCLUSION
In this system, most of the recognition algorithms are very
high results for the video was made in good environmental
conditions. This indicates that five extracted feature will
provide the reliable information if the object segmentation
works well. For the noise video in Test2 and Test3, Template
Matching algorithm gives the very bad result, the result of
ANN1 (5-features input) is also low [15]. ANN2 with inputs
are features were extracted from 20 consecutive frames and
HMM have better result for noise data because the variation of
parameters over time is expressed. However, for the healthcare
and warning system, these results are still quite low. These
results are quite low due to the object segmentation algorithm
is relatively simple, it cant extract and tracking multiple
objects at the same time. So, when the environmental
conditions are not good or there are many objects in the video,
the results will not be as expected. The results can be greatly
improved by using the better algorithm to extract the object and
develop more sustainable attributes.
ACKNOWLEDGMENT
The authors would like to thank the research group at DUT:
Hoan V. Tran, Phung K. Lai, Khue Tra, Duy H. Le.
REFERENCES
[1] Y T. Ngo; Hieu V.Nguyen; Tuan V. Pham, Study On Fall Detection
Based On Intelligent Video Analysis, IEEE The International
Conference on ATC 2012.
[2] J . Teno, D. Kiel, and V. Mor, Multiple strumbles: a risk factor for falls
in community-dwelling elderly, J . America Geriatrics Society, vol. 38,
no. 12, pp. 1321-1325, 1990.
[3] Chia-Wen Lin, Zhi-Hong Ling Automatic Fall Incident Detection in
Compressed Video for Intelligent Homecare, Computer
Communications and Networks, 2007. ICCCN 2007. Proceedings of
16th International Conference, pp. 1172-1177.
[4] Chieh-Ling Huang, E-LiangChen and Pau-Choo Chun Fall Detection
Using Modular Neural Networks With Back-Projected Optical Flow,
Biomedical Engineering: Applications, Basis and Communications, Vol.
19, No. 6 (2007) 415424.
[5] Y T. Ngo, Phung K. Lai, Tuan V. Pham, A single-camera fall detection
using neural network, United KingdomVietnamAdvance Surveillance
Systems workshop 2012.
[6] Young-Sook Lee; Wan-Young Chung, "Automated abnormal behavior
detection for ubiquitous healthcare application in daytime and
nighttime", Biomedical and Health Informatics (BHI), 2012 IEEE-
EMBS International Conference on , vol., pp.204-207, 5-7 J an. 2012.
[7] Arie Hans Nasution; and S. Emmanuel, Intelligent Video Surveillance
for Monitoring Elderly in Home Environments, International Workshop
on Multimedia Signal Processing, Greece, October 2007.
[8] D. Anderson;, Recognizing Falls fromSilhouettes, Proceedings of the
28th IEEE EMBS Annual International Conference, 2006.
[9] J . Q. Lin, The Behavior Analysis and Detection of Falling, Master
Thesis, Department of Computer Science and Information Engineering,
National Central University, Taiwan, 2004.
[10] N. Chan, A. Patel, A. Chandrasekhar, H. Lee, Fall Detection
(ECE4007L05), Georgia Institue of Technology, April 2009.
[11] Barbara Resch, Hidden Markov Models A Tutorial for the Course
Computational Intelligence..
[12] M.M.Ivor, A. Background subtraction techniques .
[13] Nait-Charif H, McKenna SJ , Activity summarization and fall detection
in a supportive home environment, in Proc of the 17th Int Conf on
Pattern Recognition 4:323326, 2004.
[14] Muhammad J amil Khan and Hafiz Adnan Habib "Video Analytic for
Fall Detection fromShape Features and Motion Gradients", Proceedings
of the World Congress on Engineering and Computer Science 2009 Vol
II, WCECS 2009, October 20-22, 2009, San Francisco, USA.
[15] Khue Tra, Viet Q. Truong, Tuan V. Pham, Threshold-based versus
Artificial Neural Network for fall detection, Danang University
Conference, 2012.
[16] Kiri Wagstaff, Claire Cardie, Department of Computer Science, Cornell
University, Ithaca, NY 14853 USA, Constrained K-means Clustering
with Background Knowledge,2001.
[17] Md. Zia Uddin, Nguyen Duc Thang, J eong Tai Kim, and Tae-Seong
Kim, Human Activity Recognition Using Body J oint-Angle Features and
Hidden Markov Model,2011.
[18] Saleh Alaliyat, Department of Computer Science and Media
Technology, Gjvik University College, Video - based Fall Detection in
Elderlys Houses, 2008.
[19] N. Chan, A. Patel, A. Chandrasekhar, H. Lee, Georgia Institue of
Technology, Fall Detection (ECE4007L05) April 2009.
[20] Hoan V. Tran; Phung K. Lai; Khue Tra; Viet Q. Truong; Duy H. Le;,
DUT-HBU database, Danang University of Technology, 2012.

Figure 12. Result of Hidden Markov Model.
Figure 13. Accuracy of HMM and ANN.
37

Measurement of Biological Concentration Using
Magnetic Agents

Cao Xuan Huu, PhamVan Tuan
Department of Electronic and Telecommunication
Danang University of Technology, UD
Danang, Vietnam
Dang Duc Long, Nguyen Thi Minh Xuan, PhamThi
KimThao
Department of Chemistry
Danang University of Technology, UD
Danang, Vietnam


AbstractDetecting and measuring the concentration of
biological agents are of interest in many biomedical applications.
This paper presents a method for measuring of that
concentration with the aid of magnetic nanoparticles. Here
biological agents (proteins or DNAs) were attached to magnetic
nanoparticles in specific binding ratios. The attached biological
agents interacted with their specific targets in solutions and
allowed those targets concentrations being measured based on
the number of nanoparticles determined via the measurement of
their induced magnetic field. A sensitive magneto-biosensor using
giant magneto-resistance sensing element together with a new
sensing technique has been introduced and employed as a novel
method for a low magnetic field measurement. The technique
demonstrates that ultralow biological concentration can be
detected and measured in a timely fashion without the need of a
special shielding medium.
Keywords-ultralow magnetic field, biosensor, sensing technique,
GMR, magnetic nanoparticle
I. INTRODUCTION
Detection and measurement of biological agents like
proteins or DNAs are of interest in most of biomedical
applications. In many methods of protein detection platforms,
the binding event of a protein to a specific recognition
molecule must be detected with a signal transducer. In ELISAs,
protein microarrays [1], and quantum dot [2] detection
platforms, the readout is based on a fluorescent or colorimetric
signal. However, many biological samples or reagents exhibit
inherent autofluorescence or optical absorption which in turn
becomes a major limiting factor of the measurement. Similarly,
other detection methods using nanowires [3], microcantilevers
[4], carbon nanotubes [5] and electrochemical biosensors [6]
rely on charge-based interactions between the protein or tag of
interest and the sensor, making each system unreliable in
conditions of varying pH and ionic strength.
Over recent years, magnetic detection of biological agents
and activities has gained much attention and be one of the top
interests. The primary attraction of the biomagnetic method is
its ability to assess non-invasively the state of health of various
physiological systems [7]. The strength of biomagnetic signals
is nearly zero (for general proteins or DNAs) or of many orders
of magnitude smaller than that of the earths magnetic field (for
other biological agents). Furthermore, magnetism of biological
agents, if exist some, will not interfere with signal magnetic
detecting transducer mechanism. Therefore, biomagnetism
detection using magnetic field becomes appropriate in
detection and measurement of biological concentration [8].
Recently, magnetic sensing techniques exploit a broad range of
ideas and phenomena from the fields of physics and material
science. For detecting and measuring biological agents like as
DNAs, proteins, etc, a diverse range of biosensor utilizing the
MTJ, GMR, GMI materials has been put into focus [9].
Normally, those sensors work with the static field of the agents
or measure statically the agents with a biasing technique. It is
important to note that geomagnetic field noise (GMN) is of
about 0.1 nT and has a 1/f-like frequency spectrum for a long
spatial range (about kilometers). Thus, to get absolute value of
biomagnetic signals, one must shield the system out of the
effect of the noise. However, the absolute values need a clear
and unchangeable off-set point, of which definitely
unreachable. In addition, either the shielding is inconvenient as
carrying out experiments in narrow shielding space or having a
construction cost for a large enough shielded space of which
few laboratories can effort. To overcome the difficulties, an
advanced method has been explored in which the difference
between the readings of two or more separated sensors has
been taken into measurement. With this technique, it is
possible to measure magnetic field changes smaller than the
GMN with high reliability and without the use of magnetic
shielding. Nonetheless, the method requires at least two sensors
and gives an output just a single reading.
Recently, researchers working in biomedicine and on
biosensor platforms have begun to examine the potential of
magnetic nanoparticles (MNPs), on the scale of 5150 nm, for
biological labeling, drug targeting, hyperthermia, and as MRI
contrast agents. For a recent review of MNPs employed in
these and other biomedical applications the interested reader is
directed to [10]. In the present study, the ultralow field
detection assay is presented of which the application is focused
onto biological agent detection with the aid of the magnetic
nanoparticles. Specifically, in this study the biological agent to
be detected and measured was proteins. The measuring
technique expressed a promising capability of overcoming the
week points mentioned above. There are a plethora of synthesis
and coating strategies, and these are covered in more detail in a
recent review article [11]. Therefore, here we will only touch
on the basic concepts of MNPs for use in biomedicine and then
introduce the method for measuring biological concentration
via the measure of an ultralow magnetic field with the aid of
MNPs.
38

II. BIOLOGICAL ATTACHMENT AND MEASUREMENT
PRINCIPLE
A. Binding Progress of Proteins onto MNPs
In the configuration illustrated in Figure 1, an immobilized
antibody adheres to the immobilization substrate and makes a
connection with biomarker and biotinylated antibody in a
sandwich structure which in turn attached to avidin and
nanoparticle by a specific reaction. For this scheme, antibodies
on the substrate specifically trap the biomarker of interest
from a sample. The substrate is then washed with a solution
containing more biotinylated antibodies, which also bind
specifically to the biomarker. These antibodies are tagged with
biotin molecules, which bind to the protein avidin. When
avidin-coated MNPs are passed over the substrate, they adhere
to the biotinylated antibody. Finished sample is then put into
sample holder and exposed to an external magnetic field
created by two permanent magnets for measurement.


(a) (b) (c) (d) (e)

Figure 1. Binding progress of a biological agent of interest to magnetic
nanoparticle.

The MNPs in attachment with biological agents are
induced by the external magnetic field, for which the
magnitude of the induced field will correlate with the number
of bound nanoparticles [12]. Therefore, the purpose of the
using MNPs is to magnetically mark the biological agents of
interest and to create an output signal. The signal stays in a
range around GMN regime. A sensor will measure the change
in magnetic induced field of nanoparticles in such small
regime using an advanced measuring technique presented in
the next section.
B. Sensor Detecting Principle
Conceptually, the easiest input circuit to consider for
detecting changes in magnetic fields is a pure magnetometer.
However, magnetometers are extremely sensitive to all
magnetic signals in the environment. This may be done
without wondering when measured inside a well-shielded
space [13]. A method which is frequently used is to employ
another sensor to measure ambient fields then taking the
difference between two sensors. However, if the magnetic
signal of interest is weak, then environmental magnetic
interference may prevent measurements. Eventually, a
detecting method has been developed in which the
measurements can be done in a common space, without
shielding, with the help of MNPs.
In order to measure the change in magnetic field created by
sample which is smaller than GMN, here we employ only one
sensor and take the advantage of a scanning system with a
controlled frequency. For this configuration, the sensor is
located at fix position and the sample is moving back and forth
with certain amplitude and frequency which is controlled by a
scanning mechanism. Figure 2 shows the principle of the
measurement using one sensor. In this setup, the measuring
signal after each cycle has been collected and then be taken
the mean value after a number of cycles for the final record.
The ultralow magnetic field of MNPs creates a change in total
magnetic field around sensor head at a frequency of 2.0 Hz.
This specific signal is taken into measurement without
difficulty.


S
Bias field Moving
sample
Scanning


Sensor
head
Magnetizing
field
N


Figure 2. Configuration setup for the measurement of MNPs' induced field.

III. MATERIALS AND MEASUREMENTS
A. Synthesis of Iron Oxide Nanoparticles
MNPs which are iron oxide nanoparticles in this study were
synthesized by coprecipitation methods from a mixture of
FeCl
2
and FeCl
3
(1:2 molar ratio). Typically, 0.86 g
FeCl
2
.4H
2
O and 2.35 g FeCl
3
.6H
2
O were mixed in 40 mL de-
ionized (DI) water and heated to 80
o
C. 5 mL of NH
4
OH was
added by syringe while vigorously stirring the reaction mixture
and then heating continued for 30 min. After that, for the
nanoparticles-coating SiO
2,
1 g of silicon oxide anhydrous in 2
mL DI water was introduced, the temperature was increased to
95
o
C and stirring continued for 30 min. The nanoparticles was
washed three times by DI water and kept at 4
o
C for long time
use.
B. Preparation of Protein-coating MNPs
Protein used in this research was enzyme protease Alcalase
purchased from Sigma Company.
Iron oxide nanoparticles with or without SiO
2
-coating were
dissolved in protein-containing buffer (100 mg/mL, pH=7.5) to
a final concentration of 10.8 mg/mL. This mixture was
incubated overnight with lightly shaking at 4
o
C. After that, the
excess protein was taken out. The iron oxide nanoparticles
were washed three times by PBS (phosphate buffer saline) and
used to determine the amount of proteins which were absorbed
on nanoparticles by Bradford assay.
To protect linking between nanoparticle and protein from
the effect of environment such as solution pH and ambient
temperature, alginate coating was employed. Briefly, total
above nanoparticle-absorbed protein complex was diffused into
39

1 ml sodium alginate solution (2 %). The mixture of
nanoparticle-protein alginate was added dropwise into calcium
chloride solution (0.2M) with continuous shaking at 4
o
C to
form beads. These beads were washed three times with NaCl
(0.9 %) solution and stored in NaCl (0.9 %) solution at 4
o
C for
long time use. The biological specimens have been prepared in
which the proteins were caught with biomarkers through
biotinylated antibody and avidin, and fixed onto MNPs by a
specific bonding mechanism [14]. Surface-functionalized
MNPs are widely employed in various protein or DNA
detection systems as immobilization platform with interest for
purication/incubation processes [15].
C. Magnetic Measurements
Sensor uses Giant Magnetoresistor (GMR) element to
detect an ultralow magnetic field caused by adhering MNPs in
a magnetizing field. GMR sensor requires an uniform bias field
created by an electromagnet couple for a linear detecting region
ranging from 800 A/m to 2400 A/m. MNPs have been induced
by a magnetizing field in vertical direction, which is produced
by two permanent magnets as shown in Figure 2. The
magnetizing field at sample has magnitude of 3.2 kOe, or 254.8
kA/m.
In order to protect unexpected contact with biological
specimen, the sensor head and the specimen were separated by
a cover glass of 120 m thick. The distance from the sample to
the upper sensor head was approximately 0.5 mm. The
specimen was attached onto the slider of the scanning system,
which was moving at 2.0 Hz with amplitude of 5 mm. The
signal detecting was taken for 1 minute, which was of 120
cycles.
IV. RESULTS AND DISCUSSION
Two types of MNPs, with and without SiO
2
, have been
fabricated. Magnetic measurements done by using Vibrating
Sample Magnetometer (VSM), as shown in Figure 3 for their
magnetization curves, identify that MNPs exhibit
superparamagnetic properties.

Figure 4 shows one of transmission electron microscopy
(TEM) images for the obtained MNPs without SiO
2
coating.
One can see that the particle size drops in range of approx. 8
12 nm. The inset in Figure 4 expresses the X-ray pattern of the
sample of which the diffraction peak (311) happened at 2
36
o
(where with arrow mark) shows clear evident for a
theoretical particle size of 8.2 nm when calculated by
Scherrers formula [16].

Figure 5 presents the amount of proteins physically
absorbed on MNPs with or without SiO
2
. Diluted concentration
range of bovine serum albumin (BSA) from 0-0.7 mg in 1 mL
NaOH 1N plus 10.8 mg MNPs were used to make the standard
line in Bradford assay (Figure 5a). The trendline and equations
were generated by excel solver. After incubated with protein
overnight, MNPs were washed and diffused in NaOH 1N.
Dilution range of MNPs in NaOH 1N to 1:2, 1:4, 1:8 were
prepared to determine concentration of proteins absorbed on
MNPs (Figure 5b) based on the equations in Figure 5a.



Figure 5: The amount of proteins physically absorbed on MNPs with or
without SiO2. Conc. stands for concentration; and Abs. for absorption.


Figure 4. TEM image of MNPs. The inset is the X-ray pattern for
MNPs.

Figure 3. Magnetization curves measured at roomtemperature for two
MNP samples
40

Experimental results shown in Figure 5 revealed that while
10.8 mg nanoparticles without SiO
2
coating absorbed 0.8 mg
proteins, 10.8 mg nanoparticles surrounded by SiO
2
shell
absorbed 2 mg proteins. This means SiO
2
shell increased
affinity between MNPs and proteins. The hydrophilic
properties of SiO
2
may explain for this phenomenon. Indeed,
Jin Chang et al. demonstrated that the core-shell MNPs-SiO
2

penetrated into tumor cells more effectively than MNPs
without SiO
2
even the size of MNPs-SiO
2
beads were more
than MNPs size [17]. Similarity, composite of methylene blue
and SiO
2
increased enzyme (horseradish peroxidase) loading
[18]. In our technique, the rate-limiting steps are the binding of
targets to the substrate and antibody-linked nanoparticles to the
target.

The beads containing MNPs and proteins still showed
superparamagnetic properties and catalyst activity of proteins,
as seen in Figure 6. Alginate shell at concentration of 2% did
not show any effect on activities of MNPs and proteins.
Due to the fix position sensor and the sample is moving
relatively to it, the signal induced on sensor head is
independent on sample scanning speed. Therefore, a higher
scanning speed will give a shorter measuring time. However,
the scanning speed of a mechanical system is limited by the
inertia of the moving part. With a constant driving power
supporting to the moving slider, higher scanning speed will
bring shorter scanning amplitude. In this work, maximum
speed is limited to about 5.0 Hz. For an optimum operation,
scanning frequency was chosen by 2.0 Hz. One of the
advantages of this measurement technique was its short
measurement time. Suppose the data will be collected after one
minute, 120 data points are taken into calculation for a mean
value.
Initial results measured for the large amount of MNP
samples (above 0.1 mM) express a good linearity of the
concentration measurement. The MNPs induced field which
was in range of 10
-4
10
-1
A/m can be detected and measured
by only one GMR sensor using a special scanning technique.
The results can be refined in the continuation of research when
the optimization is considered. Experiment also confirms that
the bio-concentration measurement with the aid of magnetic
agents, MNPs, has not been affected by magnetically
interference of biological molecules to be measured. Four
samples have been prepared with the same initial MNP amount
of 0.3 mM. Two first samples were original MNPs, one
without SiO
2
-coating and the other with SiO
2
-coating. The last
two samples were of the first samples with protein attachment.
These four samples were put into measurement at the same
settings and the obtained results were shown in Figure 7.


From Figure 7, it can be concluded that the proteins to be
measured contribute zero biomagnetism effect on MNPs
induced magnetic field. The fluctuation in signal is due to the
reduced signal of MNPs with SiO
2
coating compared to that of
MNPs without coating, as seen clearly in magnetization curves
in Figure 3. This initial result show the error bar at each data
point. It is the uncertainty determined by magnetic field
measurements where somewhat of electromagnetic noise
influenced on the output of the sensor. This implies that the
uncertainty of measurement was in order of about 10
-2
nM.
V. CONCLUSIONS
This paper presents a biophysical method for measuring
protein concentration. Magnetic nanoparticles were of iron
oxide prepared by chemical coprecipitation and then employed
as active agents for the measurement. The concentration was
determined via the amount of MNPs in attachment with
proteins. In order to show the confirmation of the correctness
of the technique, some initially obtained results were presented
which showed that proteins attached to MNPs have no
contribution to magnetism of total induced magnetic field of
MNPs. The detecting and measuring technique expressed the
advantage of a timely fashion and independent behavior of
protein on the measuring signal. The results can be improved in
the continuation of research where the optimization is being
considered. Due to several advantages, the method of
employing the MNPs as active agents for detecting and
measuring biological concentration is becoming one of
promising techniques for an application trend in biomedicine.
ACKNOWLEDGMENT
This research was supported by Ministry of Science and
Technology - Vietnam for the Potential Project implementing
in 2012. The authors are grateful to Professor Derac Son at
Hannam University for suggestions on experimental idea and
valuable discussions.
M
N
P
M
NP-SiO
2
M
NP-protein
M
NP-SiO
2-Protein
4.6
4.7
4.8
4.9
5.0
5.1
5.2
5.3
M
a
g
n
e
t
i
c

f
i
e
l
d

(
A
/
m
)
Sample
Measured data

Figure 7. Four samples with the same initial MNP amount: MNPs,
MNPs with SiO2 coating, MNPs attached with proteins, and MNPs with
SiO2 coating attached with proteins.


Figure 6: The persistence of superparamagnetic properties of MNPs core
in alginate coating
41

REFERENCES
[1] Chan, S.M. et al. Protein microarrays for multiplex analysis of signal
transduction pathways. Nat. Med. 10, 13901396 (2004).
[2] Shingyoji, M. et al. Quantum dotsbased reverse phase protein
microarray. Talanta 67, 472478 (2005).
[3] Zheng, G. et al. Multiplexed electrical detection of cancer markers with
nanowire sensor arrays. Nat. Biotechnol. 23, 12941301 (2005).
[4] J i, H. et al. Microcantilever biosensors based on conformational change
of proteins. Analyst 133, 434443 (2008).
[5] Ghosh, S., Sood, A.K. & Kumar, N. Carbon nanotube flow sensors.
Science 299, 10421044 (2003).
[6] Drummond, T.G., Hill, M.G. & Barton, J .K. Electrochemical DNA
sensors. Nat. Biotechnol. 21 11921199.
[7] L. A. Bradshaw, A. Myers, W. O. Richards, W. Drake, J .P. Wikswo,
Vector projection of biomagnetic Fields, Medical & Biological
Engineering & Computing, 43 (2005) 85.
[8] C.R. Tamanaha, S.P. Mulvaney, J .C. Rife, L.J .Whitman, Magnetic
labeling, detection, and system integration, Biosensors and
Bioelectronics 24 (2008) 113.
[9] J . Llandro, J . J . Palfreyman, A. Ionescu, C. H. W. Barnes, Magnetic
biosensor technologies for medical applications: a review, Med Biol Eng
Comput (2010) 48:977998, DOI 10.1007/s11517-010-0649-3
[10] Berry CC (2009) Progress in functionalization of magnetic nanoparticles
for applications in biomedicine. J Phys D 42: 224,003.
[11] Roca AG, Costo R, Rebolledo AF, Veintemillas-Verdaguer S, Tartaj P,
Gonzalez-Carreno T, Morales MP, Serna CJ (2009) Progress in the
preparation of magnetic nanoparticles for applications in biomedicine. J
Phys D 42:224,002.
[12] Gaster, R. S. et al. (2009) Matrix-insensitive protein assays push the
limits of biosensors in medicine, Nature Medicine,
doi:10.1038/nm2032.
[13] KarstenSternickel, and Alex I Braginski, Biomagnetism using SQUIDs:
status and perspectives, Supercond. Sci. Technol. 19 (2006) S160S171.
[14] Ilia Fishbein and Robert J . Levy, The matrix neutralized, Nature -
Analytical Chemistry, 461 (2009) 890.
[15] Arben Merkoci, Nanoparticles-based strategies for DNA, protein and
cell sensors, Biosensors and Bioelectronics 26 (2010) 11641177.
[16] Patterson, A. L,, The Scherrer Formula for X-Ray Particle Size
Determination, Physical Review, 56 (1939) 978-982.
[17] J in Zhang et al. (2006), Design nanoparticles as Drug carriers for
cancer therapy, Cancer genomics and proteomics review, 3, 147-158.
[18] Hui Yao, Nan Li, Shu Xu, J in-Zhong Xu, J un-J ie Zhu , Hong-Yuan
Chen (2005), Electrochemical study of a new methylene
blue/siliconoxide nanocomposition mediator and its application for
stable biosensor of hydrogen peroxide, Biosensors and Bioelectronics,
21, 372-377.

42

A Novel Approach to Protect Intellectual Property
Core of FPGA-Based Partially Reconfigurable
Systems
Tran Thanh
1
, Tran Hoang Vu
1
, PhamNgoc Nam
1
, Nguyen Van Cuong
2

1
School of Electronics and Telecommunications, Ha Noi University of Science and Technology, Ha Noi, Viet Nam.
Email:{nb10069, nb10002, pnnam-fet}@mail.hut.edu.vn
2
Faculty of Electronics and Telecommunications, Da Nang University of Technology, Da Nang, Viet Nam.
Email: nvcuong2000@gmail.com


Abstract Reuse of designed hardware modules, also called
Intellectual Property (IP) cores is an important part of the effort
to design and implement complex systems. An IP core design can
take months to develop, and be stolen in seconds. Therefore, IP
protection of hardware designs is the most important
requirement for many Field Programmable Gate Array (FPGA)
IP vendors. There have been already many proposals to
overcome this problem using symmetric encryption techniques
with a cryptographic key to be stored in a non-volatile memory
located on FPGA or in a battery-backed RAM as done in some of
the current FPGAs. This paper proposes a novel approach for
protecting IP cores on embedded systems based on FPGA by
building an AES decryptor in the partially reconfiguration
region. The AES decryptor can be released from the partial
reconfigurable region when it is not in use and then be replaced
with another useful application. Experimental results and
analysis show that the proposed technique reduce resource
overhead, increase the flexibility of the system and it is robust
against attacks.
Keyword- reconfigurable, bitstream security, embedded system,
partial reconfiguration.
I. INTRODUCTION
The advance in semiconductor processing technology has
led to a rapid increase in the design of complex integrated
circuit. Design methodology based on the reuse has prevailed
in the IC design field. Reuse of Intellectual property (IP) can
reduce product costs, shorten design cycles and decrease
design risk. Nowadays, reusable IP cores have become the
basic units of system design. IP cores are widely exchanged
and are being sold as an independent design. Most of reusable
IP cores can require a lot of time and effort to be implemented
and verified by IP providers, but these IP cores can be easily
copied or modified. Therefore, IP providers apply security
solutions to against the stealing and tampering their products.
In addition, users expect that purchased IP core is correct and
legal. How to protect IP core reuse can effectively has become
a serious problem and therefore an increasing number of
attention has been paid.
A. Security risks
Design outsourcing of FPGA-based systems to third-party
vendors and open source communities gives rise to network of
trust relationships between parties and toolsets. This multi-
party arrangement leads to increasing security risks [1].
Individual IP cores are typically verified, and then certified as
secure. However, controlling the consistency of a complete
design flow, from the system specification through IP core
integration and deployment of the final product, becomes
more challenging when a PR flow is considered. PR
introduces a risk of replacing part of the design with malicious
(Trojan-horse) functionality and thus a risk of compromising
the entire system [2]. During self-reconfiguration, illegal
connections (called covert channels) can potentially be set up,
allowing for unwanted ad-hoc interactions between IP cores.
This introduces the risk to IP secrets, e.g. encryption key
extraction, algorithm theft etc. IP protection methods proposed
to date can be classified into two main groups: low-cost
security and high-end security.
Low cost methods target mainly massive scale (industrial)
IP theft. This approach may not prevent the efforts of
motivated peers to obtain FPGA configuration details and to
publish results [3]. Security-by-obscurity is considered an
adequate hindrance for resource-limited and time-limited
attackers.
High-end protection includes methods aimed at providing
security of the design IP against all but sophisticated attackers
who have unlimited resources. IP protection provided by
FPGA design software [4, 5] supports IP core licensing and
Digital Rights Management (DRM), down to the design
netlist-level. The software-level protection effort is augmented
by configuration bitstream encryption, being a de-facto
industry standard provided by FPGA-specific tools [6].
However, this approach is limited and does not support PR.
When PR is used the attacker could, under certain conditions,
intercept the plain-text content of the IP core by accessing it,
using the ICAP for example.
B. AES algorithm
The Rijndael algorithm was adopted as the Advanced
Encryption Standard (AES) in 2001. The AES algorithm is a
symmetric block cipher that encrypts and decrypts electronic
data. Symmetric-key means the key used for encryption is the
same as the one used for decryption. Block cipher means the
data is processed in blocks. Symmetric-key block-cipher
encryption algorithms are widely used in many industries to
provide data security because of its high security protection
43

and efficiency, ease of implementation, and fast data-
processing speed.
AES was announced by the U.S. National Institute of
Standards and Technology (NIST) [7]. AES has a fixed block
size of 128 bits and a key size of 128, 192, or 256 bits. So far,
the AES encryption algorithm is still considered safe, and in
practice it means that data encrypted with the AES algorithm
has not been broken. The key length of AES algorithm is
enough to protect information classified as secret. For the
more safety information the key length of 192 bits or 256 bits
will be used.
The choice of key storage is the second most important
design consideration. A key is stored in either volatile or non-
volatile storage, depending on the chip vendor. Once power is
off for volatile storage, the key is lost unless there is an
external battery connected to the chip as a backup power
supply. On the other hand, non-volatile key storage hands the
designer greater flexibility.
C. Partial Reconfiguration
Partial Reconfiguration (PR) technology, offered by some
FPGA vendors, provides full access to the FPGA
configuration memory during system runtime, using for
example an Internal Configuration Access Port (ICAP). This
enables modification of the underlying hardware configuration
during runtime, e.g. insertion of IP cores, without restarting
the system.

Figure 1. The model of partial reconfigurable system
Partial Reconfiguration is the ability to dynamically
modify blocks of logic by downloading partial bit files (IP
cores) while the remaining logic continues to operate without
interruption. In this system, Figure 1:
Reconfigurable Partition (RP) is design hierarchy
instance marked by the user for reconfiguration.
Reconfigurable Module (RM) is a portion of the logical
design that occupies the RP and each RP will have
multiple RMs.
Static logic is all logic in the design that is not
reconfigurable.
The same framework and embedded platform in [8], in this
paper, we presented a new method by building a AES decoder
into partial reconfigurable region. This method is really useful
for low-cost FPGAs.
The structure of the paper is as follows: Section II reviews
the related work. Section III describes the proposed design
flow to IP core protection. Section IV details the system
implementation and evaluation. Section V concludes the
paper.
II. RELATED WORK
This section reviews current research in security of the
partial reconfigurable system and focuses on two main factors:
system integrity and design privacy protection. System
integrity protection is mainly used in systems at which high
security levels are required. Hadzic [9] describes the threat of
hardware viruses in FPGAs.
Kean [10] and Bossuet [11] highlighted vulnerability of
volatile FPGAs to IP piracy and reverse engineering, and
proposed bitstream encryption as a countermeasure. Drimer
[12] more recently examined a wide range of attack
mechanisms and countermeasures. A more general review of
security challenges facing embedded systems can be found in
[13]. Adi [14] proposed a system based on the use of public
and secret-key cryptography. In [15], Yuan summarized
current IP protection goals and proposed various solutions.
Most of the proposed approaches rely on the assumption
that a secret can be stored safely in the FPGA fabric. The
secret is usually an encryption key, and its storage is
permanent, one time programmable, or volatile.
Ideally, IP-blocks should be tested, and verified by an
external Trusted Authority. In reality, IP cores are typically
obtained through in-house design reuse, from 3
rd
party IP
vendors, or using automatic IP core generation tools.
Thompson [16] challenged us to question the security of tools.
III. PROPOSED DESIGN FLOW
We aim to implement the static system and the
reconfigurable modules in completely isolated design steps.
Then, all modules can be developed without the existence of
the final static system. Moreover, modules are encapsulated
and might be ported among different designs without any
additional synthesis or place&route step. This is possible by
statically partitioning the routing resources of the FPGA into
the resources used for the top level communication and into
resources that are used for the implementation of the static
system or the reconfigurable modules.
A. Implementing the Static System
For defining a reconfigurable partition within the static
system and for constraining the interface wires that are
allowed to cross the border for connecting a partial module,
the user has to floorplan the system. This is supported by a
comfortable GUI. For a selected reconfigurable region, our
tools will generate constraints for the physical implementation
of the static system that ensure that no logic and routing
resources will be used within the selected reconfigurable
partition.
44

The Xilinx vendor tools provide such prohibit constraints
only for logic resources. We solved the remaining problem of
constraining routing resources by generating blockers that
occupy within a definable region all wire resources that will
then not be used by the Xilinx router for implementing the
static system. For each interface signal, we leave a hole in the
blocker such that the corresponding dummy sink or source can
be connected.
As shown in Figure 2, the static system includes an
embedded processor, a reconfigurable controller with an
integrated ICAP core, UART interface, Compaq Flash card
connection, etc.

Figure 2. Block diagramof a FPGA systemcontaining reconfigurable AES
IP core controlling access to FPGA reconfiguration region via the ICAP port.
B. Implementing the Reconfigurable Modules
The implementation of the reconfigurable modules follows
the same idea of constraining than the one that has been
applied to the static system. But this time, we constrain a
particular module into an encapsulated region having the size
and containing the same resource layout than the reserved
reconfigurable partition of the static system.
Figure 2 illustrates the partial configurable region, we have
built the configurable partitions which consist of a partition to
load the AES core to perform the decoding when IP core load
to the reconfigurable partition. When updating is not required,
AES core can be released and another application can be
replaced here.
D. FPGA system multi-party environment
To implement the proposed model, we consider a multi-
party environment of IP core design and protection. The
parties and the implementation process are shown in Figure 3.
System Integrator (SysInt) designs the FPGA system and
provides it to the user. SysInt has physical access to the
product and can issue a product upgrade in the field. A typical
system consists of custom elements and multiple third-party
(IPVend) IP cores. IP cores can be distributed in various
formats: HDL sources, netlists or FPGA device-specific
partial bitstreams, depending on the level of trust between
SysInt and IPVend.
IP Vendor (IPVend) provides reusable components (IP
cores), and related data sheet, or may design an IP block to
meet a provided subsystem specification. IPVend is typically
not directly involved in the system level FPGA design
process. IPVend wishes to protect its own design secrets.
Trusted Authority (TAut) is an authorization and/or
certification center. TAut confirms the key generation process
during initial system start-up and certifies the resulting key
material. TAut is assumed to be trustworthy by all parties and
is not involved in the system development process.
As System Integrators (SysInt) need IP core for his system,
he sends requirements for an IP core to IP Vendor (IPVend).
The IPVend sends a request of encryption key to encrypt his
IP before sending it to SysInt. After receiving requests from
IPVend, SysInt will send back an encryption key of your
system, IPVend will send this key to Taut to verify whether
this key is the the key of SystInt system or not. If it is true,
Taut will send back a confirmation and IPVend will use this
key to encrypt your IP core and sent to SysInt. Finally, SysInt
integrated IP core into his system and saved in the CF card.

Figure 3. Flowchart for IP core installation
IV. IMPLEMENTATION AND EVALUATION
A. Prototype system
To test the proposed method, we have built a prototype
system(Figure 4) which consists of a reconfigurable embedded
platform based on Xilinx Virtex-6 XC6VLX240T-1FFG1156
FPGA ML605 board, a laptop and a CF card (Figure 4). The
ML605 board and the laptop are connected via a TCP/IP
connection.

Figure 4. Prototype system
45

On the Virtex-6 XC6VLX240T chip, we embedded a
MicroBlaze soft-core microprocessor using Xilinx Embedded
Development Kit (EDK) ver. 14.1 software. EDK toolset
allows designers to easily create platforms based on either
MicroBlaze. EDK offers a variety of peripherals (UARTs,
counter, Ethernet, memory controller, general-purpose I/O and
so on) and a one-connection solution based on the structure of
the IBM CoreConnect bus [17].
In this prototype system, we build AES-256 Decryptor into
a partial bitstream file saved in the CF card.
B. Resource overhead and performance of system
Table 1 shows the resources used for the modules of
hardware system based on Virtex-6 ML605 Xilinx board.
TABLE I. HARDWARE UTILIZATION OF AES DECRYPTOR CORE.
AES-256 Decryptor LUTs Registers BRAM
% of Virtex-6 (LX240T) 4 0.5 5

The proposed test program measures the execution time
consumed by the consecutive decryption of data blocks until
105KB of the partial bitstream file are processed. The data to
cipher is stored in Block RAM and configured on the fly.
As shown in Table 2, with the measured throughput from
given partial bitstream file, it takes about few milliseconds to
update the partial bitstreams for Virtex-6 LXT FPGA. This
result is acceptable for a system update.
TABLE II. PERFORMANCE OUR SECURITY SYSTEMS..
Virtex-6 (LX240T) Time Throughput
AES-256 Decryption 2.5ms 350Mbps

We also consider the resource overhead when
implementing with the Xilinx Spartan-6 LX45 and Virtex-5
LX50T FPGA. Table 3 shows that the resource overhead is
significant and the release of AES core when not being in use
provides a substantial practical benefits.
TABLE III. HARDWARE UTILIZATION OF AES CORE OF SPARTAN-6 AND
VIRTEX-5.
AES-256 Decryptor LUTs Registers BRAM
% of Virtex-5 (LX50T) 17.5 15 25
% of Spartan-6 (LX45) 8.2 4 7
V. CONCLUSIONS
In this paper, with the proposed method, all the partial
bitstream files to ensure always be encrypted when stored
outside. The flexibility of the system is that the AES core is
implemented in the partial reconfigurable region. When the
partial modules installed or updated completely, the
reconfigurable partition of AES core may be released for
another application to execute.
The proposed method for PR systems offers improved
security and convenient reuse of external third-party IP cores,
protect the IP Digital Rights Management as well as system
security and integrity in embedded design environment with
multi-party participants.
Future work will optimize the performance of the
reconfiguration flow via Ethernet interface to remote updating.
REFERENCES
[1] K. Thompson, Reflections on trusting trust, Communications of the
ACM, pp. 761-763. USA, 1984.
[2] S. T. King, J . Tucek, A. Cozzie, C. Grier, W. J iang, and Y. Zhou,
Designing and implementing malicious hardware, in Proc. First
USENIX Workshop on Large-Scale Exploits and Emergent Threats
(LEET), 2008.
[3] J . Note, E. Rannaud, Fromthe bitstream to the netlist, In Proc. of the
16th international ACM/SIGDA symposiumon FPGA, ACM, 2008,
pp. 264-264.
[4] H. Walker, Xilinx speeds HDL simulation with SecureIP and FAST
Simulation Mode Models,
http://www.fpgajournal.com/articles_2008/20080610_xilinx.htm, 2008.
[5] M. Miller, Synplicity introduces secure IP flow for FPGAs, signs
ARM, Tensilica as partners ,
http://www.edn.com/index.asp?layout=article&articleid=CA6551580
Xilinx encryption, 2008.
[6] A. Lesea, IP Security in FPGAs, unpublished, www.xilinx.com,
2007.
[7] FIPS-46: Data encryption standard, NBS, US Department of
Commerce, J anuary 1977.
[8] Tran Thanh, P. N. Nam, T. H. Vu, Ng. V. Cuong, A Framework for
Secure Remote Updating of Bitstream on Runtime Reconfigurable
Embedded Platforms, In Proceeding of the fourth International
Conference on Communications and Electronics (ICCE 2012), pp. 471-
476. Hue, Vietnam, 2012.
[9] Hadzic, I., Udani, S. & Smith, J . M., FPGA viruses, FPL '99:
Proceedings of the 9th International Workshop on Field-Programmable
Logic and Applications, Springer-Ver-lag, 1999, 291-300.
[10] Kean T., Secure configuration of Field Programmable Gate Arrays,
FPL '01: Proceedings of the 11th International Conference on Field-
Programmable Logic and Applications, Springer-Verlag, 2001, 142-
151.
[11] Bossuet L., Dynamically Configurable Security for SRAM FPGA
Bitstreams. In International J ournal of Embedded Systems, IJ ES,
Inderscience Publishers, Vol. 2, Nos. 1/2, pp 73-85, 2006
[12] Saar Drimer, Volatile FPGA design security a survey, J ournal of
Engineering, Computer Laboratory, University of Cambridge,
Version 0.96, April 17, 2008.
[13] Ravi, S., Raghunathan, A., Kocher, P. & Hattangady, S., Security in
embedded systems: Design challenges, Trans. on Embedded
Computing Sys., ACM Press, 2004.
[14] Adi, W., Ernst, R., Soudan, B. & Hanoun, A., VLSI design exchange
with intellectual property protection in FPGA environment using both
secret and public-key cryptography, Emerging VLSI Technologies and
Architectures, 2006.
[15] Yuan L., Qu G., Ghout L. & Bouridane A., VLSI Design IP
protection: solutions, new challenges, and opportunities, NASA/ESA
Conference on Adaptive Hardware and Systems, IEEE Computer
Society, 2006.
[16] Mark McLean, J . M., FPGA-based single chip crypto-graphic
solution, Military Embedded Systems, 2007.
[17] Xilinx Inc., MicroBlaze Processor Reference Guide, UG081 (v9.0),
http://www.xilinx.com/support/documentation/sw_manuals/mb_ref_gui
de.pdf
46

A DAG - SVM BASED METHOD FOR FACE RECOGNITION USING PCA
Tran Thi Minh Hanh
Danang University of Technology, Danang, Vietnam
Email: ttmhanh@dut.udn.vn


Abstract PCA (Principal Component Analysis) is a well-known
feature extraction and data representation technique widely used in
the areas of pattern recognition, computer vision and signal
processing. In this paper, PCA is implemented for face feature
extraction. These features are then used to train and test with support
vector machine (SVM) classifiers, which is based on decision-tree
pair-wise classification called Directed Acyclic Graph (DAG). A
comparative study with Principal Component Analysis (PCA) and k
Nearest Neighbor techniques is also presented. The result on two
well-known databases J AFFE and ORL show that this method could
achieve high classification rate.
.Keywords-Face recognition, Principle Component Analysis, Support Vector
Machine, Eigenfaces, Directed Acyclic Graph.
I. INTRODUCTION
Face recognition has emerged as an active research
area in the field of computer vision and pattern recognition. It
can be used in a wide range of applications such as identity
authentication, access control, and surveillance [8]. Most of
the face recognition system has two main processing steps.
The first step is feature extraction, which is performed to
provide effective information that is useful for distinguishing
between faces of different persons and stable with respect to
the geometrical and photometrical variations. In the second
step, classification, the extracted feature vector of the input
face is then matched against those of enrolled faces in the
database.

Figure 1. Face recognition systemwith PCA and DAG-SVM
Many methods have been proposed for feature
extraction, such as geometric feature-based [7], template
matching based [8], and graph matching method [3]. In
geometric feature-based methods, local features such as the
eyes, nose, and mouth are first extracted and their locations
and local statistics (geometric and/or appearance) are fed into
a structural classifier. Facial feature detection and
measurement techniques developed have not been reliable
enough to adapt the need. In contrast, template matching based
method generally is operated directly on an image-based
representation (i.e. pixel intensity array). Because the
detection and measurement of face features are not required,
this class of methods has been more practical and reliable as
compare to geometric feature-based methods.
A well-know face feature extraction using template
matching is that based on the eigenface representation
Matthew Turk and Alex Pentland [5]. Nevertheless, PCA
could not capture even the simplest invariance unless this
information is explicitly provided in the training data. To deal
with this problem, some researchers proposed other
approaches: elastic bunch graph matching Wiskott et al. [3];
Bartlett et al. [6] proposed using independent component
analysis (ICA) for face representation and reported that it
performed better than PCA; Kernel PCA is suggested by
Ming-HsuanYang [9] for face feature extraction and
recognition. However, the performance costs of ICA and
Kernel PCA are higher than PCA. In this paper, PCA is used
as the feature extraction.
There have been a few of classification methods such as k-
Nearest Neighbors (k-NN), Fuzzy C-Means (FCM) and
Artificial Neural Network (ANN). Recently, most of
researchers have focused on the classification method, Support
Vector Machine (SVM). In 1995, Vapnik and Cortes [1]
presented the foundations for SVM. Since then, it has become
the prominent method to solve problems in pattern
classification and regression. The basic idea behind SVM is,
with a set of points belonging to two classes, finding the
optimal hyperplane that separates the largest possible fraction
of points of the same class on the same side, while
maximizing the distance from either class to the hyperplane.
For linearly non-separable data, SVM maps the input to a
higher dimensional feature space where a linear hyperplane
can be found. Although there is no warranty that a linear
solution will always exist in the higher dimensional space, it is
able to find effective solutions in practice.
To deal with the face classification, many researchers [2,
4] have applied SVM in their studies and stated that the
experiment results are very positive. In this paper, an
algorithm using SVMs for face recognition is presented
(Fig.1). The SVMs based recognition is compared with the
most popular eigenface approach, which use the k -Nearest
Neighbor classification.
The remaining sections of the paper are organized as
follows. Section 2 and 3 gives details of the first module (i.e,
feature extraction) and the second module (i.e, face
classification base on SVM), respectively. Section 4 shows
implementation and experiments. Finally, the conclusion is
mentioned in Section 5.

47

II. FACE FEATURE EXTRACTION
A. Normalization
The first module is to normalize the input image. The goal
of the normalization module is to transform the facial image
into a standard format that removes or attenuates variations that
can affect recognition performance.

Figure 2. Input image (ORL database) before (a) and after (b)
normalization; global vector (c)
B. Global vector
The face images then are represented as global vectors
x
global
= (x
1
, x
2
,, x
m*n
)
T
, where m*n and x
i
is image
resolution and gray value of each pixel, respectively.
C. Feature extraction
One of well-known methods for extracting facial feature is
PCA as mentioned above. It was first applied in face
classification by Sirovich and Kirby and then Matthew Turk
and Alex Pentland [5]. Now it has become the standard
method in this field.
The input is a dataset D = { } N i x
i
,..., 2 , 1 ;
) (
=
Step 1: Compute the mean of data

=
=
N
i
i
x x
1
) (
(1)

Step 2: Compute the covariance matrix of data

=
=
N
i
i T i
x x x x
N
C
1
) ( ) (
) ( ) (
1
(2)

Step 3: Compute the eigenvectors v
i
and eigenvalues
i
of
matrix C
Step 4: Order the eigenvectors according to their
corresponding eigenvalues from high to low. Keep only K
eigenvectors associated with K largest eigenvalues
Finally, each of the centered training images (i.e, x x
i

) (
)
is projected into the eigenspace } ,..., , {
2 1 k
v v v V = . To project
an image into the eigenspace, calculate the dot product of the
image with each of the ordered eigenvectors.

) (
) ( 1 '
x x V x
i
=

(3)

The new vector } ,..., , {
'
2
'
1
' '
k x x x x = of the projected image
will contain as many values as eigenvectors.
III. CLASSIFICATION USING SUPPORT VECTOR MACHINE
In this section, the basic theory of the SVM is described
first, and then present the multi-class strategy for SVMs to
solve the multiclass recognition.
A. Two-class classification
The goal of SVM classifiers is to find a hyperplane that
separates the largest fraction of a labeled data set. The most
important requirement, which the classifiers must have, is that
they have to maximize the margin (the distance between the
hyperplane and the nearest data point of each class) (shown in
Fig. 3). This linear classifier is termed the optimal separating
hyperplane (OSH).

Figure3. A Typical SVM classifier
Consider the problem of separating the set of training
vectors belonging to two separate classes, {x
i
,y
i
}, i=1,,l,
where x
i n
R e ,
i
y } 1 , 1 { + e . Now, all hyperplanes in
n
R are
parameterized by a vector w, and a constant b, expressed in
the equation w.x +b =0. The set of vectors is said to be
optimally separated by the hyperplane if it is separated without
error and the margin is maximal. The separating hyperplane in
canonical form [10] must satisfy the following constraints,
y
i
[(w.x
i
)+b]
i
>1 , i=1,,l ( 4)
where
i
are nonnegative slack variables which measure the
degree of misclassification of the data.
Hence the hyperplane that optimally separates the data
(with a large margin and a small error penalty) is the one that
minimizes
2
1
||w||
2
+C

=
l
i
i
1
(where C is a given value) under
the constrains of equation (4). The solution to the optimization
problem is given by using Lagrange functional:
L(w,b, ) =
2
1
||w||
2
+C

=
l
i
i
1
-

=
l
i
i
i
y
1
[ {
w.x
i
) +b]-1+
i
}
(5 )
Where
i
are the Lagrange multipliers. The problem in Eq.5
is transformed to its dual problem
)} , , ( min { max
, ,


b w L
b w

Solving the dual problem,

= = =
=
l
i
j i j
l
j
i
j i
l
i
i
x x y y
1 1 1
2
1
min arg


(6 )
With constraints, C
i
s s 0 , i=1,,l
48

l
i
i
i
y
1
0

The key advantage of a linear penalty function is that the slack
variables vanish from the dual problem, with the
constant C appearing only as an additional constraint on the
Lagrange multipliers.
Solving Eq. (6) with constrains determines the Lagrange
multipliers, and the OSH is given by,

l
i
i i
i
x y w
1

and
j j
x w y b .

(7)
Where
j
x is one of the support vectors (the data that satisfy
the equality in (4)) , and
j
y =1 or -1. The obtained hyperplane
is called soft margin hyperplane.
For the new data point x, the classification is then
f(x) =sign(w. x +b) (8)
If f(x) =+1 then x belongs to positive class, if f(x)=-1 then
x belongs to negative class
B. Mapping the Inputs to other dimensions using Kernel
In most of real applications, the data could not be linearly
classified. To deal with this problem, data is transformed into
a higher dimensional feature space and assume that data in this
space can be linearly classified [1]. In fact, determining the
optimal hyperplane is a constrained optimization problem and
can be solved using quadratic programming (QP) techniques.
Let the mapping function be ) (x , to set up new optimization
problem, replacing all occurrences of x with ) (x . The
minimization problem (recall eq. 4) would still be




l
i
j i
l
j
j i
j i
l
i
i
x x y y
1 1 1
)) ( ). ( (
2
1
min arg


And w (in Eq.7) would be

l
i
i i
i
x y w
1
) (

Equation 8 would be

l
i
i i
i
b x x y sign b x w sign x f
1
) )) ( ). ( ( ( ) ) ( . ( ) (

That is, if we knew the formula (called a Kernel) for the dot
product in the higher dimensional feature space. The classifier
would be:

l
i
i i
i
b x x K y sign x f
1
) ) , ( ( ) (

Some common Kernels include Polynomial, Gaussian Radial
basic function (RBF) and Hyperbolic tangent. This paper is
only implemented with RBF- SVM classifier:
2
|| || exp( ) , (
j i j i
x x x x K for 0
C. Multiclass SVM
The earliest used implementation for SVM multiclass
classification is probably the one-against-all method. It
constructs k SVM models where k is the number of classes.
The ith SVM is trained with all of the examples in the ith class
with positive labels, and all other examples with negative
labels.


Figure 4. One-against-all (left) and one-against-one (right) multiclass
SVM []
Another major method is called the one-against-one
method, first used of on SVM was in [12]. This method
constructs classifiers where each one is trained on data from
two classes. There are different methods for doing the future
testing after all k(k-1)/2 classifiers are constructed, voting
strategy suggested in [11] is one example: if x is in the ith
class, then the vote for the ith class is added by one.
Otherwise, the jth is increased by one. Then predicting x is in
the class with the largest vote. In case that two classes have
identical votes, thought it may not be a good strategy. Platt,
Cristianini, and J. Shawe-Taylor [13] proposed decision-tree-
based pair-wise classification called Directed Acyclic Graph
(DAG). Its training phase is the same as the one-against-one
method mentioned above by solving binary SVMs. However,
in the testing phase, it uses a rooted binary directed acyclic
graph which has internal nodes and leaves. Each node is a
binary SVM of ith and jth classes.















Figure 5. Testing phase with three classes.
Fig. 5 shows the decision tree for three classes. In the
figure, i shows that associated class is not i. As the top-level
classification, we can choose any pair of classes. If x does not
belong to Class 2, thus it belongs to either Class 1 or 3 and the
next classification pair is Classes 1 and 3. In this method,
2
3 2 1

13
()
1
3

12
()

23
()
Class 2
Class 3
Class 1
49

testing time is less than the one-against-one method. DAG
method is chosen in this paper for multiclass SVM.
IV. DATABASE AND EXPERIMENTS
In this work, the experiment is performed on two face
databases. One is on the Cambridge ORL face database,
which contains 40 distinct persons. Each person has ten
different images, taken at different times. There are
variations in facial expressions such as open/ closed eyes,
smiling/ non-smiling, and facial details such as glasses/ no
glasses. All the images were taken against a dark
homogeneous background with the subjects in an up-right,
frontal position, with tolerance for some side movements.
There are also some variations in scale. The second is on
the JAFFE database of 200 images of 10 individuals,
containing six basic facial expressions (happiness, sadness,
surprise, anger, disgust and fear) and one neutral. The
performance of the SVMs based face recognition is
compared with other algorithm to show its success.

(a)

(b)

Figure 6. Some images from (a) ORL [15] and (b) JAFFE [14]
databases
A. Experiments on ORL database
In this face recognition experiment on the ORL
database, 200 samples (5 for each individual) are
randomly selected as the training set, from which the
eigenfaces and the support vector machines (SVMs) are
calculated. The remaining 200 samples are used as the test
set. In this experiment, the number of features will be
changed up to 200. And then we take a look at the result
corresponding to the number of features. In general, many
feature vectors needs more computing time but gives more
accuracy (Fig. 8). When the number of feature vectors is
chanced, input vectors of classifier are also changed, since
output passed through PCA is directly connected to the
input of the classifier.
Two classifications are implemented methods to conduct
experiments on this database to solve face recognition
problem:
1) k Nearest Neighbor (k-NN): uses distance metric L2 for
classification; with the value of k set to 1
2) Support Vector Machine (SVM): Applies nonlinear
SVM with the Kernel function RBF with the value of C (soft
margin parameter) set to 2

, 2

, 2

,, 2

,, 2

and the
kernel parameter gamma set to 2

, 2

, 2

, ,2

.
The reported results were obtained with Cross-Validation
analysis with 2-folds on the dataset.

Figure 7. Comparison of accuracy with the changes of C and gamma of the
SVM algorithmon ORL database.

Figure 7 shows the recognition rate when changing the soft
margin parameter C and the Kernel parameter . When
=2
15
and =2
11
, the recognition rate reaches a
maximum value. After choosing parameters, the classifiers
obtained from training are used to carry out recognition
on the images in the testing samples. The change of the
recognition rate with the increasing number of principle
components is shown in figure 8. Initially the recognition
rate increases with the increasing number of principle
components, and when the principle component is
increased to 80, the recognition rate reaches an optimal
value of 91.75%. Fig. 8 shows the results of using SVM in
comparison with k Nearest Neighbor. It is obvious that the
accuracy rate of SVM is higher than that of k-NN
(maximum accuracy of 85%).

Figure 8. The comparison of accuracy versus the number of eigenfaces
with NN and SVM algorithms on the ORL database.
50

B. Experiment on JAFFE database
In this face recognition experiment on the JAFFE
database, 100 samples (10 for each individual) are randomly
selected as the training set, from which the eigenfaces and the
support vector machines (SVMs) are calculated. The
remaining 100 samples are used as the test set. Two
classification methods (k-NN and SVM) are used to conduct
experiments on this database. The experiments are conducted
using Cross-Validation analysis with 2-folds.
Since JAFFE face database overall has the face images
of 10 people, the number of SVM classification functions in
the experiment will be 10. In this experiment, the accuracy
reach 100% with 14 principle components used for
training SVMs classifier (in compare with accuracy of
99.5% using Nearest Neighbor). This is likely due to the
number of individuals on this database (10 people versus
40 individuals on ORL database), the number of images are
used for training (10 images each person versus 5 images
for previous experiment). With the simulation result
shown in Fig. 9, it is obvious that SVMs based classification
still has higher accuracy rate in this experiment.


Figure 9. Comparison of accuracy versus the number of eigenfaces with NN
and SVM algorithms on the J AFFE database.
V. CONCLUSION
In this paper, a multiclass recognition strategy for the use of
conventional SVMs is implemented to solve the face
recognition problem. The face recognition experiments using
DAG RBF SVMs has been developed. The effects of SVMs
parameters are also evaluated via a range of different values.
The simulation results show that the SVMs can be effectively
trained for face recognition. As shown in the comparison with
the k-Nearest Neighbor method, the use of SVMs can achieve
high accuracy for face recognition.
REFERENCES
[1] C. Cortes and V. Vapnik, Support-vector networks, Machine learning,
vol. 20, no. 3, pp. 273297, 1995.
[2] Guodong Guo, Stan Z. Li, Kap Luk Chan, Face recognition by support
vector machines, Automatic Face and Gesture Recognition, 2000,
Proceedings. Fourth IEEE International Conference on, pp. 196 -201.
[3] L. Wiskott, J . M. Fellous, N. Kuiger, and C. von der Malsburg, Face
recognition by elastic bunch graph matching, Pattern Analysis and
Machine Intelligence, IEEE Transactions on, vol. 19, no. 7, pp. 775
779, 1997.
[4] Len Bui, Dat Tran, Xu Huang, Chetty, G., Classification of gender
and face based on gradient faces in Visual information Processing
(EUVIP), 2011, pp.269-272
[5] M. A. Turk and A. P. Pentland, Face recognition using eigenfaces, in
Computer Vision and Pattern Recognition, 1991. Proceedings CVPR
91., IEEE Computer Society Conference on, 1991, pp. 586591.
[6] M. S. Bartlett, J . R. Movellan, and T. J . Sejnowski, Face recognition by
independent component analysis, Neural Networks, IEEE Transactions
on, vol. 13, no. 6,pp. 14501464, 2002.
[7] R. Brunelli, T. Poggio, Face recognition: features versus templates,
IEEE Transactions on Pattern Analysis and Machine Intelligence 15 ,
1993, pp.1042-1052
[8] R. Chellapa, C. L. Wilson, S. Sirohey, Human and machine recognition
of faces: a survey, Proceedings of IEEE 83, 1995, pp.705-741
[9] Yang Ming-Hsuan, Kernel eigenfaces vs. kernel fisherfaces: Face
recognition using kernel methods, in Automatic Face and Gesture
Recognition, 2002. Proceedings. Fifth IEEE International Conference
on, 2002, pp. 215220.
[10] V.N. Vapnik, Statiscal Learning Theory, Wiley, New York, 1998
[11] J . Friedman. (1996) Another Approach to Polychotomous Classification.
Dept. Statist., Stanford Univ., Stanford, CA. [Online]. Available:
http://www-stat.stanford.edu/reports/friedman/poly.ps.Z
[12] U. Kreel, B. Schlkopf, C. J . C. Burges, and A. J . Smola, Pairwise
classification and support vector machines, in Advances in Kernel
MethodsSupport Vector Learning, , Eds. Cambridge, MA: MIT Press,
1999, pp. 255268..
[13] J . C. Platt, N. Cristianini, and J . Shawe-Taylor, Large margin DAGs
for multiclass classification, in Advances in Neural Information
Processing Systems. Cambridge, MA: MIT Press, 2000, vol. 12, pp.
547553.
[14] The J AFFE database. Available: http://www.kasrl.org/jaffe.html
[15] The ORL database. Available:
http://www.cl.cam.ac.uk/research/dtg/attarchive/facedatabase.html


51


A SPEAKER RECOGNITION SYSTEM USING COMBINATION METHOD
BETWEEN VECTOR QUANTIZATION AND GAUSSIAN MIXTURE MODEL

Ngo Quoc Hung
Electronic and Telecommunication Eng. Department
Danang University of Technology
Danang, Vietnam
Ngoquochung.dtvt@gmail.com



Pham Van Tuan
Electronic and Telecommunication Eng. Department
Danang University of Technology
Danang, Vietnam
pvtuan@dut.udn.vn



Abstract Speaker recognition is a biometric technique to
recognize peoples identity based on their voice signal. A
recognition system has two main requirements, which are high
accuracy recognition rate and short processing time under large
amount of training data. Many researchers have proposed
various speaker recognition techniques; and the two most
popular methods are Vector Quantization (VQ) and Gaussian
Mixture Model. Each method has its own advantage. VQ method
can perform simply and has fast computation time. However, the
main drawback is that its recognition accuracy rate is not high,
especially with large data sets. Meanwhile, the GMM-UBM has
greater accuracy rate than VQ though, for long processing time,
this process does not always produce satisfying result in practice.
Accordingly, this paper proposes a method to solve the two above
requirements by performing a combination of two advantages of
each VQ and GMM model to provide a new model, which can be
called a Hybrid VQ/GMM-UBM model. This model not only
takes the advantage of high accuracy in GMM method but also
improves the accuracy rate and reduces the amount of
computation of the system when combined with VQ model. The
efficiency of the model is evaluated by computational time and
accuracy rate compared to GMM models. Experimental results
showed that the hybrid VQ/GMM-UBM model had better
accuracy.
Keyword - Speaker recognition, Vector Quantization,
Gaussian Mixture Model
I. INTRODUCTION
Speaker recognition is a biometric technology derived
from areas of speech processing. The speaker recognition field
has over 50 years of research and development. The general
idea of speaker identification tasks is to assume that the voice
of human is unique to each individual, and it can be used as a
distinguished characteristic to identify the owner of that voice
among other individuals.
Speaker recognition system has two operation phases,
training phase and testing phase. In both phases, speech signal
is preprocessed to improve the voice quality and reduce noise.
It then was extracted characteristics to obtain the set of feature
vector. In the training phase, the characteristic vector is used
to train the speaker model. Many methods are used to train
speaker model, from the simplest one which is used to build
codebook model using vector quantization (VQ) (yet the
accuracy of this method is not high) to complex methods such
as Gaussian Mixture Model Universal Background Model.
The overall structure of speaker recognition system is depicted
in Figure.1

Figure 1. General Speaker Recognition System [3]
II. PRE-PROCESSING AND FEATURE EXTRACTION
A. Amplitude Normalized
Voice data was obtained with the amplitude fluctuation.
Even if the speaker says with a standard volume, the
amplitude of obtained signal can still be unstable. This easily
happens when the speaker slightly turns away or moves the
microphone closer to his mouth or pulls it away more than a
few centimeters. This fluctuation affects the recognition
results.
The normalization is necessary. However, it does not
require the signal amplitude to be good, not too small to lose
its characteristics. Thus, we can simply implement by
multiplying each point with an appropriate coefficient k.

(32767/ 2) 100
max ( )
k
s n

52

B. Silience Removal
Speech signal usually contains many silence intervals at
various points such as at the beginning points of the signal,
between words of the sentence or at the end of the signal. If
the signal contains silence intervals without treatment, it will
occupy resources of system to process on these signal
intervals. The silence intervals, however, do not have any
contribution to the identification, even it can interfere to the
processing. Hence, the silence intervals must be treated and
eliminated before implementing feature extraction. Nowadays,
a number of met-hods can effectively solve this problem as
voice activity detection VAD [9][9][9], short time energy or
spectral centroid.
C. Feature Extraction
The extraction of the best parametric representation of
acoustic signals is an important task to produce a better
recognition performance. The efficiency of this phase is
important for the next phase since it affects its behavior.
MFCC is based on human hearing perceptions which cannot
perceive frequencies over 1Khz. In other words, in MFCC is
based on known variation of the human ears critical
bandwidth with frequency is employed. MFCC has two types
of filter which are spaced linearly at low frequency below
1000 Hz. A subjective pitch is present on Mel Frequency
Scale to capture important characteristic of phonetic in speech.
Logarithmic spacing is above 1000Hz.
The overall process of the MFCC is shown in Figure
2.

mel
cepstrum
mel
spectrum
frame continuous
speech
Frame
Blocking
Windowing FFT spectrum
Mel-frequency
Wrapping
Cepstrum

Figure 2. Computing of mel-cepstrum
III. MODEL TRAINING
A. Vector Quantization

Figure 3. Vector Quantization based Codebook of two
speakers
Vector quantization (VQ) is a process of mapping
vectors from a vector space to a finite number of regions
in that space. These regions are called clusters and are
represented by their centroids. A set of centroids, which
represents the whole vector space, is called a codebook. In
speaker identification, VQ is applied on the set of feature
vectors extracted from the speech sample and as a result, the
speaker codebook is generated. Such codebook has a
significantly smaller size than extracted vector set and is
referred as a speaker model. This codebook is generated by
many algorithms such K-mean, LBG
During the matching, a matching score is computed
between extracted feature vectors and every speaker codebook
enrolled in the system. In this paper, matching score is a
Euclean distance [1] between feature vectors and codebook of
speaker as formula:

2
1
1
, min
N
j i i
i
D X C x c
N

where X is a set of N extracted feature vectors, C is
a speaker codebook, x
i
are feature vectors, c
i
are
codebook centroids.
B. Gaussian Mixture Model Universal Background Model
Gaussian Mixture Model is a type of statistical model
which was first introduced by Reynolds [7]. In this approach,
UBM which is a large GMM trained to represent the speaker
independent distribution of features is used. UBM can be
gender independent/dependent model and use EM algorithm to
training [5][6]. After UBM was trained, speaker dependent
models are derived from the UBM by maximum a posteriori
(MAP) adaptation [7]. To form a speaker dependent model,
rst, the log-likelihood of each gender dependent model given
the input data is calculated. The gender is determined by
selecting the gender-model with the higher score. The
corresponding gender dependent UBM is used to adapt a
speaker dependent model (Figure 4) [7]. Regarding speaker
adaptation three EM-steps and a weighting factor of 0.6 for the
adapted model and correspondingly 0.4 for the UBM are used
to merge these models to nal speaker dependent model [8].

Figure 4. Adaptation of speaker model from UBM [7]
53

C. The combination of VQ and GMM-UBM (VQ/GMM-UBM)
As mentioned before, VQ based solution is less accurate
than that of the GMM. In this paper, a method took the
superiority of VQ, which is simplicity computation to
distinguish between male and female speaker. After, we use of
GMM merits to identify the speaker identity in the smaller
subgroup.
In this approach, a testing processing was built on
three stages. In the first stage, feature vectors of testing
speaker were compared with male codebook and female
codebook using Euclean Distance to decide gender of testing
speaker. Male codebook was trained from a large data of male
speakers to represent the male speaker; the same procedure for
female codebook. In the second stage, after knowing gender of
testing speaker, feature vectors of testing speaker were
compared to each VQ model of trained speaker in same
gender group to define ten trained speaker which had the
highest matching scores. In the third stage, ten trained
speakers were computed the log-likelihood with feature
vectors of testing speaker using GMM speaker to define a
final speaker model who had the highest matching score.
After, a threshold was applied to decide accept or reject.
Figure 5 represents speaker identification processing with
combination of VQ/GMM-UBM.

Fig 5. Speaker identification processing with
combination of VQ/GMM-UBM.
Since the idea is using both models of VQ and GMM-
UBM, in training phase, two speaker model groups were built
for male speaker and female speaker as figure 6. Each group
will contain VQ model and GMM-UBM model for each of
training speaker.

Fig 6. Building of two speaker model groups
IV. EXPERIMENTAL SETUP AND RESULT
Speaker database was collected from 150 speakers (70
males, 80 females) whose voices were recorded under the low-
noise environment conditions. The audio files were recorded
from Adobe Audition program, using PCM, sampling
frequency was16000Hz, 16bit. The recording was done
because of two purposes: preparing database for the training
and identification processes.
- For the training process: in this research 100 people
were recorded (50 males and 50 females), each one will be 45
seconds.
- For the identifying process, testing database was taken
from 150 people, including 100 people recorded in the training
process who were identified to be the interested ones, the other
50.
In this paper, data was characteristically extracted with 39
characteristics per frame. For VQ, used size of codebook was
128. For GMM, model had 25 gaussian mixtures.
The time result of identifying process uses VQ, GMM-
UBM and combination of VQ and GMM is shown in Figure.7.
VQ based system had shortest calculating time when
compared with other models. This was the main advantage of
model using VQ, but its accuracy is so low (Figure. 8).
Although GMM model processed the computation for a long
time, merit of GMM model had higher accuracy. Therefore,
with the idea of combining two merits of the two previous
models, time processing of VQ-GMM model was shorten (a
reduction identification time up to 26% is reached) but system
performance is still improved (Figure 7).
54


Fig 7. Identifying time for each speaker with each testing database

Fig 8. DET curve of different modeling techniques

V. CONCLUSION
In this paper, the combination of two techniques has
been executed. From obtained results, we observe that the
combination approach between VQ and GMM is the good
approach due to their different ways of classifying the data.
With this combination, data was classified better in order to
improve the calculating time as well as improve the system
performance.
REFERENCES
[1] Homayoon Beigi (2011), Fundamental of speaker
recognition, Spinger, New York Dordrecht Heidelberg
London.
[2] Piyush Lotia, M.R. Khan (2011), Multistage VQ Based
GMM For Text Independent Speaker identification
System, International Journal of Soft Computing and
Engineering (IJSCE), Vol. 1 (No. 2), pp 21-26.
[3] J oseph Campbell (1997), Speaker Recognition: A
Tutorial, Proceedings of IEEE, Vol. 85 (No. 9), pp 1437-
1462.
[4] Rafik Djemili, Mouldi Bedda, and Hocine Bourouba
(2007), A Hybrid GMM/SVM System for Text
Independent Speaker Identification, World Academy of
Science, pp 448-454.
[5] Richard O. Duda, Peter E. Hart, David G. Stork (2001),
Pattern Classification, Willey Interscien, 2
nd
.
[6] Douglas A. Reynolds, Richar C. Rose (1995), Robust
Text-Independent Speaker Identification Using Gaussian
Mixture Speaker Model, IEEE Transaction on speech and
audio processing, Vol. 3 (No 1), pp 72-83.
[7] Douglas A. Reynolds, Thomas F. Quatieri, Robert B. Dunn
(2000), Speaker Verication Using Adapted Gaussian
Mixture Models, Digital Signal Processing,Vol.10(1-3),
pp.19-41.
[8] Tuan V. Pham, Michael Neffe, Gernot Kubin, Horst Hering
(2007), Speaker Segmentation for Air Traffic Control,
Speaker Classification II, LNAI 4441, pp. 177-191.
[9] Tuan V. Pham, Michael Neffe, Gernot Kubin (2007),
Robust Voice Activity Detection For Narrow-Bandwidth
Speaker Verification Under Adverse Environments,
Interspeech, ISSN: 1990-9772.
[10] Tuan V. Pham(2008), Wavelet Analysis for Robust Speech
Processing and Applications, VDM Verlag Dr. Muller
Aktiengesellschaft & Co. KG, Dudweiler Landstr. 125 a.
[11] Theodoros Giannakopoulos, A method for silence removal
and segmentation of speech signals, implemented in
Matlab Department of Informatics and
Telecommunications, University of Athens, Greece
[12] Vaishali Kulkarni, H. B. Kekre (2010), Speaker
Identification by using Vector Quantization, International
Journal of Engineering Science and Technology, Vol. 2, pp
1325-1331.
[13] Archana Shende, Subhash Mishra , Shiv Kumar (2011),
Comparison of different parameters used in GMM based
automatic speaker recognition, International Journal of
Soft Computing and Engineering (IJSCE), Vol. 1 (No. 3),
pp 14-18.
[14] F. K. Soong, A. E. Rosenberg, L. R. Rabiner and B. H.
J uang (1987), A Vector Quantization Approach to the
Speaker Recognition, AT&T Technical Journal, Vol. 66,
pp. 14-26.
[15] M.G.Sumithra, A.K.Devika (2012), Performance Analysis
of Speaker Identification System Using GMM with VQ,
International Journal of Computer Network and Security
(IJCNS), Vol 4. No 1., pp 14-19.









55

DESIGN OF TELEVISION REMOTE CONTROL
BY VIETNAMESE SPEECH
Nguyen Tu Ha
Danang University of Technology
Da Nang, Viet Nam
haspkt2000@yahoo.com

Pham Van Tuan
Electronics and Telecommunication Department
Danang University of Technologyy
Da Nang, Viet Nam
phamvt1976@yahoo.co.uk



Abstract- The applications of speech recognition for remote
control devices have received increasing attentions from many
scientists as well as manufacturers all over the world nowadays.
Catching up with this trend, this paper presents a study of
designing a Vietnamese speech recognition system for television
remote control. In this research, the recognition of Vietnamese
speech is designed based on the combination of Vector
Quantization (VQ) method and Hidden Markov Models
(HMMs).

Keywords- SPEECH RECOGNITION; MFCC; VECTOR
QUANTIZATION; HMM; IR REMOTE CONTROL

I. INTRODUCTION
Speech recognition is a process which is utilized for
recognizing speech uttered by a speaker. This technology has
been studied for more than five decades since 1950s. Voice
communication is the most effective mode of communication
used by humans. Speech recognition is an important and
emerging technology with great potential. The significance of
speech recognition lies on its simplicity. This simplicity,
together with the ease of operating the speech recognition
devices, has created a great deal of advantages in creating and
using this technology. It has been applied in various fields,
such as security devices, household appliances, cellular
phones, ATM machines, and computers.
The speech recognition technology has been recently applied
in designing voice remote control for television (TV). The two
popular TV producers, Samsung and LG, have presented
various smartTVs which are voice controlled. However,
hitherto, there is no TV production that is controlled by
Vietnamese voice in the market yet.
In this paper, the design of a TV remote control by
Vietnamese speech recognition is presented. The speech
recognition process is based on extracting the speech features
by MFCC method, and recognizing by applying the
combination of Vector Quantization (VQ) method and Hidden
Markov Models (HMMs). A TV controlling module is
designed using control signal emission principle of TV
remote. This module communicates with computer by a
RS232 interface.

Figure 1: Overview about remote control TV by speech
II. THEORETICAL BASIS
2.1. TV remote control signal
Infra-red is frequently used in communication and control,
since it is easily generated and does not suffer from
electromagnetic interference. The limitation of this signal,
however, is that some other light emissions could also contain
infrared, and these can interfere with the communication and
control of devices. In order to allow a good communication
using infra-red, and avoid those "fake" signals, it is imperative
to use a "key" which can detect and inform the receiver what
is the real data transmitted and what is fake.
Remote controls utilize the 36kHz (or the approximation) to
transmit information. Infra-red light emitted by IR Diodes
pulsates at 36 thousand times per second, while transmitting
logic is at level 1 and silence is 0 [9].
Characteristics of remote control signal are shown in Figure 2
below:

Figure 2: Remote control signal characteristics

2.2. Speech recognition system
The speech recognition based on the combination of VQ
method and HMMs is illustrated in Figure 3.
There is a vocabulary of V words to be recognized, and each
word is modeled by a distinct HMM. The training sets consist
of K utterances of each word, pronounced by one or more
56

speakers. In order to obtain a word recognizer, the following
steps are performed.

Figure 3: Schematic diagram of the VQ/HMM
isolated word recognition system
Preprocessing
After being received, voice signal is preprocessed before its
features being extracted. The purpose of this pre-processing
voice signal is to eliminate interference, standardize
amplitude, clarify signal, determine the controlling orders, and
detect the endpoint.
Feature extraction
Feature extraction is a process of doing analysis aimed at
defining important, typical and stable data of voice signal.
This process helps minimize a large set of data in training and
recognition and create a significant decreasing in the number
of estimations in the system. It also clarifies the differences
between this voice and another, at the same time, reduce the
differences between two different pronunciations of a word. In
this study, we chose MFCC approach for extraction of speech
feature.
VQ Codebook
In discrete HMM system, the continuous feature space is
subdivided by a vector quantizer into M non-overlapping
subsets and each subset is represented with a codeword m
(1mM). The set of available codewords is called the
codebook. The VQ codebook is constructed by an
unsupervised cluster algorithm, the Linde-Buzo-Gray (LBG)
algorithm [4] [6].
Re-Estimation of HMM
For each word of vocabulary, a HMM was built. We estimated
the model parameters that optimize the likelihood for the
training set of observation sequences. There are many criteria
that can be used for this problem. In this study, the Baum-
Welch algorithm [5] [7] developed by Baum which is one of
the most successful optimization methods was used.
Recognition
For each unknown word to be recognized, the model
likelihood for all possible models was calculated, and the
model with the highest likelihood was then selected. The
probability calculation was performed using the Viterbi
algorithm [7] [8], more precisely the logarithm of maximum
likelihood.
III. DESIGN OF TV REMOTE CONTROL MODULE
Based on reception and emission of infra-red signal principle
in the TV control remote, a module is designed in order to be
able to learn TV controlling infra-red signal from remote,
subsequently re-emitting the signals to control TV. Process of
learning infra-red signals is operated by counting the time
interval between level 1 and level 0 of infra-red signal series
emitted from TV remote. It then stores into the memory.
When receiving any order of controlling functional buttons
from computer, this module will re-emit the corresponding
series of controlling order that it learned before.
Block diagram of remote control TV module is shown in
figure 4 below:

Figure 4: Block diagram of remote control TV module


Figure 5: Remote control TV module
IV. RESULT
A set of 18 keywords was chosen for this study, namely tt,
bt, tivi, tng, gim, chuyn, m, knh and mt, hai, ba,
bn, nm, su, by, tm, chn, khng. Another set for
purpose of identifying TV controlling orders was also
included.

Speech database was collected from 150 speakers (70males,
80 females) whose voices were recorded in low-noise
environment condition. The sound files were recorded by
adobe audition software, using PCM and obtaining sample at
frequency of 16000 Hz, 16 bit. Sound recording was carried
out with 2 purposes: preparing database for training and for
recognition process. In experimental evaluation process, the
commands which were spoken by 100 people were recorded.
For training and evaluating the identification, the results were
divided into 2 groups: 100 in trained group, and 50 in
untrained group.
57

The recognition results of the words
Table 1: The recognition results of the words

The identification results of the commands
Table 2: The recognition results of the commands (100 people
trained)

Table 3: The recognition results of the commands
(50 people untrained)

V. CONCLUSION
The study has successfully designed the Vietnamese speech
recognition system by applying the combination of VQ
method and HHMs. The speech recognition was indicated to
perform well and rather stable in low-noise environment
condition. TV control module which communicated with
computer by RS232 interface also performed well. It could
"learn" and control almost all kinds of TV.
In order to achieve the better recognition and be able to apply
in practice, it is necessary to build a bigger database with a
more diverse set of vocabulary. Other noise reduce methods,
moreover, should be applied in pre-processing stage.
REFERENCES
[1] Gales. M. and S. Young, The Application of Hidden
Markov Models in Speech Recognition, Foundations and
Trends in Signal Processing, Vol.1, No.2, 2007, p.p 195-304.
[2] L. R. Rabiner, A tutorial on hidden Markov models and
selected applications in speech recognition, Proceedings of
IEEE, vol. 77, no. 2, February 1989, pp. 257286.
[3] Juang, B. H. and L. R. Rabiner, Hidden Markov Models
for Speech Recognition, Technometrics, Vol.33, No.3,
August 1991, pp. 251-272.
[4] Linde Y., Buzo A., and Gray R. M., An Algorithm for
Vector Quantizer, IEEE Transactions on Communication,
Vol.28, No.1, 1980, pp. 84-95.
[5] Segura J. C., Rubio A. J., Peinado A. M., Garcia P., and
Roman R., Multiple VQ Hidden Markov Modeling for
Speech Recognition Speech Communication, Vol.14, 1994,
pp. 163-170.
[6] Balwant A. Sonkamble, D. D. Doye , Speech Recognition
Using Vector Quantization through Modified K-means LBG
Algorithm, Computer Engineering and Intelligent Systems,
ISSN 2222, Vol.3, No.7, 2012, pp.137-144.
[7] Lawrence Rabiner and Biing-Hwang Juang, Fundamentals
of speech recognition, Prentice-Hall International, Inc, 1993.
[8] Huang X. D., Hon H., Hwang M., and Lee K., A
Comparative Study of Discrete, Semi Continuous, and
Continuous Hidden Markov Models, Computer Speech and
Language, Vol.7, 1993, pp. 359-368.
[9] http://www.ustr.net/
58

ECG Signal Transmission Using Wireless
Technology in Patient Health-care and Monitoring
System
Duong Trong Luong, Nguyen Duc Thuan
Dept. of Biomedical Engineering
Hanoi university of Science and Technology
Hanoi, Viet Nam


Nguyen Hung
Dept. of Biomedical Engineering
Hanoi university of Science and Technology
Hanoi, Viet Nam


Abstract- The transmission and reception of biosignals from
human body using wireless technology have been developed
several recent years and designed mainly for center patient
monitoring system that includes a center monitor and station
monitors in hospital (so-called Patient Monitoring system). Due
to the development of telecommunication and Internet, biosignal
transmission using wireless and Internet will lead to big utilities
in patient treatment and examination in real time without
limitations in time and distance. With the aim of developing ECG
signal monitoring system using wireless technology, this research
focuses on the best method of designing an ECG signal
transmission-reception system using Zigbee standard. The system
was tested with simulated ECG signal at a distance of 40 m.
Results obtained show that ECG signals received have quality
equivalent to that of input ECG signals from transmission
module. This is a basic for the author group to do research on
developing and insuring the reliability, reducing the level of
interference in signal transmission and implementing
transmission in long distances and connected via the Internet.

Keywords- ECG signal transmission using Zigbee; patient
remote monitoring system
I. INTRODUCTION
ECG signals in particular and biosignals from the human
body in general are the most important physiological signals of
human. According to several studies, millions of people die of
cardiovascular diseases each year all over the world and also
there are many patients in danger due to heart attacks and
strokes due to late diagnosis and treatment. People with heart
diseases and people due to late detection cannot afford to go to
the hospital regularly. This leads to the need for a remote ECG
signal monitoring system.
In Vietnam, the study of biological signals from the human
body in general and the ECG signals in particular using
wireless technology to connect to the Internet has great
significance in the field of diagnosis and patient health-care.
However, up to now no publication on this field has been
reported. Only one project was proposed in 2012 to coordinate
between the company NTT EAST-Japan and Vietnam Post and
Telecommunications (VNPT) in Vietnam in remote health
management services (ehealth-care management), which
measure the parameters of weight, blood pressure, heart rate,
blood sugar, blood fat, etc. These quantities are expressed in
text and stored in database system and transmitted over a
telecommunication network of VNPT, without ECG signal
processing and transmission (graphic form). In fact, there are
several types of noises intermingle with ECG signals during
recording and transmission. Some of the noises affecting ECG
signals are AC 50 Hz noises, the others are noises due to
patients movement, electrical interference in the propagation
environment, baseline noises (low-frequency noises influence
the process of receiving and analysis of ECG signals;
especially when surveying the ST-T wave of the ECG signal).
Therefore, in addition to system design problems on ECG
signal acquisition and transmission by wireless technology, the
issue of noise reduction is very important in the process of
receiving and processing the ECG signal. In the framework of
research and publication in this paper, we only focus on
designing method of ECG signal transmission system using
Zigbee standard and displaying on the computer.


II. THEORETICAL BASIS
2.1. ECG signal
ECG signal is one of the important parameters of human
and is a typical signal for study and analysis of pathological
and biological mechanisms of the heart. ECG signal frequency
is in range of from 0.05 Hz to 100 Hz. This frequency range is
used for diagnostic applications of cardiovascular diseases. In
addition, the frequency range of ECG signals from 0.05 Hz to
35 Hz is used to monitor the patient's health relating to heart
diseases.
Figure 1.ECG signal
[
[


A
m
p
l
i
t
u
d
e

(
m
V
)

Time(ms)
59

Basic characteristics of the ECG signal are shown in Table
1 and Table 2 below:
TABLE 1. THE AMPLITUDE OF WAVES IN THE ECG SIGNAL
Waves in the ECG signal Amplitude (mV)
P wave 0.25
R wave 1.6
Q wave 0.4
T wave 0.1 0.5

TABLE 2. TIME INTERVALS BETWEEN THE WAVES IN THE ECG
SIGNAL
Waves in the ECG signal Time intervals between
the waves (S)
P-R 0.12 0.2
Q-T 0.35 0.44
S-T 0.05 0.15
P wave 0.05 0.11
QRS 0.05 0.09


2.2. Zigbee (IEEE 802.15.4 standard)
Regarding the design and application of signal transmission
in wireless personal local area network (WPAN or WBAN),
the basic trends are: data transfer at low speed, medium, and
high speed. In particular, Bluetooth data transfer at medium
speed, Zigbee at low data transfer rate. There is no available
IEEE standard for high-speed data transmission in this
network. Zigbee is an architecture that was developed based on
the IEEE 802.15.4 standard, which has lots of advantages
compared to Bluetooth. The comparison of some of the main
parameters between Zigbee and Bluetooth standards is shown
in Table 3. In addition, Zigbee is the perfect solution for
applications based on sensor network. Zigbees data transfer
speed in range of from 20 Kbps to 250 Kbps with different
frequency bands which is shown in Table 4.

TABLE 3. COMPARISON OF ZIGBEE/ IEEE 802.15.4 WITH
BLUETOOTH /IEEE 802.15.1
Standard Zigbee/ 802.15.4 Bluetooth/802.15.1
Transmission
distance (m)
1-75 1-10
Time-life battery
(date)
100-1000 1-7
Number of nodes
in network
>64000 7
Application Monitoring and
control
Web, email, video
Stack size (Kb) 4-32 250
Data rate (Kbps) 20-250 720


TABLE 4. DATA TRANSFER RATE OF ZIGBEE WITH THE DIFFERENT
FREQUENCY BANDS
Bandwidth Applicability Data transfer
rate
channels
2.4GHz World 250Kbps 16
915MHz USA, Japan 40Kbps 10
868MHz Europe 20Kbps 1

Zigbee has three basic topologies: Star, mesh and Cluster-
tree topology. Each topology has its own characteristics, and is
suitable for certain applications. Mesh or Cluster-tree topology
is used for expanding the scope of activities of the network in
about one kilometer.
Following are some characteristics of Zigbee:
Data transfer rate: 20 Kbps to 250 Kbps
Low power consumption. In particular, the capacity
of the receiver / transmitter signal is between 25 and
35 mA, the capacity to "stand by is 3 A.
Transmission distance in sight: 75 m
High reliability
Low costs
Extended ability to 65000 nodes

III. DESIGN OF ECG SIGNAL ACQUISITION AND TRANSMISSION
SYSTEM USING ZIGBEE STANDARD
3.1. Block diagram of the system

















3.2. Function of each module.
Electrodes: measure the ECG signal with the necessary
lead.
Amplifier and noise filter:
Amplifier circuit includes: pre-amplifier and
instrumentation amplifier circuit. Pre-amplifier circuit will
amplify the signal obtained from electrodes. In this circuit,
OP07 IC is used to control wireless base stations.
Instrumentation amplifier uses INA141 IC, which is a IC of
Figure 2. Block diagramof ECG signal transmission and acquisition system
using Zigbee standard associated with the computer.


Patient


Electrodes


Amplifier &
noise filter


PIC
18F26K20


Transmissio
n module
Zigbee
ECG

Acquisition
module
Zigbee

IC
max 232

Computer
60

high accuracy ( 0.05%), low power consumption, power
supply from 2.25V to 18V, and is used in general biosignal
instrumentation amplifier.



VCC+
J 3
TAY PHAI
1
2
3
-
+
U6
OP07
3
2
6
7
4 8
1
VCC-
R5 390K
VCC+
R6 390K
VCC-
VCC+
VCC-
VCC+
OUT1
VCC+
-
+
U3
OP07
3
2
6
7
48
1
R3
22K
R4
10K
SW1
VG
R2
22K
J 4
CHAN PHAI
1
2
3
VCC-
U1
INA141
7
1
8
6
4
2
3
5
V
+
G1
G2
Vout
V
-
Vin-
Vin+
Vref
-
+
U2
OP07
3
2
6
7
48
1
-
+
U7
OP07
3
2
6
7
4 8
1
J 2
TAY TRAI
1
2
3
VCC-








Noise filter circuits include: high-pass filter circuit with
cutoff frequency of about 0.05Hz, low-pass filter with cutoff
frequency of about 100 Hz, band-pass filter and notch filter to
eliminate noise from the AC power supply with frequency
50Hz.









PIC18F26K20: is a microcontroller from Microchip. This
is a programmable IC, it performs the conversion of the signal
from amplifier and noise filter module into digital format
before the signal is fed to the transmission Zigbee module.
Module Zigbee includes PIC18F4620 and MRF24J40MA
IC from Microchip. PIC18F4620 is a multifunction Integrated
Circuit, it has digital and analog input/output pins. This IC
supports RS485, RS232 standards for connections to the
computer. On the other hand, it can work in Master/Slaver
mode. MRF24J40MA performs reception-transmission of
signal via Zigbee standard (802.15.4) with some characteristics
such as operation frequency from 2.405 GHz to 2.48GHz, data
transfer rate of 250Kbps, using carrier sense multiple
access/collision avoid method, auto acknowledgment, check
data frame transmission, ..... A Zigbee module is used as a
station (node) to receive-transmit signal.

Load
-Vin
+Vin
G =10 with pin 1
no connect to pin
8.
G =100 with pin 1
connect to pin 8.
Vout
C6
0.047uF
J 13
CON1
1
J 14
CON1
1
C5
0.047uF
R11
0
R15
+
-
U6
OP07
3
2
6
71
48
Figure 7. Schematic of Band pass filter circuit
J 13
CON1
1
Vcc-
R10
23.7K
R9
23.7K
J 12
CON1
1
0
+
-
OP07
3
2
6
71
48
C3
0.1uF
C4
0.047uF
Vcc+
Figure 6. Schematic of Low pass filter circuit
Vcc+
C7
R12
R14
R16
C8
Vcc-
Vcc+
R13
J 15
CON1
1
0
+
-
U7
OP07
3
2
6
71
48
Vcc-
R15
0
+
-
U6
OP07
3
2
6
71
48
J 16
CON1
1
0
Figure 8. Schematic of Notch filter circuit
Figure 3.Circuit diagram pre-amplifier and ECG signal instrumentation
amplifier.
Figure 4. Equivalent circuit diagraminside instrumentation amplifier IC
INA141.
Vcc+
0
R8
453K
J 10
CON1
1 +
-
OP07
3
2
6
71
48
J 11
CON1
1
Vcc-
C2
10uF
C1
10uF
R7
226K
Figure 5. Schematic of high pass filter circuit
61






When Zigbee module performs the transmission function
(as shown in the block diagram of the system Fig.2), PIC
18F4620 will packet and receive data from 18F26K20 IC data
and then create data frame in Zigbee standard and send to
MRF24J40MA IC which converts received data into frequency
to transmit from antenna. When Zigbee module performs the
receiving function, MRF24J40MA will receive the signal from
the transmission module by antenna and convert it into digital
format and send to PIC 18F4620. This IC transfers and
separates data to the computer by RS 232 connectivity standard
which performed by IC Max 232. ECG signal measured will be
displayed on the computer and can be transmitted to other
terminals via the Internet.
Figure 9 describes the data transmission algorithm from
the ADC inside PIC18F26K20 to the transmission module
Zigbee.
IV. RESULTS
The design of the proposed system was implemented in the
laboratory of biomedical signal measurement of the
Department of Electronics Technology & Biomedical
Engineering, School of Electronics and Telecommunication,
Hanoi University of Science and Technology. The system has
been tested under the following conditions: the distance
between the receiver and the transmitter is 40m in sight,
without any obstruction; ECG signal used in this system is
generated from the ECG signal simulation circuit. Received
ECG signal is displayed on the computer. This signal has the
same quality with the signal from the input of the transmission
module. The result is shown in Figure 10 and Figure 11.





V. CONCLUSION
The author group have researched, designed and tested a
system of ECG signal transmission using Zigbee standard.
Initial result shows that the designed circuits work well with
acceptable accuracy and credibility. In a subsequent study, the
author group will continuously study the noise reduction
solution, and carry out tests and transmission with the ECG
signals from the human body in the natural environment at
larger distances, in the presence of obstruction. The evaluation
of results will also be made.
REFERENCES
[1] Balambigai Subramanian, Efficient Zigbee Based health Care system
for Arrhythmia Detection In Electrocadiogram, European J ournal of
Scientific Research ISSN 1450-216X Vol.69 No.2 (2012), pp. 180-187.
[2] F. Vergari, V. Auteri, C. Corsi, C. Lamberti, A ZigBee-based ECG
transmission for a low cost solution in home care services delivery,
Viale Risorgimento, 2. 40136 BOLOGNA.
[3] H.Labiod, H.Afifi, C. Desantis, WifiTM, BluetoothTM, ZigbeeTM and
WimaxTM. Spinger, 2007.
Figure 11. Testing of ECG signal transmission - receiving and
displaying.
ECG signal
monitoring
Figure 10. Testing of system.
start
Initialize
Check
Stopbit
i=0;

i<=64?
ADC on,
++;
Array[i-1]=
ADC_result;
Zigbee is
ready?
i=0;

No
End
Yes
Yes
Sent array to
Zigbee module;
ADC off;
Clear array;

Yes
No
No
ADC_result 10bits
divided into 2 bytes,
one byte 5bits
+header classify high
byte, low byte
Figure 9. The data transmission algorithm fromthe ADC in PIC 18F26K20
to the transmission module Zigbee.
62

[4] S.Chaudhuri et al., Ambulation Analysis in wearable ECG. Springer
Science+Business Media, LLC 2009.
[5] BURR-BROWN, Precision, Low power Instrumentation Amplifier.
Http://www.burr-brown.com/
[6] 28/40/44-Pin Enhanced Flash Microtrollers with 10-bit A/D and nano
Watt Technology. Microchip Technology Inc, 2004.
[7] 2.4GHz IEEE Std. 802.15.4 RF Transceiver Module. Microchip
Technology Inc, 2008.


63

Combination of analog and digital solutions
for Wireless ECG Monitor
Khue Tra Phung T.KimLai KhiemDuy Nguyen Tuan Van Pham
Electronic & Telecomm. Engineering. Dept., Da Nang University of Technology, Danang
t r akhue08dt 3@gmai l . com, phung08dt 3@gmai l . com, nguyendkhi em89@gmai l . com, pvt uan@dut . udn. vn

Abstract Signal processing, in various forms, is the main part of
a large number of medical equipment today, and it continues to
play a crucial role in the analysis of medical data. In this paper,
combination of analog and digital solutions in wireless
Electrocardiogram (ECG) monitor is presented. The received
ECG signal from patient is fed into pre-processing unit which
has amplification and analog filter such as the Sallen-Key filter,
RC high pass filter and then it is post-processed with digital
filter. In this way, the ECG signal is processed a better way that
still maintains the compact size of the device for portable
applications. Last but not least, five important parameters of
ECG which are heartbeat, QRS duration, ST interval, R
amplitude and T amplitude, are extracted to help doctor to
diagnostic disease.

Keywords Electrocardiogram (ECG), monitoring system,
MSP430, wireless ECG, digital ECG signal
I. INTRODUCTION
Novel methods in medicine which are created by the
application of technology science, have large influences on
characteristics of healthcare system. Nowadays, along with
usefulness of digital signal in processing and storing signal,
digital signal processing is applied popularly, specially in
medicine [1] [2].
In older systems of Electrocardiogram (ECG)
measurement, the ECG signal is processed in analog region.
For example with ECG 6851, most of implementations are
based on analog technology [3]. In the ECG 6851, the filters
such as highpass filter, lowpass filter, Notch filter are
implemented by analog solutions. The filters contain RC
filters or LC filters quite complex, cumbersome.
On the other hand, there are many implementation for
ECG signal in digital region. They are quite flexible and can
store in long time. However, sometime digital ECG signal
processing make distortion and the received ECG parameters
will be affected.
In this paper, combination of analog and digital processing
is presented. With the hand held device, this is quite perfect.
Our device with a moderate price of about one million VND,
in accordance with the affordability of small hospitals,
patients can afford an ECG monitoring circuit that they can
use in their houses. Besides, device is designed with small and
light circuit by using three electrode for measuring a lead.
Patients carry it along and are followed while doing their
normal activities without any feeling of trouble. Instead of
forcing one to lie on a position to measure and record the ECG
signal in a long time, transmission of signals based on
wireless technology allows people to move and act freely. The
ECG signal is still recorded as normal within a relatively wide
range (several tens of meters). In addition, extremely low
power consumption will help to save power for a long-time
using, which reinforces the feasibility of our project.
The paper is organized as follows: Section II presents
system architecture. Afterward, the analog solutions is shown
with the sensor node in Section III. Analysis of digital method
is carried out in Section IV. Finally, conclusion and future
work in Section V conclude the last paper.


















Implementa
tion
Amplifier
(INA118)
Filte
r
ADC
10bit
(MSP430
F2274)
Voltage
Supply
Driver Leg
Circuit
(OP07D)
Wireless
Trans-
mission
(CC2500)
EC
G
sig
nal
Sensor
Wireless
Receiver
(CC2500)
Interface
with
CC2500
(MSP430
F2274)
Interface
with
computer
(TUSB-
3410UF)
Voltage
Supply
Receiver Monitor
Receiver
Process
Show
ECG
signal
Fig. 1 Wireless ECG monitor system diagram.
64

II. SYSTEM ARCHITECTURE
Three main parts of a wireless ECG monitor system are
shown in Fig.1: sensor node, receiver node and monitor node.
The sensor node, worn by the person, is quite small and
portable. In this part, the ECG signal is captured by three
electrodes, and then amplified and filtered by the analog
circuits. Then the MSP430 in the sensor node implements
analog signal to digital signal conversion. At the same time, it
sends all the digitized data to the eZ430-RF2500 wireless
module, which transmits data through 2.4GHz channel [4].
The received ECG data in the receiver node is sent to PC via
serial port connection. In monitor part, PC can be used to
implement digital processing and extract five parameters of
the ECG signal. Then the PC makes unified database
management and information display. Fig. 1 shows a kind of
point to point communication system.
III. ANALOG SOLUTIONS
The ECG signal detected at the electrode is very small. In
theory, the ECG signal at the electrode is about 1mV but in
realistic working condition, it can be smaller [5]. Furthermore,
the ECG signal is affected by various kinds of noise such as
noise 50Hz from powerline noise (in European countries and
the United States is 60 Hz), high frequency noise due to the
vibration of the muscles, low frequency noise makes signal
drift, etc. In this way, in ECG measurement the biggest
challenge is the elimination of the bodys DC oset and the
50/60Hz hum that the human body - as an antenna - collects.
From those above, ECG detection is very difficult and how to
collect ECG signal properly or remove many kinds of noise
more difficult. If the ECG does not collect or collect
incorrectly, all continue processing will not implement.
Besides, if digital processing is abused too much, the ECG
signal will distortion and the received parameters will be
affected. Therefore, analog ECG signal detection is very
necessary and analog solutions is used in here is properly.
A. The Instrumentation Amplifier
In the ECG 6851 catalog [3], amplifier is implemented
using three op-amps with other complex components. This
helps the signal is amplified with amplifier of small noise. As
for the purposes, the INA118 precision instrumentation
amplifier was chosen. We only just with the INA118, those
tasks is finished perfectly and simply. With the gain of signal
amplifier in the INA118 can be changed flexibility by Rg
input resistor. Besides, it has differential inputs with very high
common-mode rejection ratio and integrated three op-amps
for two amplifier stage. This helps cancel common noise
effectively. According to the standard around the world,
current load the person is less than 20A to prevent electrical
shock for patients. The INA118 FET-Input instrumentation
amplifier may provide lower noise and extremely high input
impedance, input current approximately zero, it is very
satisfiable for medical applications [6]. When the input signal
needs amplify with gain is 1000 into 1V to keep output signal
within the input voltage range of the analog-digital converter
(ADC). Thus, two 33 resistors are enough to set the gain of
the instrumentation amplifier.












B. Implementaion of Analog Filters
Since the ECG signal has small amplitude under the
influence of many kind of noise such as the 50/60 Hz noise
from the AC network, the DC noise in contacting body with
electrode or noise for tremor. Those noise occur in the
recorded environment and in the transceiver, the ECG signal
is fed into filters for noise suppression before coming the
analog-digital converter (ADC). The ECG signal has a wide
frequency range but the useful information of ECG signal in
medical is usually in the range of frequency from 0.05Hz to
100Hz [7]. In other paper, the filter is implement with a large
amount of resistor and capacitor. In this paper, the Sallen-Key
topology which is second order active filters is used [8]. This
filter is the super-unity-gain amplifier which allows for very
high Q factor and pass band gain without the use
of inductors. The image of the Sallen-Key filter is shown in
Figure 3.







The cut-off frequency can be chosen below the lowest
frequency component of the ECG signal, which is 0.05Hz in
diagnostic mode and under the highest frequency component of
the ECG signal, which is 100Hz. The cut-off frequency is
calculated with the following equation:
cuto =
1
2nR1R2C1C2

With the high pass filter, R1 =R2 =10k, C1 =C2 =
10F.
C. The Microcontroller
The aim of using microcontroller in this project implements
ADC and transmits ECG signals. With portable application,
this design needs small and ultra-low power. Furthermore, due
to the ECG signal is effect by many type of noise, specially
AC noise, we use battery supply provide whole circuit to
decrease noise. In this application, we use MSP430F2274 is a
part of the family MSP430 designed 16-bit RISC processors
for ultra-low power applications from Texas Instruments. The
MSP430F2274 has 10-bit 200ksps Analog-to-Digital
Converter (A/D) with internal reference, sample and hold,
autoscan, and data transfer controller. Specially the
MSP430F2274 supports UART for interface with the
computer and SPI protocol for RF transceiver.

Fig. 2 The INA118 amplifier.
Fig. 3 The lowpass Sallen-Key filter.
65

The MSP430F2274 is integrated on eZ430-RF2500 module
[4]. This is available and makes a convenient for using RF
transceiver whose figure is shown in Figure 4.


D. The Driver Leg Circuit
Besides two wires are connected as shown in Figure 5, we
can use isolation amplifiers for the direct patient connection.
The body is connected to output of the op-amp. The OP07 is
unsurpassed for low-noise, high-accuracy amplification of
very-low-level signals which is applicated in this circuit. This
reduces the pickup as far as the ECG amplifier is concerned
and effective grounds the patient. This connection prevent
accidental internal cardiac shock and protect to the circuits
and equipment from damage. If abnormal high voltage should
appear between the patient and ground due to electrical
leakage or other means, the auxiliary op-amp saturates.
Ordinary, people connects the third electrode into the right
foot. Performance of elimination noise is better than the
previous method. However, the purpose of this application is
the ECG signal is recorded frequently and patients feel
comfortable with carrying this device regularly. Therefore, the
third electrode is connected as the Figure 5.


E. The Transceiver Chip
The CC2500 is a low-cost 2.4 GHz transceiver designed
for very low-power wireless applications. High sensitivity of
the CC2500 is 104 dBm at 2.4 kBaud and 1% packet error
rate. Although it is good news for those who build high
quality surface mounted boards, some of the potential
customers may fall away as a result of difficulties with
soldering. In the CC2500, frequency range is 2400 2483.5
MHz. This is a standard RF. Hence project development in the
future create a network monitor is implemented easily. The
CC2500 is integrated on eZ430-RF2500 module [4] as
showed in Figure 3.

F. The Voltage Supply
The supply is large enough to control the operation circuit.
Firstly, the INA118 needs it provides stable +2.5V for the
amplifier and ensures that the negative supply stays at -1.9V.
Actually, the amplifier should work well until supply absolute
value is higher than 1.35V. Secondly, the ADC10 module
operates properly which needs higher supply voltage (2.2V)
than other peripherals in the MCU. At minimum voltage is
1.8V, CPU works, but the ADC10 did not. Failures like this
are annoying and can corrupt the samples sent to the
computer.
To require a dual voltage from a single battery, the voltage
inverter is used as shown in Figure 6. This circuit is simple
and inexpensive. We use a 6V battery. After the voltage is fed
through the voltage inverter, it is divided into 3V. The
LM386 is a power amplifier designed for a saving power
consumption application. With purpose creates the same
output voltage, we chose R1 =R2 =470 k and C1 =C2 =
10F.









IV. DIGITAL SOLUTIONS
The ECG signal is affected by various kinds of noise such
as AC 50/60Hz noise, low frequency noise makes signal drift,
etc. They are removed mostly in sensor node with analog
method. However, nothing is absolute, little of noise still exist
with useful signal. Furthermore, in wireless transmission,
noise is added the ECG signal. Hence, digital signal
processing in computer must be used and it occupy important
role.
A. Notch Filter
In this part there is described noise elements filtering with
digital filter. The main noise element is power supply network
50 Hz frequency. This noise has to be removed before the
signal is used for next data processing like heart rate
frequency determination. At computer, this aim is
implemented by Notch filter [9]. The Notch filter is a filter
that passes all frequencies except those in a stop band centered
on a center frequency. The amplitude response form of Notch
filter is drawn in Fig. 7.













Fig. 4 The eZ430-RF2500 module.
Fig. 5 The model of ECG amplifier.
Fig. 6 The model of ECG amplifier.

Fig. 7 The response of Notch filter with cutoff frequency is 50Hz.

66

In here, cut-off frequency of the Notch filter is 50Hz. After
that, the ECG signal before and after Notch filter is shown in
Fig. 8 and Fig. 9.




Ripples in the ECG signal as the Fig. 7 present powerline
interference 50Hz which is eliminated by Notch filter as the
Fig. 8.
B. Parameters Detection
1) QRS Detection
The QRS complex represents the most important
component within the ECG signal [10]. Its detection is the
first step of all kinds of automatic feature extraction. QRS
detector must be able to detect a large number of different
QRS morphologies.
Algorithm of QRS detection is presented in the figure below
[11].


The R peak is the highest peak of the ECG signal. We can
rely upon this peak to detect the rest ones. From the power
spectral analysis of the various signal components in the ECG
signal, a filter can be designed which effectively selects the
QRS complex from the ECG. The received signal is
differentiated to provide information about the slope of the
QRS complex. Then we use threshold to detect the R peak. In
here, adaptive threshold is implemented. We also use a time
constant is 200ms to remove inappropriate R peak. If we have
two R peaks in 200ms, we remove the second peak. Then
adaptive threshold is adjusted automatically.
2) Heartbeat
We realize that RR interval is the one heartbeat. From
detection of R peak, we calculate heartbeat by following
formula.
Ecortbcot =60/ RRintcr:ol
3) P, R, T peak amplitude
For QRS detection, the R peak is detected. We get its
information to show R amplitude.
To detect P peak, we get the peak just before QRS complex.
And between two consecutive R peaks, the maximum value is
T peak.
4) ST interval
From S peak is detected in the QRS complex detection and
T peak is detected as above algorithm, we calculate ST
interval easily.
V. RESULTS AND EVALUATION
The final version of project is depicted in behind figure.


The first method in experiment results of system is
implementation with ECG simulators at different modes, such
as difference for frequency, amplitude, etc. In this case,
frequency we set for the ECG simulator is 60Hz, 75Hz or
35Hz, and on the tested waveform at the output, we found that
the frequency is also approximately the set frequency. In
figure 8, the ECG signal is set at 60 Hz and heartbeat is
calculated from the wave is 5*60/4 =75 times / 1 minute.
Besides, for a standard ECG signal, the QRS, ST and PR
parameters must satisfy at some specific intervals such as the
QRS complex about 50-100 ms, ST interval about 200-300 ms
and PR is 60-120 ms [12]. With these values, doctors can base
on them for heart diseases detection. On experimental results
with simulator, we also found the ECG parameters intervals
satisfy with simulating signal. For the purpose checks
amplitude, the ECG signal from the ECG simulator is set at
1mV and gain of amplifier is 1000. We see that at the ECG
output after amplifying circuit 1V peak. This shows our
amplifying system works fine and stable with simulating
signal.


The tested results as discussed above show the capability
of our system with ECG simulator which produces ECG
signal found in a normal, healthy person. The received ECG
Fig. 8 The ECG signal before Notch filter.

Fig. 9 The ECG signal after Notch filter.
Fig. 10 The QRS algorithm.
Fig. 11 Final project.
Fig. 12 The ECG signal in oscilloscope.
67

signal in oscilloscope is shown in Fig. 12. At this point, we
can conclude that: the acquisition data block, wireless
transmission and PC monitor are stable and accurate for ECG
simulator.
The second method is testing with real persons. We
conducted experiments with different scenarios including: 6
healthy students, 2 students after drinking coffee and 2
persons after waking up. This result is depicted in Table I.
TABLE I
SYSTEM RESULT FOR TESTING ON REAL PERSON
ID student Heartbeat
Student 1 75
Student 2 78
Student 3 70
Student 4 75
Student 5 80
Student 6 70
Student 7 72
Student 8 73
Student 9 76
Student 10 80
According to the tables results, we can ensure that our
system works appropriately with real situations. For practical
uses, one of the most challenging issues for any lab product is
noise, however we was successful in solving noise
cancellation on our device as shown on the table and this will
secure the its application to daily life.
For the amplifier, in this project, we use 3V rail-to-rail
amplifiers. This is one advantage comparing to ECG 6851
commonly found in hospital [3], which use 9V voltage
supply. Based on our expected functions for the device, it will
consume less energy, thus is more power-efficient and the
analog part is less burdensome than other ECG systems. Also,
our device use batteries as the main power supply instead of
AC source like other ECG system sold available in the market,
therefore, we do not need any isolation circuit for it and keep
our device as small as possible. Furthermore, most of the
current ECG systems use analog implementation for filtering
[3]. This is one of the reason why the size for those systems
are not small enough for portable applications. Our device, on
the other hand, applies the combination of analog and digital
filters according to the research of Mr Tamas Hornos [13].
This clever solution enables us to keep the device small for
hand held purposes. The wave we receive at the PC after
digital processing and feature extraction is shown in Figure 13.













In an indoor corridor, maximum transfer distance of more
than 20 meters is fairly good. Because in a real operation
situation, due to absorption of structures or more radio noise
can be present, the signal can be attenuated, more attention
should be paid in future into improving the reliable transfer
distance.
VI. CONCLUSION AND PERSPECTIVES
A new approach for ECG signal monitoring using wireless
technology was designed and implemented for homecare
system of elderly person or patients in this work. The focus of
this paper is to design a compact wireless ECG monitoring
device using commercially available electronic components.
The wave form in screen is pretty accurate and smooth,
showing how well our device works. Wireless
communications implementation in this paper is believed that
it will bring benefits not only to the hospital care systems, but
also to those who need 24 hours of ECG monitoring. In
conclusion, the system implemented has the ability to redefine
not only for current hospital care, but also for work, home and
other activities.
To make the device become more effective and applicable,
future work will develop this project the wireless sensor
network or publish ECG on to a website or mobile phone via
bluetooth connection. As a result, doctors can take care of
many patients at the same time just with a portable device.
Besides monitoring the ECG signal, this includes automatic
ECG recognition to interpret the ECG of the user in
determining the present or upcoming heart problems. These
will create a larger application for this project.
ACKNOWLEDGMENT
The authors would like to thank Mr Hieu V. Nguyen, Da
Nang University of Technology, Mr. Hung T. Le, Texas
Instruments, Vietnam and Texas Instruments Incorporated.
REFERENCES
[1] Cristian Vidal Silva, Andrew Philominraj, A DSP Practical
Application: Working on ECG Signal, Chile, November 23, 2011.
[2] http://www.ntnu.edu/studies/miel/components/signal-processing-in-
medical-applications.
[3] Field,Goth R., Conely J ., ECG recorder : Nihon Kohden ECG-6851
K.
[4] eZ430-RF2500 Development Tool User's Guide, Texas Instruments
Incorporated, SLAU227E September 2007 Revised April 2009.
[5] Abhishek J oshi, Sourabh Ravindran, EKG-Based Heart-Rate Monitor
Implementation on the LaunchPad Value Line Development Kit Using
the MSP430G2452 MCU, Texas Instruments Incorporated, March 2011.
[6] ECG and EEG Applications Quick Reference Guide, Texas Instruments
Incorporated.
[7] Sutar,R.G. ,Electron. Dept., Mumbai Univ., Nagpur, India, ECG
Feature Extraction Using LCAD, May 2012.
[8] Sonia Behra,PremprakashMourya,KamleshChaudhary,Vikas Mishra,
Design of SallenKey Low Pass Filter for High Bandwidth, 2012.
[9] Piskorowski, I., Powerline interference removal from ECG signal
using notch filter with non-zero initial conditions, 18-19 May 2012.
[10] Zahia Zidelmal, Ahmed Amirou, Mourad Adnane, Adel Belouchrani,
QRSdetection based on wavelet coefficients, September, 2012.
[11] Valtino X. Afonso, ECG QRS Detection.
[12] DSaskia A.B.E., Klaas, Monohydroxyethylrutoside as protector against
chronic doxorubicin-induced cardiotoxicity, 19 J ul 2012.
[13] Tamas Hornos, Department of Electronics and Electrical Engineering
University of Glasgow, Wireless ECG/EEG with the MSP430
Microcontroller, Thesis submitted for the degree of Master of Science,
2009. Fig. 13 The ECG signal in PC.

68
Biodegradation of Phenol by Native Bacteria Isolated
from Dioxin Contaminated Soils
Bui Ba Han, Nguyen Thi Lan, Dang Duc Long
*
Department of Biotechnology, Faculty of Chemistry
Da Nang University of Technology, The University of Da Nang
54-Nguyen Luong Bang Street, Da Nang City, Vietnam
*Email: ddlong@dut.edu.vn

Abstract - In this investigation, aerobic bacteria in soil
contaminated with dioxin (taken from Da Nang airports area in
Vietnam) were isolated and selected for their ability to degrade
phenol using an enrichment technique containing phenol as the
sole source of carbon and energy (100 mg/L phenol in a mineral
salt medium). Four strains (designated D1.1, D1.3, D1.4, and
D1.6) were obtained and characterized. The results showed that
these bacteria were highly effective for the removal of phenol.
After 120 hrs of culture, strain D1.4 degraded 54.84% and
44.19% phenol from the initial concentrations of 100 mg/L and
1,000 mg/L, respectively; strain D1.6 degraded 66.45% of phenol
from the initial concentration of 1,500 mg/L. The combination of
those bacteria in the same medium had a positive effect on the
phenol degradation activity. The outcome of the study can
contribute new useful resources for treatments of wastewater and
soils contaminated with phenolic wastes.
Keywords: dioxin-contaminated soil, phenol degradation, aerobic
bacteria, isolation.
I. INTRODUCTION
In the development of the world today, human health and
the environment have become the most pressing issues. Due to
those concerns, biodegradation of aromatic compounds have
received a great deal of attention because of their toxicity and
persistence in the environment. Among all those compounds,
phenol and their derivatives are amongst the most prevalent
forms of chemical pollutants since they are commonly used to
produce many resins, dyes, paints, varnishes, detergents,
herbicides and pharmaceutical drugs. They are also by-
products of many big industries such as petroleum processing,
coke conversion, and steel manufacturing [1, 2, 3]. Phenol can
occur naturally in some agricultural products, animal wastes
and decomposition of organic materials [4]. However, it is
documented to have harmful effects on human health and the
environment. Phenol is a water-soluble and highly mobile
neurotoxin and can cause damage to the human body and other
living organisms through ingestion, inhalation or contact [1,
5]. As a potent air pollutant, phenol can damage structures and
the ozone layer and may reduce visibility and the heat balance
of the atmosphere [5]. Therefore, phenol has been declared to
be a hazardous substance and a hazardous air pollutant by the
United States Environmental Protection Agency [6].
A variety of physical, chemical and biological methods
have been used for the safe removal of such a chemical from
the environment. One of the cheapest and safest solutions for
this is by bioremediation using microorganisms [7]. In general,
it is difficult to degrade phenol by biological methods when its
concentration is above 200 mg/L and thephenol-degradation
activity of microorganisms are completely deactivated at
concentrations larger than 3,000mg/L [8]. Nevertheless,
microbial degradation of phenol with different initial
concentrations ranging from 50-2,000 mg/L have been carried
out in various reactor systems (e.g., shake flask, fluidized-bed
reactor, continuous stirred tank bioreactor) using a variety of
fungi and bacteria such as Candida tropicalis, Acinetobacter
calcoaceticus, Alcaligenes eutrophus, Pseudomonas putida,
Burkholderia cepacia G4, Bacillus stearothermophilus [9, 10].
Those microorganisms have usually been isolated from
environmental samples with high concentrations of pollutants.
In Vietnam, due to historical reasons, there are many sites
contaminated heavily with chemical reagents, especially
dioxin. In those sites there are serious environmental
problems, but the sites also harbour microorganisms with the
ability to metabolize toxic chemicals.
In this study we aimed to isolate naturally occurring
bacterial strains present in those contaminated sites and
characterizing their efficiency of phenol degradation.
II. MATERIALS AND METHODS
A. Soil Samples and Chemicals
Soil samples were collected in Da Nangs airport area in
Vietnam in the spring of 2010. This soil is heavily
contaminated with dioxin from Agent Orange used in the
Vietnam War (1964-1973). At present, the dioxin
concentration in soil there is up to approximately 200 pg
TEQ/g soil [11]. Phenol was obtained from Sigma-Aldrich
Co., UK. The mineral salt medium (MM) used for bacterial
enrichment and isolation was modified from the one used by
Fortnagel and coworkers [12] and comprised (per liter): 3.5g
of Na
2
HPO
4
.2H
2
O, 1.0g of KH
2
PO
4
, 0.5g of (NH
4
)
2
SO
4
; 0.1g
of MgCl
2
.6H
2
O, 50mg of Ca(NO
3
)
2
.4H
2
O, 1 ml of vitamin
B12, and 1 ml of trace salt solution. The final pH of the
medium was 7.2. The trace salt solution contained 0.01g
MoO
3
, 0.07g ZnSO
4
.5H
2
O, 0.005g CuSO
4
.5H
2
O, 0.01g
H
3
BO
3
, 0.01g MnSO
4
.5H
2
O, 0.01g CoCl
2
.6H
2
O and 0.01g
NiSO
4
.7H
2
O in 100 ml water. The MM solid medium
contained 10 g of agar (Meck Co., USA) per liter.
B. Isolation procedure
Soil samples were passed through a sieve (1.7 mm mesh) to
remove large pieces of debris and vegetation. A 0.5% agar
medium (agar in distilled water) was autoclaved and cooled
down (45
o
C). A 100 ml of agar medium was mixed well with

69
1 g of the sieved soil samples. Then 5 ml of the resulting
suspension was spread evenly onto Petri dishes containing the
sterile solid MM medium, which was supplemented with phe-
nol (100 mg/L). Those dishes were cultured in an aerobic con-
dition at 30C for 120 hrs. After this incubation, individual
colonies were transferred onto new MM agar plates with and
without phenol (also 100 mg/L) as the carbon source. After
incubating at 30C for 120 hrs, we selected the colonies that
grew on the medium with phenol, but were unable to grow on
the medium without phenol. Well grown colonies were
maintained on nutrient agar slants and stored at 4C 1
o
C
until used for further experiments. Basic biological
characteristics of the isolated bacterial strains were carried out
by standard laboratory procedures [13].
C. Growth of the bacterial strains in phenol and aromatic
compounds
The isolated cultures were used to inoculate MM
containing phenol at different concentrations and grown at
28C on an orbital shaker at 150 rpm. Similar experiments for
MM containing benzene, aniline and dioxin-contaminated soil
as the carbon source were also conducted. Control samples of
MM without phenol or the other organic substrates were
prepared for reference. All the cultures were limited in
exposure to light. Samples were aseptically removed at regular
intervals and analyzed for growth, substrate removal and pH.
The growth was measured by analysing cell numbers, total
protein concentration, and biomass. To estimate the cell
number, 5 ml of the culture medium was centrifuged at 4,000
rpm for 20 minutes at 4C 1
o
C, then the obtained pellet was
resuspended in 4 ml of phosphate buffer (0.05M, pH 7.2) and
worked up by the method of Bedard et al. [14].
For analysis of the total protein concentration (mg/ml
medium), the samples after the centrifugation were washed a
few times with fresh (phenol-free) MM to remove the
substrate. After cell lysis in the presence of NaOH (0.15 M)
for 5 min at 95
o
C [14], the total protein concentration was
determined by calibration with bovine serum albumin (BSA)
standards according to Biure [15]. The biomass (g/L of
medium) was estimated by a dried weight method [10].
Specifically, 10 ml of the culture broth was centrifuged as
mentioned above, then the pellet was washed twice and finally
transferred from the tube into a pre-weighed 1.2 m pore filter
paper (Whatman GF/C). The filter paper were dried in an oven
at 105
o
C for between 72 hrs, cooled in a desiccator at room
temperature and reweighed to estimate the dry weight of the
biomass. The phenol concentration was determined using a 4-
aminoantipyrine colorimetric approach [16]. The supernatant
of the centrifuged culture medium was reacted with 4-
aminoantipyrine at pH 7.9 0.1 forming a brownish-orange
compound. Subsequently sample absorbance was measured at
500 nm. The phenol concentration was calculated by referring
to the standard curve.
D. Determination of enzyme activities in cell extracts
Manganese peroxidase (MnP) is an enzyme catalyzing the
Mn(II) and H
2
O
2
-dependent oxidation of lignin and a variety
of phenols. We measured the MnP activities of the isolated
bacteria using an assay based on the oxidative coupling of 3-
methyl-2-benzothiazolinone hydrazone (MBTH) and 3-(di-
methylamino)benzoic acid (DMAB) [17]. The reaction of
MBTH and DMAB in the presence of H
2
O
2
, Mn(II), and MnP
gives a deep purple-blue color with the absorption peak at 590
nm. Catalase and oxidase activities were measured by tradi-
tional biochemical methods [18].
III. RESULTS
A. Enrichment, isolation and characterization of phenol-de-
grading bacteria
Starting with two different dioxin-contaminated soil sam-
ples at Da Nang airport (D1 and D2 location), we have been
able to isolate four bacterial strains (D1.1, D1.3, D1.4, and
D1.6) from the D1 sample (Figure 1). The isolates were all
Gram-positive, formed yellowish and slimy colonies, and grew
under strictly aerobic conditions. The strains reacted nega-
tively in an acid resistance test, produced no H
2
S and NH
3
,
showed catalase, but not oxidase activity (Table I). From the
second location (D2), we failed to isolate any aerobic bacteria.
B. The growth of the bacteria in media containing phenol or
other organics


Figure 1. Colonies of the bacterial strain on MM agar. A-D1.1 strain, B-D1.2
strain, C-D1.3 strain, D-D1.6 strain.

TABLE I. MORPHOLOGICAL AND BIOCHEMICAL CHARACT-
ERISTICS OF THE PHENOL-UTILISING BACTERIAL STRAINS
Characteristics D1.1 D1.2 D1.3 D1.6
Gramreaction + + + +
Cell shape Coccus Coccus Oval Oval
Oxygen Requirement Aerobic Aerobic Aerobic Aerobic
Glucose test + + + +
Maniltol test + + + +
Lactose test + + +
Glucose fermentation + + + +
H2S product
NH3 product
Catalase reaction + + + +
Oxydase reacton
Aerobic growth + + + +
Polyphosphate + + +
Acid resistance


70
The growth of the four bacterial strains on MM containing
phenol as a the sole carbon and energy source are shown in
Figure 2. The bacterial strains D1.3, D1.4, and D1.6 rapidly
reached log phase (several hours), while strain D1.1 exhibited
slower growth and reached a lower cell density with lower
protein concentration. It should be noted that the amounts of
protein produced by the bacteria were not tightly linked to the
cell numbers, especially in the case of the strain D1.3. After
120 hrs of incubation, the biomass of strain D1.6 was highest
(0.8 g/L), followed by D1.4 (0.73 g/L), and the lowest value
was observed for strain D1.1 (0.24 g/L). Those growth profiles
indicated that strain D1.1 grew more poorly in the phenol
medium. Also, after cell incubation, the phenol amounts that
were degraded by the four strains- D1.1, D1.3, D1.4, and D1.6
were 31.98%, 51.64%, 54.84% and 47.07%, respectively.
Clearly, the phenol-degradation capabilities of the strains were
not proportionally related to their growth. The reason is that
the phenol degrading activity is dependent on amounts and
characteristics of metabolic enzyme systems in those strains,
not on their cell numbers. To know more about the metabolic
characteristics of the bacterial strains, we grew them in various
liquid media containing other organic carbon sources, from
simple sugars to toxic aromatic compounds. The strains grew
well on media containing glucose, lactose and mannitol,
respectively. They also showed the ability to grow in media
containing benzene and aqueous extract of the dioxin-con-
taminated soil, but not in aniline medium (data not shown).
C. Enzyme activities in crude cell extracts
After 120 hrs of culture, the biomass of each strain was
obtained and the cells were lysed. The activities of MnP of the
four cell extracts showed that strain D1.4 had the highest
enzyme activity, followed by strain D1.3, then strain D1.6, and
strain D1.1 had the lowest activity (Table II). This pattern was
similar to the variation of the phenol-degrading capability of
those strains.
TABLE II. MNP ENZYME ACTIVITY (U/L)
a
OF THE BACTERIA AFTER
120 HRS OF CULTURE ON LIQUID MM
Bacterial strains isolated D1.1 D1.3 D1.4 D1.6
MnP enzyme activity (U/L) 2.9690 5.9791 7.4738 5.2317
0 24 48 72 96 120 144
0
20
40
60
80
100
Strain D1.1
Time (h)
P
h
e
n
o
l

(
m
g
/
l
i
t
e
r
)
0.00
0.02
0.04
0.06
0.08
0.10
0.12
C
e
l
l

n
u
m
b
e
r
x
1
0
8
/
m
l
P
r
o
t
e
i
n

(
m
g
/
m
l
)
0.00
0.01
0.02
0.03
0.04


0 24 48 72 96 120 144
0
20
40
60
80
100
Strai n D1.3
Time (h)
P
h
e
n
o
l

(
m
g
/
l
i
t
e
r
)
0.00
0.05
0.10
0.15
0.20
P
r
o
t
e
i
n

(
m
g
/
m
l
)
C
e
l
l

n
u
m
b
e
r
x
1
0
8
/
m
l
0.00
0.01
0.02
0.03
0.04
0.05

0 24 48 72 96 120 144
0
20
40
60
80
100
Time (h)
P
h
e
n
o
l

(
m
g
/
l
i
t
e
r
)
0.00
0.05
0.10
0.15
0.20
Strai n D1.4
C
e
l
l

n
u
m
b
e
r
x
1
0
8
/
m
l
P
r
o
t
e
i
n

(
m
g
/
m
l
)
0.00
0.01
0.02
0.03
0.04
0.05
0.06
0.07


0 24 48 72 96 120 144
0
20
40
60
80
100
Strain D1.6
Time (h)
P
h
e
n
o
l

(
m
g
/
l
i
t
e
r
)
0.00
0.05
0.10
0.15
0.20
0.25
P
r
o
t
e
i
n

(
m
g
/
m
l
)
C
e
l
l

n
u
m
b
e
r
x
1
0
8
/
m
l
0.00
0.01
0.02
0.03
0.04
0.05


Figure 2. The growth of four microorganismstrains (D1.1, D1.3, D1.4, and D1.6) on mineral salts liquid medium. The culturing process was conducted by
culturing 1 ml of the sample of each strain in 50 ml of mineral salts liquid mediumwith 100 mg/L phenol as only the sole source carbon and energy. After
an appropriate time, the phenol concentrations (mg/L), the total cells (cells10
8
/ml), and the total protein (mg/ml) were analyzed as described in Materials
and Methods. Symbols: , phenol concentrations in the culture medium; , total cell numbers of bacteria; , total protein
concentration.

71
a
An unit of theenzymeactivity was defined as theamount of enzymerequired for theoxidation of
1 M of phenol red per min measured at the wavelength of 610 nm.
D. Ability to grow in different phenol concentrations of the
bacterial strains
All four isolates grew in a series of liquid MM with the
phenol concentrations varying from 100 mg/L, to 300, 1000,
and 1500 mg/L. Based on the resulting cell numbers, protein
concentrations, biomass, and phenol concentrations, we could
observe that the bacterial strains D1.1 only grew well in 100
mg/L phenol, but grew poorly in the higher concentrations of
phenol. Meanwhile, the strain D1.3 only grew poorly when the
concentration of phenol was up to 1,500 mg/L, and the strains
D1.4 and D1.6 grew well in the whole range of the phenol
concentration. The growth of two bacterial strains (D1.4 and
D1.6), which seemed to adapt best to the phenolic envi-
ronment (Figure 3), growing in 1,000 mg/L (strain D1.4) and
1,500 mg/L (strain D1.6).
E. Phenol degradation ability of a mixed culture of the bacte-
rial strains and the native soil microorganism systems
A synergistic action of microorganisms can be helpful in
metabolizing and degrading organic chemicals [5, 19]. To test
this with our isolates, we mixed the four strains together in
equal numbers and grew the mixed culture in MM containing
100 mg/L. For comparison, the two initially dioxin-
contaminated soil samples were also cultured in the same
medium and the amounts of soil were calculated so that they
carried similar numbers of bacterial cells. The phenol
degradation activities of those cultures during 120 hrs of incu-
bation are shown in Figure 4. Kinetics of phenol degradation
of D1, D2 and D1.1-6 were different. For D1, the phenol
degradation was slowly at first, however, at the latter stage
(after 96 hrs of culture) it occured very quicky. Meanwhile,
D1 and D1.1-6 degrade phenol very quickly at the first 72 hrs
after that the degradation was slowed down greatly. However,
all the systems of the microorganisms degraded similar
amounts of phenol after 120 hrs.
F. Influence of environmental factors on the phenol-degrading
capability of the bacterial strains
From the results mentioned above, we found that strain
D1.4 had the best ability to degrade phenol in MM medium.
Therefore, we investigated the influence of environmental
0 24 48 72 96 120 144
0
200
400
600
800
1000
Strain D1.4, 1000 mg/liter
Time (h)
P
h
e
n
o
l

(
m
g
/
l
)
0.00
0.02
0.04
0.06
0.08
0.10
0.12
C
e
l
l

n
u
m
b
e
r
x
1
0
8
/
m
l
P
r
o
t
e
i
n

(
m
g
/
m
l
)
0.00
0.01
0.02
0.03
0.04
0.05

0 24 48 72 96 120 144
0
400
800
1200
1600
Strain D1.6, 1500 mg/liter
Time (h)
P
h
e
n
o
l

(
m
g
/
l
)
0.00
0.03
0.06
0.09
0.12
0.15
C
e
l
l

n
u
m
b
e
r
x
1
0
8
/
m
l
P
r
o
t
e
i
n

(
m
g
/
m
l
)
0.00
0.01
0.02
0.03
0.04
0.05


Figure 3. The growth of two microorganisms (D1.4 [1,000 mg/L] and D1.6 [1,500 mg/L]) in mineral salts liquid medium. Phenol concentrations and the
growth of the microorganismwere determined as described in the Material and Methods. Symbols: , phenol concentrations in the culture medium;
, total cell numbers of bacteria; , total protein concentrations.

Figure 4. The growth of microorganism strain D1.1-6 system and dioxin
contaminated soil (D1 and D2). Mineral salts liquid mediumwas added to the
phenol concentrations of 100 mg/liter, and was cultured at 28C, with a
shaking speed at 150 rpm. Phenol concentrations and the growth of the
cultured bacteria were determined as described in the Material and Methods.
Symbols: , dioxin soil sample D1; , dioxin soil sample D2;
, the mixture of microorganism strain (D1.1, D1.3, D1.4 and D1.6).
Values are the means of 3 replicates and error bars represent the standard
deviations.

72
factors on the phenol-degrading activity of this strain. This set
of experiments would be helpful to identify optimal conditions
of phenol degradation using the isolated bacteria. Figure 5
shows the effects of NaCl, glucose concentrations and pH of
the growth medium on the final phenol concentration after 120
hrs of incubation of the strain D1.4 in the MM containing 100
mg/L phenol.
Low or average concentration of NaCl (from 1 to 10%)
showed a small effect on the capability to degrade phenol of
the bacteria (Figure 5A), only high concentration of NaCl
(15%) began to inhibit the activity of the bacteria. That
probably dues to sterilizing effect of NaCl at that high
concentration. Figure 5B shown that glucose (in the NaCl-free
mediun) had an optimum range of the concentration for the
activity of D1.4 bacteria (around 0.75% w/v).
Higher or lower concentrations of glucose inhibited
significantly the phenol-degrading activity of this strain. At the
condition of 0% NaCl and 0.75% w/v glucose, the optimum
pH value for the acitivity of D1.4 was about 8 (Figure 5C).
IV. DISCUSSIONS
Much research on biodegradation of phenol using pure or
mixed microorganisms isolated from various environmental
sources have been reported [2-5, 7-10, 19-22]. However,
studies on phenol degrading ability of microorganisms from
dioxin-contaminated soils have not been reported. In this
study, we have isolated four aerobic bacterial strains from
dioxin-contaminated soil that can all use phenol as the source
of carbon and energy for their growth.
In addition, those bacteria can also utilize several toxic
aromatic compounds, such as benzene and dioxin derivatives.
Due to this kind of metabolic capability, those bacterial strains
can degrade phenol and different aromatic contaminants. The
phenol degradation capabilities were not the same for those
isolates. Through culturing the strains in a wide range of
phenol concentration, we selected two strains, D1.4 and D1.6,
which can grow and metabolize a very high concentration of
phenol (1,500 mg/L). Our assay suggested that the phenol-
degrading activity of the bacteria were closely linked to the
activity of manganese peroxidase, an important enzyme for



Figure 5. Effect of some factors on phenol degradation by D1.4. (A) The influence of NaCl, (B) glucose concentrations, (C) and pH on the ability of phenol
decomposition by strain D1.4 in MM containing 100 mg/L phenol. A mineral salts mediumwas supplemented as Materials and Methods. After 120 hrs
culturing, the remaining phenol concentrations was determined. Values are the means of 3 replicates and error bars represent the standard devations.


73
degradation of a wide variety of chemicals and polymers.
These strains can be used as a biological reagent for
processing phenol and aromatic chemicals in soil and
wastewater. Overall, our study here demonstrates that dioxin-
contaminated soils are valuable sources of microorganisms
that can be beneficial in environmental protection as well as in
other fields.
However, we have not been able to isolate all the phenol-
degrading microorganisms existing in the soil samples. For
example, the soil sample D2 showed a strong phenol-
degrading activity in its native state, but we were unable
toisolate any bacteria using our selection procedure. Since the
procedure was designed for isolating aerobic bacteria, it is
possible that in soil sample D2, the dominant phenol-
degrading microorganisms are different, such as anaerobic
bacteria, or fungi. The mixed culture of the strains had better
phenol degradation than each individual bacterium, showing
that the synergistic actions of the native organisms could be
very important in the degrading capability. We have not
investigated thoroughly this kind of synergy in this study, but
it will be an important topic so that the best biodegradation
reagent for phenol can be constructed from the soil samples.
Furthermore, identifying the best biological agent is
required as well as identifying the optimal condition of the
growth medium. In general, the presence of NaCl in the
medium had a positive effect on the capability of phenol
degradation. The bacteria grew better in MM with glucose (in
the range of 0.25 0.75% w/v) than without glucose. pH of
the medium could also influence the ability to degrade phenol
of the strain D1.4.
With the results in our study, we have established the
fundamental research and benefits of utilizing heavily dioxin-
contaminated soil microorganisms for bioremediation of sites
and the treatment of industrial wastewater contaminated with
phenolic wastes.
REFERENCES
[1]. Weber M, Weber M, Kleine-Boymann M, Phenol. Ullmann's
Encyclopedia of Industrial Chemistry, 2004.
[2]. Ajaz M, Noor N, Rasool SA, Khan SA, Phenol Resistant Bacteria
from Soil: Identification, Characterization and Genetical Studies,
Pakistan J ournal of Botany, Vol 36, pp. 415 424, 2006.
[3]. Gayathri KV, Vasudevan N, Enrichment of Phenol Degrading
Moderately Halophilic Bacterial Consortium from Saline Environ-
ment., J . Bioremed Biodegrad, , Vol 1, pp. 1-6, 2010.
[4]. Prpich GP, Daugulis AJ , Enhanced biodecomposition of phenol by a
microbial consortiumin a solid liquid two phase partitioning
bioreactor, Biodecomposition, Vol 16, pp. 329 339, 2005.
[5]. J ame SA, Rashidul Alam AKM, Fakhruddin ANM, Alam MK,
Degradation of Phenol by Mixed Culture of Locally Isolated Pseudo-
monas Species, J . Bioremed Biodegrad Vol 1, pp. 1-4, 2010.
[6]. United States Environmental Protection Agency (US-EPA), Codes
of Federal Regulations, 40 CFR 116.4, 40 CFR 372.65, 40 CFR
302.4 and 40 CER 355. http://www.epa.gov/epacfr40/chapt-
1.info/chi-toc.htm, 2006
[7]. Chakraborty S, Bhattacharya T, Patel TN, Tiwari KK, Biodegra-
dation of phenol by native microorganisms isolated from coke
processing wastewater, J ournal of Environmental Biology, Vol 31,
pp. 293-296, 2010.
[8]. J iang H, Fang Y, Fu Y, Guo QX, Studies on the extraction of phenol
in wastewater, J ournal of Hazardous Materials, Vol 101, pp. 179-
190, 2003.
[9]. Nair CI, J ayachandran K, Shashidhar S, Biodecomposition of
Phenol, African J ournal of Biotechnology, Vol 7, pp. 4951-4958,
2008.
[10]. Agarry SE, Solomon BO (2008) Kinetics of batch microbial degrada-
tion of phenols by indigenous Pseudomonas fluorescence. Interna-
tional J ournal of Environmental Science and Technology 5: 223-232.
[11]. Huu NB, Ha DTC, Characterization of bacterial community
structure in bioremediation treatment of herbicide/dioxin contami-
nated soil at field trials, Vietnam J ournal of Biology, Vol 32, pp. 88-
93, 2010.
[12]. Fortnagel P, Harms H, Wittich RM, Krohn S, Meyer H, Metabolism
of Dibenzofuran by Pseudomonas sp. Strain HH69 and the Mixed
Culture HH27, Applied and Environmental Microbiology Vol 56,
pp. 1148-1156, 1990.
[13]. Brenner DJ , Krieg NR, Staley J T (Eds.), Bergeys Manual of
Systematic Bacteriology, Volume 2, 2nd ed., Springer, 2005.
[14]. Bedard DL, Haberl ML, May RJ , Brennan MJ , Evidence for novel
mechanisms of polychlorinated biphenyl metabolismin Alcaligenes
eutrophus H850, Applied and Environmental Microbiology Vol 53,
pp. 1103-1112, 1987.
[15]. Layne E, Spectrophotometric and Turbidimetric Methods for
Measuring Proteins, Methods in Enzymology, Vol 10, pp. 447 455,
1957.
[16]. APHA (1998) Standard Methods for the Examination of Water and
Wastewater, 20th edn. American Public Health Association, Wash-
ington, D.C.
[17]. Castillo MP, Stenstrm J , Ander P, Determination of manganese
peroxidase activity with 3-methyl-2-benzothiazolinone hydrazone and
3-(dimethylamino) benzoic acid, Analytical Biochemistry Vol 218,
pp. 99-404, 1994.
[18]. Benson HJ , Microbiological Applications: A Laboratory Manual in
General Microbiology, Complete Version, 8th Edition, McGraw-Hill
Science/Engineering/Math, USA, 2001.
[19]. Marrot B, Barrios-Martinez A, Moulin P, Roche N, Biodegradation
of high phenol concentration by activated sludge in an immersed
membrane bioreactor, Biochemical Engineering J ournal Vol 30, pp.
174-183, 2006.
[20]. Stephen AB, Shelton DR, Berry D, Tiede J M, Anaerobic Biode-
composition of phenolic compounds in digested sludge,
Applied and Environmental Microbiology, Vol 46, pp. 50-54, 1983.
[21]. Paula M., van Schei PM, Young LY (1998) Isolation and
Characterization of phenol degrading denitrifying bacteria.
Applied andEnvironmental Microbiology 64: 2432 2438.
[22]. Adav SS, Chen MY, Lee DJ , Ren NQ, Degradation of Phenol by
Aerobic Granules and Isolated Yeast Candida tropicalis, Biotech-
nology and Bioengineering Vol 96, pp. 844-852, 2007

74

Biopolymer Film from Chitosan for Shelf-life
Extension of Fruits

NGUYEN, Thi Xuan Lam; NGUYEN, Thi Minh Xuan; DANG, Duc Long
Biotechnology Department, Faculty of Chemistry Engineering
Danang University of Technology, the University of Danang
Danang, Vietnam
ntmxuan@dut.udn.vn; ddlong@dut.udn.vn



Abstract After harvesting, fruits and vegetables deteriorate
rapidly. The main causes of their deterioration are dehydration
and attack of microorganisms. Therefore, the use of edible
coatings on the surface, followed by a cold storage, will prolong
shelf-life of fruits and vegetables. Chitosan, a derivation of
deacetylated chitin, a high bioactive substance, becomes a
promising alternative treatment to toxic chemical in
preservation. However, the studies in Vietnam only focused on
initial investigation of using chitosan film to extend storage life of
these perishable commodities, but not factors boosting bioactivity
of chitosan. Our research studies the effect of deacetylation
degree, molecular weight of chitosan and soaking time in chitosan
solution on mango preservation. Our results showed mangoes
dipped into chitosan with low molecular weight at high degree of
deacetylation for 10 minutes prolonged shelf life greater than
control without chitosan coating
Keywords-biopolymer film, chitosan, shelf-life extension,
deacetylation degree, mango storage
I. INTRODUCTION
Fruits play an important role in the Vietnam economy.
Revenues from exporting fruits during period 2004-2011, have
increased growth rates that average of 20% each year and
revenues reached USD 600 millions in 2011. Today, our fruits
are sold in over 50 countries around the world. The largest
markets are developed countries such as Japan, Germany,
England, Canada, Singapore In which, mangoes are one of
the most cultivated fruit in the tropics, including Vietnam.
FAO, 2009 predicted that the growth rate of mangoes will
accelerate in the following years. Recent year, many potential
markets have opened for importing our mangoes such as New-
Zealand, South Korean and Spanish.
However, all fruits are still alive after harvesting, that
means they continue to respire and transpire. This leads to lose
water which usually causes shrinkage. Moreover, respiration
induces transformation of organic matter to CO
2
and water.
Therefore, weight loss, color changes, softening and microbial
spoilage occur during transportation and marketing. Treatment
with chemical is usually used to extend shelf-life of fruits.
However, nowadays, consumers around the world require
high-quality food, without traditional preservatives, leading to
increased effort in discovering new natural preservative.
Chitosan is a linear polysaccharide composed of randomly
distributed -(1-4)-linked D-glucosamine (deacetylated unit)
and N-acetyl-D-glucosamine (acetylated unit). It is made by
treating shrimp and other crustacean shells with the alkali
sodium hydroxide. Chitosan became an ideal choice to
alternate other chemical preservative because of its film-
forming, biochemical properties which has led to prolonged
storage life, and total natural product. Indeed, chitosan coating
were demonstrated that they can decrease the respiration rate of
fruits [13] and also delay the postharvest development of
microbe causing diseases for fruits [10].
In Vietnam, there have been some initial achievements in
using chitosan to extend storage life of fruits. Nguyen Van
Toan et al. (2009), proposed a preserving-banana process
which increased their shelf-life three times by chitosan.
Tomatoes, grapefruits stored by films from chitosan have
showed significant extension of their shelf-life compared to
controls. However, these studies only investigated chitosan
concentration, methods for making film (spraying or dipping)
to raise storage capacity. While chitosan at higher degree of
deacetylation and lower molecular weight have boosted the
antimicrobial activity of chitosan rather than lower degree of
deacetylation and higher molecular weight [1, 4, 11, 12, 14].
Our research considered these effect factors on bioactivity of
chitosan in mango preservation.
II. MATERIALS AND METHODS
Mangoes (Mangifera cambodiana) were harvested at Cam
Hai, Cam Ranh, Khanh Hoa province, Vietnam. After
transported to our laboratory, their surfaces were pretreated
immediately by CaCl
2
and H
2
O
2
before coating or non-coating
with chitosan.
Chitosan used are a gift from Nha Trang University.
A. Determining weight loss
Postharvest mangoes pretreated surfaces were numbered
and weighed. After 4, 8, 12, 14 storage days at room
temperature, mangoes with or without chitosan-coating were
weighed again. Weight loss was calculated by difference of
weight of mangoes between before and after storage time.
75

B. Determining total sugar content
Total reducing sugar content was determined by Bertrand
method. The general principle of this method is based on the
oxidation reaction between sugar and metal ion under alkaline
condition.
C. Determining vitamin C content
Vitamin C content was measured based on ability of
vitamin C that can reduce iodine. Briefly, 5 g sample were
grinded with sand in 5ml HCl 5% solution. Then, add distilled
water to 50 ml. 20 ml of this solution was used to titrate by
iodine with soluble starch as indicator.
III. RESULTS
A. Chemical composition of mangoes
The alteration in compositions of mango fruits during
ripening in non-preservative condition was revealed in table I.
These results were going to be used to compare with mangoes
stored by chitosan films later.
Table I: Chemical compositions of green and ripe mango
Compositions Green mango Ripe mango
Water (%)
84.02 82.34
Dissolved solid
content (Bx)
9.50 17.00
Reducing sugar
content %
6.30 16.15
Total acidity %
2.59 0.17
Vitamin C (mg) 44.64 21.80
pH
2.36 5.02
B. The effect of chitosan with different degree of deacetylation
on mango storability
The antimicrobial activity of chitosan was stronger at
higher degree of deacetylation of chitosan [11, 12, 14].
Chitosan with a deacetylation degree of 92.5% was more
effective than chitosan with deacetylation degree of 85% [15].
This may protect fruits from the attack of microbe and prevent
deterioration of fruits by microbial spoilage. However there
was also a report that the antimicrobial activity of chitosan did
not depend on change in degree of deacetylation of chitosan
(73, 84 and 95%) [7].
To clarify the effect of deacetylation degree of chitosan on
shelf-life extension of mangoes, chitosan with degree of
deacetylation of 93.32%, 85.8% with the same molecular
weight of 200 kDa were used in our researches. Chitosan were
prepared by dissolving 1g chitosan in 100 ml 1% acetic acid
solvent in 2 hours with stirring until obtaining homogeneous
solution. Mangoes pretreated with CaCl
2
and H
2
O
2
before
coating or non-coating with chitosan. The day, when mangoes
were dipped into prepared chitosan solution in 10 minutes and
stored at room temperature, was initial time of the process of
preservation.
To assess storage ability of chitosan with different
deacetylation degree, weight loss, sugar content and vitamin C
content were examined after 4, 8, 12 and 14 days of treatment
with or without chitosan.
Chitosan with higher deacetylation degree prevented
weight loss of postharvest mangoes better.
Weight loss is an important parameter to evaluate the
quality of postharvest fruits and vegetables because of
reduction of economic benefit and sensory value of products
from this loss. This loss causes by evaporation and
metabolism of organic matter in respiratory of fruits and
vegetables.
The ability of chitosan at different degree of deacetylation
in maintaining weight of mangoes was determined by
weighing mangoes treated with chitosan solution after 4, 8, 12,
and 14 days of treatment. Below figure has showed difference
of mango weight loss between before and after dipped into
chitosan solution.

Figure 1: The weight loss of mangoes after 4, 8, 12, and 14 days of treatment
into chitosan solution with deacetylation degree of 93.32% (DH, blue line,
diamond) or 85.8% (DL, red line, square); and mangoes without chitosan-
coating (CT, green line, tangle) as control.
Figure 1 illustrated that the weight loss of mangoes coated
chitosan film was significant less than the loss of mangoes
non-coated. After 9 storage days, all mangoes non-treated with
chitosan were rotten, while shelf-life of chitosan-coated
mangoes prolonged over 14 days. Moreover, chitosan with
various deacetylation degrees also conserved mango weight
light differently. The more chitosan was deacetylated, the less
weight loss during storing was.
The hydrolysis of carbohydrate occurred slower in
postharvest mangoes coated chitosan with more
deacetylation.
During ripening, sugar content increase by hydrolysis of
hemicelluloses, pectin and starch. Sugar content also indicates
ripe degree of fruits. To prevent the loss of sugar, storage
methods need to restrict this sugar produced.
Determination of reducing sugar content in 100g sample
was carried out by Bertrand method. The method is based on
the redox reaction of reducing sugars (glucose, fructose,
maltose, etc.) with oxide copper II to generate oxide copper I
under the alkaline condition. Oxide copper I, on which the
76

quantity of reducing sugar can be calculated, has red color like
brick.

Figure 2: Sugar content in 100g of mangoes after 4, 8, 12, 14 days of
treatment into chitosan solution with deacetylation degree of 93.32% (DH,
blue line, diamond) or 85.8% (DL, red line, square); and mangoes without
chitosan-coating (CT, green line, tangle) as control.
Sugar content of mangoes coated chitosan film increased
markedly slower than sugar content in mangoes non-coated
after 4, 8, 12, 14 days of treatment. After 14 storage days,
sugar content of mangoes with chitosan film was equivalent to
that of control at storage day of eighth. The similar to weight
loss, the more chitosan was deacetylated, the slower
hydrolysis was.
Decline of vitamin C were impeded more effectively by
chitosan-coating film with the higher degree of
deacetylation.
Human beings cannot synthesize vitamin, a vital
substance. Almost vitamins, including vitamin C, have been
absorbed from fruits and vegetables. That is a reason why
vitamin C content becomes an important index to indicate the
quality of fruits.
Vitamin C content in mangoes treated or non-treated with
deacetylated chitosan at different degree were determined by
redox titration with iodine.

Figure 3: Vitamin C content (mg) of mangoes after 4, 8, 12, 14 days of
treatment into chitosan solution with deacetylation degree of 93.32% (DH,
blue line, diamond) or 85.8% (DL, red line, square); and mangoes without
chitosan-coating (CT, green line, tangle) as control.
Vitamin C content lost during storage even mangoes
treated or non-treated with chitosan solution. However, this
loss was inhibited significantly by chitosan with higher
deacetylation degree. Indeed, after 14 days of treatment,
vitamin C content of mangoes dipped into deacetylated
chitosan of 93.32%, 85.8% were 33.12 mg, and 23.23 mg,
respectively.
C. The effect of different molecular weight chitosan on
mango storability
Recent researches have explored the effectiveness of
different molecular weight chitosan in controlling fruit decay
caused fungi as in vitro and in vivo [1, 4]. Coating fruits with
chitosan of lower molecular weight was more effective in
controlling the growth of fungi than chitosan of higher
molecular weight. However, chitosan of low molecular
weight was difficult to form film, and so hinder them from
fruit preservation.
To determine appropriate molecular weight of chitosan
for mango storage, chitosan of 200 kDa, 12.96 kDa, 6 kDa
and mixture of 200 kDa and 6 kDa with ratio 1:1 were used
in our researches. All chitosan with different molecular
weight have the same degree of deacetylation and were
prepared as above describe.
The effect of different molecular weight chitosan on
weight loss of postharvest mangoes.

Figure 4: The weight loss of mangoes after 4, 8, 12, 14 days of treatment
into chitosan solution with different molecular weight of 200kDa (PH, blue
line, diamond); or 12.96 kDa (PM, red line, square) or 6kDa (PL, green
line, tangle);or into mixture of chitosan of 200kDa and 6kDa with ratio 1:1
(PHL, purple line, cross); and mangoes without chitosan-coating (CT, light
blue line, asterisk ) as control
This result proved again that chitosan inhibited
significantly weight loss of mangoes after harvesting.
However, different molecular weight chitosan had different
effect on this inhibition. Interestingly, the most success in
maintaining weight of postharvest mangoes was mixture of
chitosan of high and low molecular weight with ratio 1:1.
Indeed, after 14 days of treatment, in 100 g of mangoes
treated with chitosan of 12.96 kDa and 6 kDa were lost 5.47
g and 8.89 g, respectively, while mangoes treated with
mixture were lost only 4.65 g.
The effect of different molecular weight chitosan on sugar
content of postharvest mangoes.
As the above explanation, reducing sugar content is an
important indicator of ripe degree of postharvest fruits. The
faster sugar content increase, the more difficultly fruits can
be stored.
77


Figure 5: Sugar content in 100g of mangoes after 4, 8, 12, 14 days of
treatment into chitosan solution with different molecular weight of 200kDa
(PH, blue line, diamond); or 12.96 kDa (PM, red line, square) or 6kDa (PL,
green line, tangle);or into mixture of chitosan of 200kDa and 6kDa with
ratio 1:1 (PHL, purple line, cross); and mangoes without chitosan-coating
(CT, light blue line, asterisk ) as control.
Figure 5 illustrated again that chitosan could prevent
hydrolysis of carbohydrate in postharvest fruits. Chitosan of
12.96 kDa (PM) and 6 kDa (PL) showed stronger inhibition
compared to chitosan of 200 kDa (PH). However, the mixture
of high and low molecular weight chitosan was also the most
effective during storing mangoes.
The effect of different molecular weight chitosan on
vitamin C content of postharvest mangoes

Figure 6: Vitamin C content (mg) of mangoes after 4, 8, 12, 14 days of
treatment into chitosan solution with different molecular weight of 200kDa
(PH, blue line, diamond); or 12.96 kDa (PM, red line, square) or 6kDa (PL,
green line, tangle);or into mixture of chitosan of 200kDa and 6kDa with
ratio 1:1 (PHL, purple line, cross); and mangoes without chitosan-coating
(CT, light blue line, asterisk ) as control.
The results revealed that chitosan impeded the loss of
vitamin C in mangoes during storing, but different molecular
weight chitosan did not have significantly difference in
conserving vitamin C content. Nevertheless, mixture of high
and low molecular weight chitosan showed the best storability
of vitamin C.
D. The effect of soaking time in chitosan solution on mango
storability.
From the above results, low molecular weight chitosan was
demonstrated more effective than high molecular weight
chitosan in fruit preservation. However, with low molecular
weight chitosan, fruits need soaking more time into chitosan
solution to form film. This can damage chlorophyll in green
mango peel. Consequently, sensory value of mangoes and
storage life was shortened by this damage. Besides that,
different degree of deacetylation of chitosan can also affect
generation of film.
To find out the appropriate soaking time, prepared
mangoes were dipped into different degree of deacetylation of
chitosan and different molecular weight chitosan solution for
2, 5, 10 minutes. Storage life of mangoes was from time
mangoes dipped into chitosan solution until mangoes appeared
phenomena such as dark yellow peel, wrinkled and soft fruits.
The effect of time mangoes soaked into chitosan with
different degree of deacetylation on shelf-life
Table II: Shelf-life of mangoes coated chitosan of different degree of
deacetylation for different soaking time
Degree of deacetylation
of chitosan
Soaking time
(minutes) Shelf-life
93.32%
2 11
5 12
10 16
85.80%
2 12
5 12
10 13
Without chitosan 0 9
Storage time of mangoes dipped into chitosan with the
same degree of deacetylation increased with soaking time. In
all samples, shelf-life of mangoes soaked into chitosan of
highest deacetylation degree of 93.32% for 10 minutes was
longest.
The effect of time mangoes soaked into chitosan with
different molecular weight on shelf-life
Table III: Shelf-life of mangoes coated chitosan at different molecular weight
for different soaking-time
Molecular weight of
chitosan
Soaking time
(minutes) Shelf-life
200 kDa
2 12
5 12
10 13
12.96 kDa
2 12
5 14
10 14
6 kDa
2 11
5 14
10 14
Mixture of chitosan of 200
and 6 kDa
2 14
5 16
10 18
Without chitosan 0 9
Storage time of mangoes dipped into chitosan with the
same molecular weight increased with soaking time. With
chitosan at molecular weight of 12.96 and 6 kDa, shelf-life of
mangoes did not show any difference between soaking time of
5 or 10 minutes. In all samples, shelf-life of mangoes soaked
into mixture of chitosan of 200 and 6 kDa for 10 minutes was
longest.
IV. DISCUSSION
Our research demonstrated that mango storability by
chitosan with high degree of deacetylation were better than
chitosan with low degree by maintaining weight, vitamin C
78

content, delaying transformation of carbohydrate to sugar, and
extending storage-life of mangoes. The similar results have
been documented in preserving fresh strawberries [5, 13].
Some reports suggested that effects of chitosan may be
associated with direct anti-microbe properties against the
postharvest pathogen [1, 2, 11]. The number of amino groups
on chitosan increase with increased degree of deacetylation.
As a result, chitosan with higher number of amino group
dissolves completely, which increase chance of interaction
between chitosan and microbial cell walls.
The molecular weight of chitosan was also showed effect
on their antimicrobial activity in numerous reports [3, 6, 8, 9,
17]. Generally low molecular weight chitosan revealed greater
effectively than high molecular weight. Indeed, coating fruits
with chitosan of 15 kDa was more effective in controlling the
growth of Penicillium digitatum and P. expansum causing fruit
decay than chitosan of 357kDa [4]. In addition, Jeon et al. [8]
suggested that better antimicrobial activity of chitosan was at
molecular weight more than 9 kDa. Indeed, chitosan of
9.3kDa inhibited effectively the growth of Escherichia coli,
while that with molecular of 2.2kDa enhanced growth of the
same bacteria [16]. Our results displayed the same trend in
mango storability and shelf-life extension. Chitosan films of
12.96 kDa presented a better effective in slowing down weight
loss, production of reducing sugar, and increasing storage-life
of mangoes than chitosan of 200 kDa and 6 kDa. However, the
difference between 12.96 kDa and 6 kDa was not markedly,
especially in maintaining vitamin C. Small molecule of
chitosan may penetrate into microorganism more easily than
large molecule. However, film coating on the surface of
mangoes are difficult to be made from small molecules. To
overcome this difficulty, we mixed chitosan of high and small
molecular weight. This mixture had significantly effective on
mango storage compared to individual kind of chitosan.
In conclusion, chitosan have greater effective in extending
shelf-life and maintaining the quality of postharvest mangoes.
However, the effect of chitosan at different degree of
deacetylation and molecular weight on fruit preservation was
different. Higher degree of deacetylation and lower molecular
weight of chitosan have better effective. Interestingly, we find
out that the mixture of chitosan at 200 kDa and 6 kDa with
93.32% degree deacetylation presented the most effective in
fruit storability.
V. REFERENCES
[1] Badawy, M.E.I. and E.I. Rabea, "Potential of the biopolymer chitosan with
different molecular weights to control postharvest gray mold of tomato
fruit". Postharvest Biology and Technology, 2009. 51(1): p. 18.
[2] Badawy, M.E.I. and E.I. Rabea, "A Biopolymer Chitosan and Its
Derivatives as Promising Antimicrobial Agents against Plant Pathogens
and Their Applications in Crop Protection". International J ournal of
Carbohydrate Chemistry, Review article, 2011. 2011: p. 29.
[3] Chang, D.S., et al., "A development of food preservative with the waste of
crab processing". Bulletin of the Korean Fisheries Society, 1989. 22: p. 9.
[4] Chien, P.J ., F. Sheu, and H.R. Lin, "Coating citrus (Murcott tangor) fruit
with low molecular weight chitosan increases postharvest quality and shelf
life". Food Chemistry, 2007. 10(3): p. 5.
[5] Ghaouth, A.E., et al., "Chitosan coating effect on storability and quality of
fresh strawberries". J ournal of Food Science, 1991. 56: p. 4.
[6] Hirano, S. and N. Nagao, "Effects of chitosan, pectic acid, lysozyme, and
chitinase on the growth of several phytopathogens". Agriculture and
Biological Chemistry, 1989. 53: p. 2.
[7] Ikinci, G., S. enel, and H.A.e. al., "Effect of chitosan on a periodontal
pathogen Porphyromonas gingivalis". International J ournal of
Pharmaceutics, 2002. 235(1-2): p. 7.
[8] J eon, Y.J ., P.J . Park, and S.K. Kim, "Antimicrobial effect of
chitooligosaccharides produced by bioreactor". Carbohydrate Polymers,
2001. 44(1): p. 6.
[9] Kendra, D.F. and L.A. Hadwiger, "Characterization of the smallest
chitosan oligomer that is maximally antifungal to Fusarium solani and
elicits pisatin formation in Pisumsativum". Experimental Mycology, 1984.
8(3): p. 6.
[10] Li, H. and T. Yu, "Effect of chitosan on incidence of brown rot, quality
and physiological attributes of postharvest peach fruit". J ournal of the
Science of Food and Agriculture, 2001. 81(2): p. 6.
[11] Liu, X.F., et al., "Antibacterial action of chitosan and carboxymethylated
chitosan". J ournal of Applied Polymer Science, 2001. 79(7): p. 12.
[12] Morimoto, M. and Y. Shigemasa, "Charaterization and bioactivities of
chitin and chitosan regulated by their degree of deacetylation". Kobunshi
Ronbunshu, 1997. 54(54): p. 11.
[13] Park, S.I., et al., "Antifungal coatings on fresh strawberries (Fragaria x
ananassa) to control mold growth during cold storage". J ournal of Food
Science, 2005. 70(4): p. 6.
[14] Shimojoh, M., et al., "Bactericidal effects of chitosan fromsquid pens on
oral streptococci". Nippon Nogeikagaku Kaishi, 1996. 70(7): p. 13.
[15] Simpson, B.K., et al., "Utilization of chitosan for preservation of raw
shrimp (Pandalus borealis)". Food Biotechnology, 1997. 11(1): p. 20.
[16] Tokura, S., et al., "Induction of drug specific antibody and the controlled
release of drug by 6-O-carboxymethyl-chitin". J ournal of Controlled
Release, 1994. 28(1): p. 7.
[17] Ueno, K., et al., "Antimicrobial activity by fractionated chitosan
oligomers". Advances in Chitin Science, 1997. II: p. 6.






















79

EFFECT OF CO
2
UTILIZATION ON THE
GROWTH OF CHLORELLA VULGARIS FOR
FOOD TECHNOLOGY
Nguyen Hoang Minh, Nguyen Thi Thanh Xuan, Dang Kim Hoang, Nguyen Dinh Phu
Chemistry Faculty, Danang University of Technology, The University of Danang
Danang city, Vietnam
nhminh@dut.udn.vn



AbstractCO
2
fixation by Chlorella vulgaris cultivation is one of
positive solutions which contribute to greenhouse effect reduction
and high quality biomass production. C. vulgaris is a potential
microalgae due to its great content of essential nutrition, faster
growth, easier cultivation. More importantly, the microalgae are
able to use high concentration of CO
2
, one of major contributor
to greenhouse effect and global warming, as carbon source for its
growth. Therefore, it is very necessary to study the effect of CO
2

on growth of C. vulgaris. The results indicated that the supplying
CO
2
with 30 mL.min
-1
of flow-rate in culture media could
significantly improve biomass yield and amount of valuable
nutrition (chlorophyll, protein, lipid). Optimization of growth
conditions (CO
2
flow-rate and light intensity) also was carried
out. C. vulgaris is likely to prefer to grow in the media which is
set up with 20 mL.min
-1
CO
2
flow-rate and 4000 Lux light
intensity. Especially, isolated microalgal strains have
morphological properties similar to original C. vulgaris.
Keywords- Chlorella vulgaris; CO
2
; food technology; microalgae;
greenhouse effect; biomass.
I. INTRODUCTION
Greenhouse effect causing from the increase in atmosphere
CO
2
level, has become a serious worldwide problem [1, 2, 3].
According to scientific literature, the CO
2
emission increased
by 3% in 2011, reaching an all-time high of 34 billion tones
[1]. More importantly, major reason for this CO
2
increase is
attributable to human activities which mostly relates to burning
fossil fuels [2]. Therefore, to prevent severe damages due to
excess greenhouse gases, human behavior should be seriously
changed to mitigate CO
2
emission in atmosphere.
Interestingly, many scientists have concentrated on
converting CO
2
into valuable productions through microalgaes
photosynthesis [4]. Among all microalgae, Chlorella vulgaris
is the preferable object. C. vulgaris is a green spherical single
celled fresh water microalgae [5, 6]. It is widely produced and
marketed as a food supplement in many countries, including
China, Japan, Europe and the US. Chlorella is being
considered as a potential source of a wide spectrum of nutrients
(protein, lipid, carotenoid, vitamins, minerals) being effectively
applied in the healthy food market as well as for animal feed
and aquaculture [7]. Many researches also proved that
Chlorella is important as a health promoting factor on many
kinds of disorders such as gastric ulcers, wounds,
constipation [8]. Moreover, Chlorella has fast growth rate,
and withstand high carbon dioxide levels in simple growth
culture. Therefore, C. vulgaris has attracted an increasing
amount of attention.
On the other hands, Chlorellas growth depends on culture
conditions (nutrients, CO
2
, light intensity). Many studies
showed that Chlorella grew better in elevated CO
2
aerating
culture [6, 9]. However, each microalgae requires certain CO
2

concentration for highest biomass production [9]. Therefore,
the aim of the research is to determine effect of CO
2
utilization
on the growth of C. vulgaris.
In this study, C. vulgaris strain is cultured in appropriate
media. Then, microalgaes growth was investigated in cultures
containing different CO
2
concentrations. After harvesting
microalgae with highest productivity, concentrations of
chlorophyll a, protein, lipid in Chlorella were measured and
evaluated. Optimization of CO
2
concentration and light
intensity was performed in this study.
II. MATERIALS AND METHODS
A. Microalgae source
The microalgal strain used in this study is Chlorella
vulgaris, which was bought from Hanoi Institute of
Biotechnology.
B. Investigation of algal growth
The seed culture was grown in 1000 ml Erlenmeyer flasks
containing 500 ml of Antoine medium and subcultured by
transferring 10% (v/v) inoculum at 5-day intervals. The
experiments were exposed to the 20W fluorescent lamps
daylight in 12:12 circadian cycles at room temperature. In
order to investigate the role of CO
2
on Chlorellas growth,
growth culture was performed in 3 different cases: without
aerating used as control; with aerating under air flow rate of
650mL.min
-1
, corresponding to ambient CO
2
; and with CO
2
-air
mixture (30 mL.min
-1
pure CO
2
gas flow-rate mixed with 650
mL.min
-1
air flow-rate, corresponding to 5% CO
2
).
Microalgaes growth was tracked by measuring optical
density (OD) of the culture with a spectrophotometer at 420
nm. Algal growth curve was built up with the OD values
recorded at the same time every day.
80

C. Evaluation of algal quality
1) Measuring of chlorophyll-a concentration by Avigad
Vonshak method [10]
Measuring chlorophyll-a concentration is carried out by
using spectrophotometer, following 3 steps: (1) separation of
microalgae cells from the medium by centrifugation (15000
rpm at 4
O
C for 5 min); (2) extraction of pigments with acetone
and (3) spectrophotometric determination of the concentration
of Chlorophyll-a in the extract at 664 nm and 647 nm
wavelength. Chlorophyll a was determined using the following
formula:
Chlorophyll a (g/ml) = 11.93. E
664
1.93.E
647

2) Measuring of protein concentration by Bradford assay
[11]
After 14 days, microalgae biomass was collected by
centrifugation (Hettich Zentrifugen centrifuge) at the speed of
6000 rpm for 10 min. In this experiment, ultra-sonication was
used to release protein fractions from the whole cell. Then,
protein extraction obtained by centrifugation was measured by
Bradford method [8]. In briefly, 10 L of protein extract was
mixed with 90 L of the diluted reagent. The blue mixtures
absorbance at wavelength 595nm was measured by
spectrophotometer. As the result, protein concentration was
determined based on standard curve plotted from concentration
series (0.1-1.0 mg.mL-1).
3) Measuring of lipid concentration [12]
The lipid fractions were extracted using the method which
was considered the most effective extraction method by Attilio.
Following the previous method, the dried microalgae biomass
samples were pulverized and extracted using mixture of
chloroform: methanol (2:1, v/v). About 50 ml of solvents were
used for every gram of dried sample in each extraction step.
After stirring by ultra - sonication at 20 kHz frequency for 10
min, distilled water was added to the samples which were then
centrifuged at 6000 rpm for 10 min. The lipid soluble in the
chloroform was removed by using micropipette.
D. Optimization of growth conditions [13]
1) Optimization of CO
2
flow rate
The effect of flow rate of CO
2
on biomass production was
investigated by varying CO
2
flow-rates: 20, 40 and 60 mL.
min
-1
, corresponding to 3, 6, and 9.2% CO
2
, respectively. The
experiments were performed in the same conditions: the
temperature of 27
O
C and light intensity of 7000Lux.
2) Optimization of light intensity
The effect of light intensity on biomass productivity was
investigated by varying the light intensity with the values of
4000, 7000 and 13000 Lux. The luminous flux was measured
by Advance Light Meter.
E. Microalgal isolation
Water samples were collected aseptically from two Ham
Nghi and 29-3 lakes at Danang city. The samples were
transferred to 10 tubes containing Antoine nutrient medium and
incubated at 25
o
C under 4,000 Lux illumination and 12h
light/dark photoperiods. After 10 days, 100 L of samples was
inoculated onto petri plates containing Antoine nutrient
medium solidified with 2.0% of bacteriological agar. The
incubation was performed under the same condition as the
tubes for 15 days. Then, colonies from the petri plates were
mixed into distilled water contained in test tube. Finally, algas
morphology was observed by using microscope with a
magnification of 40x.
III. RESULTS AND DISCUSSION
A. The role of CO
2
on the growth of Chlorella vulgaris
Generally, phototrophic microalgal growth requires a
supply of carbon dioxide as a carbon source [3]. Microalgae
grown in elevated CO
2
environments typically exhibit
increased rates of photosynthesis and biomass production. In
this experiment (Fig. 1), without aeration, microalgaes growth
rate was likely to be very weak. With aeration, Chlorella cells
slowly developed during 11 days batch cultivation. Since
twelfth day, they quickly developed and achieved the maximal
value after 20 days. Interestingly, C. vulgariss growth rate in
5% (v/v) CO
2
aerated culture was exponentially elevated after
first 5 days, and reached to value similar to those obtained
under ambient CO
2
after 12 days. The results implied that
Chlorellas growth rate depends on CO
2
concentration in the
culturing medium. C. vulgaris exhibited an increase in
performance under elevated CO
2
condition. Senthil also
cultured C. vulgaris in medium supplying with varying
concentration of CO
2
. Data revealed that the highest
chlorophyll and biomass, which were 60 and 20 times more
than that of C. vulgaris at ambient CO
2
, were recorded at 6%
CO
2
level [9]. Such CO
2
fixation by photoautotrophic algal
culture has the great potential to not only produce high biomass
productivity, but also diminish the release of CO
2
into the
atmosphere, helping alleviate the trend toward global warming.

Figure 1. C. vulgariss growth at 3 different cases: without air, with air and
with CO2-air mixture
Besides, in order to produce a high valuable microalgae
biomass production serving for food technology, CO
2

concentration in the culture should be controlled seriously. Too
high CO
2
concentration could produces inhibition and, on the
other hand, too low concentration could limits growth [6].
These thresholds vary from one species to another and are not
yet adequately known. Therefore, it is necessary to investigate
81

the role of CO
2
and optimize CO
2
concentration in C.
vulgariss growth.
B. Evaluation of algal quality
1) Measuring of chlorophyll-a concentration
Chlorophyll- a is a primary photosynthetic pigment in all
algae, in which Chlorella sp. is considered as a microalgae
strain containing highest Chlorophyll amount. This is an
indicator exhibiting microalgaes growth. In food technology,
chlorophyll- a is used as food and pharmaceutical colorants. It
also can exhibit health promoting activities. Chlorophyll has
wound healing and tissue regeneration properties and acts like
a natural antibiotic [7]. It also plays a significant role in cancer
prevention. For these reason, it is necessary to evaluate
influence of CO
2
in Chlorophyll- a formation of C. vulgaris.
After 7 days, chlorophyll- a obtained from 5% (v/v) CO
2

aerated culture (0.93 g.mL
-1
) was significantly higher than
that under ambient CO
2
and control (0.20- 0.26 g.mL
-1
)
(Table I). The result indicated that ambient CO
2
was not
enough to support growth of C. vulgaris. Higher CO
2
recorded
vigorous growth. It implied that supply of elevated CO
2
helped
microalgae form more chloroplasts to effectively convert and
use light energy.
2) Measuring of protein concentration
In order to obtain valuable algal biomass applied for food
technology, protein concentration should consist of high ratio
in alga. According to Dang Dinh Kim et al. [14], protein
content of C. vulgaris is high, about 40 - 60 %. Razif Harun
reported that the protein concentration is about 51 58%,
compared to dried weight [15]. On the other hand, Hee sun lee
et al assumed that highest protein amount of Chlorella obtained
is 60.6% [16].
In this study, protein obtained under elevated CO
2
was
quite high (53%) and similar to Razif Haruns results. Under
ambient CO
2
, protein content was 31% only (table I). This
reveals that the increase in the microalgal performance at
higher levels of CO
2
may be attributable to the enhanced
availability of dissolved CO
2
.
3) Measuring of lipid concentration
Apart from protein, lipid concentration is also important
target which should be evaluated. Previous studies showed that
polyunsaturated fatty acids of algae could offer many health
benefits, such as hypercholesterolemia, hyperlipidaemia and
atherosclerosis [5].
C. vulgaris cultured under 5% CO
2
showed significantly
higher amount of lipid value (13.25 %), compared to other
culture conditions (4- 9 %). This was similar to finding of
Dang Dinh Kim who found that the lipid content of Chlorella
was about 10-15%.
On the other hand, the collected lipid from CO
2
bubbled
sample was analyzed by HPLC. The result showed that the
peak of triglyceride appeared at approximately 10.42 minutes
retention time and the triglyceride in the mixture presents in
quite high concentration (data not shown). Moreover, the
triglycerides composition analysis results taken by GC-MS
revealed that there are 2 typical peaks of methylester of fatty
acids palmitic acid (C16:0) and oleic acid (C18:1). These 2
types of fatty acids are valuable for pharmaceutical purpose.
TABLE I: Chlorophyll-a, protein, lipid concentration of C. vulgaris were
measured in the 3 different aerated cultures.

In conclusion, C. vulgaris grew well under 5% CO
2
with
optimum nutritional compositions. Chlorophyll-a, protein, lipid
content were dependent on the amount of CO
2
in the culture,
meaning that it is possible to modify the nutritional contents of
Chlorella by changing culture conditions.
C. Optimization of growth conditions
Previous studies indicated that it is very necessary to
increase both CO
2
concentration and light intensity to obtain
high biomass production [17]. In this study, the highest dry
weight of C. vulgaris was obtained under growth culture
supplied with 3% of CO
2
and 4000 Lux of light intensity. Algal
biomass decreased when CO
2
and light intensity were higher
than optimal values.
The growth response of C. vulgaris was studied under
varying concentrations of carbon dioxide (3, 6, 9 %). The result
of table II indicated that at higher CO
2
concentration, algals
biomass began to slowly reduce, but they continued to remain
greater than those observed in ambient air (0.52 mg.mL
-1
). This
is in agreement with Senthil Chinnasamy et al. [9], in which
they explained that increase in the algal performance at higher
level of CO
2
may be because of the enhanced availability of
dissolved CO
2
and the toxicity of a very high CO
2
concentration is due to the lowering of pH.
TABLE II: Effect of CO2 concentration on biomass production of C.
vulgaris


Effect of light intensity on algal growth was showed in
table III. When using higher light intensity (7000 Lux),
Chlorellas dried biomass weight slightly reduced. However,
13,000 Lux light gave strong inhibition on algal growth. The
result of light intensity in this study (table III) seems to be
contradictory with that of Siranee et al. who assumed that algal
culture at 8000 Lux gave significantly higher amount of algal
biomass, compared to those under 3000 and 5000 Lux [18].
Moreover, according to findings of Anondho Wijanarko et al.
[4], illumination using UV light indicates a decrease in cell
amount. This susceptibility to photo inhibition is caused by the
blockage of chloroplast-encoded protein synthesis, and a severe
damage occurs in photo-system II reaction center of micro
82

algae chlorophyll a. This implied that using very high light
intensity (13000 Lux) in this experiment could cause
phenomenon of photo inhibition.
TABLE III: Effect of light intensity on biomass production of C. vulgaris.


D. Microalgae isolation at Danang city
Microalgae isolation from natural rivers at Danang city
plays an important role in discovering and possessing novel
microalgae strains which are able to produce valuable biomass
and tolerate to high CO
2
concentration. Many reports showed
the possibility of the isolation of such naturally occurring algal
forms [9]. Some microalgae could quickly grow under 40% of
CO
2
. Komada et al. reported that Chlorella littorale developed
better at 30
o
C and 20% CO
2
concentration and another marine
alga could perform fast even at 60% CO
2
[19]. This shows the
changes of using such organisms to produce high-quality
biomass.
In this experiment, natural strains isolated from Ham
Nghi and 29-3 lakes have similar properties with original
Chlorella vulgaris strains morphology, such as green color,
spherical. Since we do not know whether they are Chlorella
vulgaris or not, further study will focus on identifying isolated
strains genotype.

Figure 2. Microagal morphologies were observed under microscope with a
40x magnification
IV. CONCLUSION
In this study, Chlorella vulgaris could grow better under
elevated CO
2
. The highest chlorophyll-a, protein, lipid
concentration were recorded under 5% of CO
2
, compared to
ambient CO
2
(0.036%). The highest dry weight of C. vulgaris
was obtained under growth culture supplied with 3% of CO
2

and 4000 Lux of light intensity. Furthermore, the first result of
microalgae isolation from natural lakes at Danang city
contributes to creating bank of microalgal strains which could
not only produce high value molecules but also mitigate CO
2

emission in atmosphere. More importantly, this study is the
central foundation for carrying out further studies, such as
investigation of microalgae harvesting, establishment of model
using CO
2
flue gas from industrial parks for large-scale
Chlorella vulgaris cultivation at Danang city.
REFERENCES
[1] J os G. J . Olivier, Greet J anssens-Maenhout, J eroen A. H. Peters. 2012.
Trends in global CO2 emissions: 2012 report. PBL Nethrlands
Environmental Assessment Agency.
[2] Nhan T. Nguyen, Minh Ha- Duong, The potential for mitigation of CO2
emissions in Vietnams Power sector. Depocen. Working paper series
No. 2009/22.
[3] Nguyen Thai Hoa, Kei Gomi and Yuzuru Matsuoka. A scenario for
sustainable low-carbon development in Vietnamtowards 2030
[4] Anondho W., Dianursanti. 2004. Effect of photoperiodicity on CO2
fixation by Chlorella vulgaris Buitenzorg in bubble column
photobioreactor for food supplement production. Makara technology,
vol 8, No. 2 35-43
[5] Becke W. 2004. Microalgae aquaculture in Handbook of Microalgae
culture. Richmond A. p: 380-392.
[6] Fadhil M. Salih. 2011. Microalgae tolerance to high concentrations of
carbon dioxide: A review. J ournal of Environmental Protection, 2, 648-
654.
[7] Gouveia. L A. P. Batista, Souse I. 2008. Microalgae in Novel Food
Products. Food Chemistry research development. Nova Science
Publisher. Chapter 2.pp:1-37.
[8] Gouveia L. , B. P. Nobre, F. M. Marcelo, S. Mrejen, M. T. Cardoso, A.
F. Palavra, R. L. Mendes. 2007. Functional food oil coloured by
pigments extracted from microalgae with supercritical CO2. Elsevier,
Food Chemistry 101:717-723.
[9] Senthil Chinnasamy, Balasubramanian Ramakrishnan and Keshav C.
Das. 2009. Biomass Production Potential of a Wastewater Alga
Chlorella vulgaris ARC 1 under Elevated Levels of CO2 and
Temperature. Internation J ournal of Molecular Sciences. Int. J . Mol.
Sci. 2009, 10, 518-532
[10] Vonshak A. (1997), Preface in: spirulina plantesis: physiology, cell-
biology and biotechnology.
[11] Bradford, M. M. 1976. A rapid and Sensitive-Method for the
Quantitation of Laemmli, U. K.1970. Nature. Vol 227. pp: 680-685.
[12] Attilio Converti. 2009. Effect of temperature and nitrogen
concentration on the growth and lipid content of Mannochloropsis
oculata and Chlorella vulgatis for biodiesel production. Chemical
Engineering and Processing; Process Intensification, Elsevier, 48, 1148-
1151
[13] Nguyen T. T. Xuan, Dang K. H. , Nguyen M. T., Le T. B. Y. , Nguyen
H. M. .2012, A study of Chlorella vulgaris microalgae cultivatoin for
biodiesel production. Science and Technology. The University of
Danang. Vol. 8[57].
[14] ng nh Kim. 1999. Cng ngh sinh hc vi to, NXB Nng nghip
H Ni.
[15] H. Razif, S. Manjinder. 2010. Bioprogress engineering of microalgae to
produce a variety of consumer products, J . Elsevier, Renewable and
sustainable Energy reviews 14:1037-1047
[16] Hee Sun Lee, Hoon J ung Park and Mi Kyung Kim. 2008. Effect of
Chlorella vulgaris on lipid metabolismin Wistar rats fed high fat diet.
The Korean Nutrition Society. 204-210.
[17] D. Sasi, E.E. Powel and G.A. Hill. 2009. Effect of light intensity and
CO2 on growth of Chlorella vulgaris. University of Saskatchewan,
Saskatoon, Saskatchewan, S7N 5A9mCanada.
[18] Siranee S. , Preeda P. 2007. Nutrient recycling by Chlorella vulgaris
fromseptage effluent of the Bangkok city, Thailand. ScienceAsia 33:
293-299
[19] Komada, M. , Ikemoto, H. 1993. A new species of highly CO2
tolerant fast growing marine microalga suitable for high density
culture. J . Mar. Biotechnol. 1. 21-25

83

EFFECT OF CARBON SOURCES ON
PROLIFERRATION OF ZEDOARY (CURCUMA
ZEDOARIA ROSCOE) CELL SUSPENSION
CULTURES
Vo Chau Tuan, Tran Quang Dan
Faculty of Biology and Environmental Science
College of Education, Da Nang University,
Da Nang, Vietnam
Email: vochautuan@gmail.com



AbstractZedoary (Curcuma zedoaria Roscoe) plant, a vegetatively
propagated species of the Zingiberaceae family, is an aromatic
herbaceous plant with a rhizome growing mainly in China, Vietnam,
India, J apan. Cell suspension culture is the prior choice in large-scale
production of some valuable secondary metabolites thanks to its great
advantages over other technologies. We report here the effect of
different sugars on the growth of cells in a shake liquid culture
condition. As a single carbohydrate source in medium, sucrose
exhibited a better effect on the cellular growth when compared with
fructose or glucose. Zedoary cells reached a maximum biomass of
10.44 g fresh cell (approximately 0.66 g of dry cell) after 14 days of
culture at sucrose concentration of 30 g/L. The combination of two
hexoses (glucose and fructose) at different concentrations were not
effective for proliferation of cell suspensions. Results from this study
might be a foundation for further studies on Curcuma zedoaria
Roscoe in order to produce its secondary metabolites in a large scale.
Keywords- Callus, cell biomass, cell suspension, Curcuma
zedoaria, medicinal plant
I. INTRODUCTION
Zedoary (Curcuma zedoaria Roscoe) plant, a vegetatively
propagated species of the Zingiberaceae family, is an aromatic
herbaceous plant with a rhizome growing mainly in China,
Vietnam, India, and Japan [7]. Zedoary is a valuable medicinal
plant; the essential oil obtained from rhizome has been reported
to have antimicrobial activity and be clinically used in the
treatment of cervical cancer; the water extract of zedoary
demonstrated antimutagenic activity [24]. It has been also used
for treatment of stomach diseases, hepatoprotection [22],
treatment of blood stagnation, and promoting menstruation as a
traditional medicine in Asia [15]. Furthermore, the zedoary has
anti-inflammatory potency related to its antioxidant effects
[22].
The presence of valuable metabolites in plants has
stimulated interests of the industries in the fields of
pharmaceuticals, agrochemicals, nutrition and cosmetics. The
bulk of the market products, such as secondary metabolites
from higher plants, are collected from plants growing in the
wild or from field cultivated sources [9]. In recent decades,
interests in chemopreventive products from natural plants has
grown rapidly [14]. However, there are many challenges in the
production of plant-based medicines, many of which put both
the consumers and the plant populations at risk [2]. Therefore,
an alternative, economically viable, and environmentally
sustainable production source of desired secondary metabolites
is of a great interest. In this regard, plant cell cultures can be an
attractive alternative as a production system, as well as a model
system, to study the regulation of natural product biosynthesis
in plants so as to ultimately increase yields [9].
Sugar plays a central role in plant life. They enter the
metabolism pathways and the transformation of energy, which
are required for growth of cell. In plant tissue cultures, sugar
serves as a carbohydrate supply to provide an optimal culture
condition for cells. For most of plant cell lines, sucrose is an
important carbon and energy source, and its initial
concentration can affect such parameters as growth rate and the
yield of secondary metabolites [11]. Utilization of
monosaccharides by plant cells has been extensively
investigated. Many types of plant cells and tissue cultures
commonly utilize sucrose, glucose, fructose and some other
monosaccharides equally when they are added to the culture
medium as a sole carbon source [8]. Even though
carbohydrates are of prime importance for cell growth,
maintenance and differentiation in vitro, the fundamental
aspects of carbon utilization and metabolism in cell and tissue
cultures have yet to be fully understood [3].
In this study, we investigated the effect of three kinds of
sugar (sucrose, glucose, fructose) in separation or in
combinations on the cell growth of Curcuma zedoaria cell
suspension culture for improvement of culture medium and
further aims for secondary metabolites production in large-
scale.
84

II. MATERIALS AND METHODS
A. Callus formation
Leaf-base explants of 0.50.5 cm were excised from
zedoary plants in vitro [13]. The explants were placed on the
Murashigie and Skoog (MS) solid medium supplemented with
2% (w/v) sucrose, 0.25-4.0 mg/L 2,4-dichorophenoxyacetic
acid (2,4-D), and 0.25-4.0 mg/L benzyladenine (BA) for callus
induction. White and soft calli (primary calli) were transferred
were transferred on the MS medium supplemented with 0.5-4.0
mg/L 2,4-D, and 0.5-4.0 mg/L BA for secondary calli
induction. The secondary calli which were light yellow, hard
and friable inducted on the MS medium supplemented with 0.5
mg/L 2,4-2,4-D and 0.5 mg/L BA after for weeks of culture
were maintained in fresh medium with the same composition
every two weeks. The pH of the medium was adjusted to 5.8,
and then it was autoclaved at 121
o
C for 15 min. The cultures
were incubated at 252
o
C under an intensity of 2,000-3,000 lux
with a photoperiod of 10-h day light.
B. Initiation of cell suspension cultures and test carbon
sources
Cell suspension cultures were initiated from callus which
were light yellow, hard and friable (Fig 1A) by culturing callus
(3 g) in 250 mL Erlenmeyer flasks containing 50 mL of liquid
medium with the same composition as the callus inducing
medium under a rotary shaker at 120 rpm, for 10 days until a
suspension of free cells formed. From the resulting cell
suspensions, 3 g of cells was transferred on the MS medium
supplemented with 1.5 mg/L 2,4-D and 0.5 mg/L BA but
varied carbon concentration and sources: sucrose, glucose,
fructose (20-70 g/ L) or combination of glucose and fructose
rates (20/40; 30/30 and 40/20 g/L) for biomass production. The
flasks were placed on a rotary shaker (Fig 1B ) at 120 rpm
under the same physical conditions as described for callus
cultures except an intensity of 500 lux.
C. Growth measurement
Samples were obtained after 14 days of culture to
determine the cell biomass in both fresh and dry weights. For
measurement of fresh cell weight (Fig 2A), the cells in the
suspension culture were filtered, washed with distilled water,
collected, and weighed. The dry cell weight was determined by
drying the fresh cell biomass at 50
o
C until a constant weight
was attained (Fig 2B).
D. Statistical analysis
The experiments of cell suspension culture were conducted
with a minimum of three replicates. All experiments were
repeated three times. The data were analyzed by mean
standard error followed by comparison of the means with the
Duncans test at p<0.05.
III. RESULTS AND DISCUSSION
A. Effect of sucrose on cell growth
As shown in Table I the initial concentration of sucrose
significantly affect the biomass accumulation of the cell
culture, highest fresh cell weight were attained in media
containing sucrose concentration of 30 g/L, the cell biomass
reached with 10.44 g of fresh cell weight (approximately 0.66 g
dry cell weight). However, despite sucrose being an
indisputably important carbon and energy source, increasing
its concentration from 40 to 60 g/L resulted in fresh cell
weight reductions and significantly reduced at sucrose
concentration of 70 g/L. This decline in cell growth might be
attributed to the inhibition of nutrient uptake as the osmotic
potential was enhanced and the medium became more
viscous. Do et al. (1991) have shown that in Vitis vinifera a
higher concentration of sucrose can act as an osmotic
agent, with mannitol having a similar effect on growth [4].
Additionally, this retardation in growth could be caused by
a cessation in the cell cycle when nutrients are limited and
sucrose concentrations are higher [6, 26].
TABLE I. EFFECT OF DIFFERENT INITIAL SUCROSE
CONCENTRATIONS ON THE GROWTH OF
ZEDOARY CELL SUSPENSIONS
Sucrose concentrations
(g/L)
Cell biomass (g)
fresh weight dry weight
20 7.22
c
0.55
d
30 10.44
a
0.66
b
40 8.85
b
0.64
b
50 8.80
b
0.65
b
60 8.75
b
0.70
a
70 6.75
d
0.60
d

Different letters indicate significantly different means using
Duncans test (p<0.05)

Even though at high concentrations of sucrose (40-60 g/L),
the fresh cell biomass obtained was lower than compared with
the response to 30 g/L sucrose, but cells tend to increase the
dry biomass accumulation. At 60 g/L sucrose concentration,
fresh weight of cells reached only 8.75 g, lower than on the
medium with 30 g/L sucrose (reaching 10.44 g fresh cell), but
the dry weight was higher (0.70 g dry cell compared to 0.66 g
dry cell). These results are in agreement with some previous
reports for other plant species such as Solanum chrysotrichum
[25], Centella asiatica [20]. The role of sucrose in promoting
dry cell weight can be explained by its effect on tubulin, a
ubiquitous protein responsible for growth and development of
cells. Tubulin controls cell shape and chromosome separation
and provides cytoskeletal tracks for intracellular vesicular
transport via Tual and Incw1 genes. In the absence of sucrose,
these genes are not expressed but addition of sucrose turns
on their expression [20].
B. Effect of glucose on cell growth
Results presented in the Table II shows that the change of
initial glucose concentrations had also affected on cell growth.
When the sugar concentration is in the range of 20-30 g/L, the
biomass accumulation of cells is relatively low. Increasing the
85

concentration of sugar to the range of 40 to 60 g/L, cell
biomass accumulation increased, the cell biomass reached
maximum value with 7.32 g of fresh cell weight
(approximately 0.65 g dry cell weight) at sugar concentration.
of 60 g/L after 14 days of culture. The growth of cells reduced
with increasing sugar concentration up to 70 g/L. Generally, at
different sugar concentrations, cell growth in medium
containing glucose was remarkably lower than that in medium
supplemented with sucrose (Table I).

TABLE II. EFFECT OF DIFFERENT INITIAL GLUCOSE
CONCENTRATIONS ON THE GROWTH
OF ZEDOARY CELL SUSPENSIONS
Glucose concentrations
(g/L)
Cell biomass (g)
fresh weight dry weight
20 3.15
e
0,22
d
30 6.05
d
0,42
c
40 7.12
c
0,53
b
50 7.20
b
0,62
a
60 7.32
a
0,65
a
70 6.85
c
0,57
e


Different letters indicate significantly different means using
Duncans test (p<0.05)

The carbon source in plant cell suspension media are
usually supplied as carbohydrates, with the most common
sugars being sucrose and glucose. Several previous reports
showed that glucose was suitable carbon source in cell and
tissue cultures [8, 23, 10]. Our study showed that glucose was
less effective for proliferation of zedoary cells compared with
sucrose. Similar results were reported by Ling et al (2008)
working with Ficus deltoidea [12] and Shinde et al (2009) with
Psoralea corylifolia [21].
C. Effect of fructose on cell growth
The effects of different fructose concentrations of 20,
30, 40, 50, 60 and 70 g/L on the growth of zedoary cell were
shown in Table III. The initial concentration of fructose was
influence on the accumulation of cell biomass. The rise of the
fructose concentration from 20 to 60 g/L increased the cell
biomasses for both fresh and dry weight. Zedoary cells reached
a maximal biomass of 8.40 g fresh cell (approximately 0.62 g
dry cell) after 14 days of culture at the initial fructose
concentration of 60 g/L. However, as the concentration of
fructose increased to 70g/L, the growth of cells was inhibited
(only reached 7.65 g fresh cell and 0.57 g dry cell).
The absorption lines of plant cells in culture depends on the
sensitivity of each species of plants to the carbon source or on
differences in the formation of products of metabolic processes
[16]. Soluble sugars such as glucose, fructose and sucrose are
often referred to as the substances responsible for osmotic
adjustment in tissues under osmotic stress and the best carbon
sources for the growth of most plant cell cultures [12].
Abdullah et al. (1998) showed that 3% and 5% of fructose
and mannitol were able to promote growth of Morinda elliptica
cell cultures [1]. Ling et al. (2008) working with Ficus
deltoidea found that fructose treatment gave the highest growth
increase followed by sucrose, sorbitol and glucose,
respectively [12]. The kinetics of sugar uptake by maize
endosperm suspension cultures resembled those observed for
asparagus cell cultures, in that fructose was transported most
rapidly, followed by glucose and then sucrose [5]. The
consumption of fructose by Solanum. eleagnifolium suspension
culture was found to be 27% higher than sucrose [19].
Although fructose managed to produce the highest growth
rate, it was anticipated that at higher concentration, it might
have a growth-inhibitory effect. Such inhibitory effect at higher
concentration of fructose was well document in Nicotiana
tabacum and Cinchona succirubrum with using high fructose
concentration [19].
TABLE III. EFFECT OF DIFFERENT INITIAL FRUCTOSE
CONCENTRATIONS ON THE GROWTH
OF ZEDOARY CELL SUSPENSIONS
Glucose
concentrations
(g/L)
Cell biomass (g)
fresh weight dry weight
20 3.15
e
0,22
d
30 6.05
d
0,42
c
40 7.12
c
0,53
b
50 7.20
b
0,62
a
60 7.32
a
0,65
a
70 6.85
c
0,57
e

Different letters indicate significantly different means using
Duncans test (p<0.05)
D. Effect of glucose and fructose in combinations of different
concentrations on cell growth
Table IV indicated that the combinations of glucose and
fructose at different concentrations have no effect to promote
on the cell growth. The cell biomass reached a highest value
with 6.95 g of fresh cell (approximately 0.65 g dry cell) in
media containing of 40 g/L fructose and 20 g/L glucose.
Generally, cell biomass was lower than cells cultured on the
medium containing sole sugar after 14 days of culture.
TABLE IV. EFFECT OF GLUCOSE AND FRUCTOSE IN
COMBINATIONS OF CONCENTRATION ON
GROWTH OF ZEDOARY CELL SUSPENSIONS

Glucose and fructose rates
(g/L)
Cell biomass (g)
Glucose Fructose fresh weight dry weight
40 20 6.60
b
0.60
b
30 30 6.75
b
0.61
b

20 40 6.95
a
0.65
a

Different letters indicate significantly different means using
Duncans test (p<0.05)

86








Figure 1. Callus and suspension cell of zedoary. A Light yellow, compact
and friable callus, B Suspension cell





Figure 2. Cell biomass of zedoary. A: Fresh cell biomass, B: Dry cell
biomass
According to Duong Tan Nhut et al. (2006), the
combination of two hexoses (glucose and fructose) at different
concentrations, the best proliferation of cell was obtained at the
combination of 30 g/L glucose and 30 g/L fructose, higher than
compared with single sugar [18]. Results of our study is
completely opposite. Thus, to achieve the optimum cell
growth, different plants required different carbon sources due
to the different enzymatic metabolism. Moreover, the
availability of sugars and its derivatives would initiate
different responses and would affect plant metabolism, growth
and development [12].
IV. CONCLUSION
The suspension culture offers many advantages to
scale-up production of secondary metabolites in plant cells of
interest. In this study, sucrose exhibited a better growth of cell
when compared with fructose or glucose. The combination of
two hexoses (glucose and fructose) at different concentrations
were not effective for proliferation of cell suspensions. Results
from this study might be a well established foundation for
further studies on Curcuma zedoaria Roscoe in order to serve
as a potential source for secondary metabolites production in
large scale.
ACKNOWLEDGMENT
This study was supported by a grant from the Basic
Research Program in Natural Science of the Vietnamese
Ministry of Science and Technology (2006-2008).
REFERENCES
[1] Abdullah MA, Lajis NH, Ali AM, Marziah M, Sinskey AJ , Rha CK
(2005). Issues in plant cell culture engineering for enhancement of
productivity. Development of chemical engineering process 13: 1-15.
[2] Cole J B, Saxena PK, Murch SJ (2007). Medicinal biotechnology in the
genus scutellari. In Vitro Cell Dev Biol-plant 43: 318-327
[3] De Mello MO, De Campos AF, Murilom(2001). Sucrose metabolizing
enzymes in cell suspension cultures of Bauhinia forficata, Curcuma
zedoaria and Phaseolus vulgaris. Pesq. agropec. Bras 36(9): 1085-
1092.
[4] Do CB, Cormier F (1991). Effects of low nitrate and high sugar
concentrations on anthocyanin content and composition of grape (Vitis
vinifera L) cell suspension. Plant Cell Rep 9: 500-504.
[5] Felker FC, Miernyk J A, Crawford CG (1989). Characterization of Maize
endosperm-derived suspension cells throughout the culture cycle and
enhancement of tissue friability. Plant Cell, Tissue and Organ Culture
18: 153-165.
[6] Gould AR, Everett NP, Wang TL, Street HE (1981) Studies on the
control of cell cycle in cultured plant cells: Effect of nutrient limitation
and nutrient starvation. Protoplasma 106:1-13 .
[7] Hong CH, Noh MS, Lee WY, Lee SK (2002). Inhibitory effects of
natural sesquiterpenoids isolated from the rhizomes of Curcuma
zedoaria on prostaglandin E2 and nitric oxide production. Planta Med
68: 545-547.
[8] Kato A, Tohoyama H, J oho M, Inouhe M (2007). Different effects of
galactose and mannose on cell proliferation and intracellular soluble
sugar levels in Vigna angularis suspension cultures. J Plant Res
120:713-719.
[9] Kirakosyan A, Kaufman PB (2009). Recent advances in plant
biotechnology. Springer Dordrecht Heidelberg London New York, USA.
[10] Koch K.E. (2004) Sucrose metabolism: regulatory mechanisms and
pivotal roles in sugars using and plant development. Curr Opin Plant
Biol 7: 235-246.
[11] Lee EJ , Mobin M, Hahn EJ , Paek KY (2006). Effects of sucrose,
inoculum density, auxins, and aeration volume on cell growth of
gymnema sylvestre. .Journal of Plant Biology 49(6): 427-431.
[12] Ling OS, Kiong ALP, Hussein S (2008). Establishment and
optimisation of growth parameters for cell suspension cultures of Ficus
deltoidea. American-Eurasian Journal of Sustainable Agriculture 2(1):
38-49.
[13] Loc NH, Duc DT, Kwon TH, Yang MS (2005). Micropropagation of
zedoary (Curcuma zedoaria Roscoe)-a valuable medicinal plant. Plant
Cell Tiss Organ Cult 81:119-122.
[14] Matkowski A (2008). Plant in vitro culture for the production of
antioxidants-A review. Biotechnology advances 26: 548-560.
[15] Matsuda H, Ninomiya T, Yoshikawa M (1998). Inhibitory effect and
action mechanism of sesquiterpenes from Zedoariae rhizoma on D-
galactosamine/lipopolysaccharide-induced liver injury. Bioorg Med
ChemLett 8: 339-344.
[16] Mello MO, Dias CTS, Amarai AF (2001a). Growth of Bauhinia
forficata, Curcuma zedoaria, and Phaseolus vulgaris cell suspension
cultures with carbon sources. Sci Agric 58: 481-485.
[17] Murashige T, Skoog F (1962) A revised mediumfor rapid growth and
bioassays with tobacco tissue culture. Physiol Plant 15: 473-497.
[18] Duong Tan Nhut, Nguyen Trinh Tramand Nguyen Trinh Don (2006).
Effect of sucrose, glucose and fructose on proliferation of Hymalaya
Yew (Taxus wallichiana Zucc.) cell suspension cultures. A potential
source for Taxol production. Proceedings of International Workshop on
Biotechnology in Agriculture, Nong Lam University Ho Chi Minh City,
Vietnam142-145.
[19] Nigra, HM, Alvarez MA, Giolietti AM (1990). Effect of carbon and
nitrogen sources on growth and solasodine production in batch
suspension cultures of Solanum eleagnifolium Cav. Plant Cell, Tissue
and Organ Culture 21: 55-60.
[20] Omar R, Abdullah MA, Hasan MA, Marziah M. (2004). Development of
growth medium for Centenlla asiatica cell culture via response surface
methodology. American Journal of Applied Sciences 1(3): 215-219.
[21] Shinde AN, Malpathak N, Fulzele DP (2009). Studied enhancement
strategies for phytoestrogens production in shake flasks by suspension
culture of Psoralea corylifolia. BioresourceTechnology 100: 1833-
1839.
B A
A B
87

[22] Yoshioka T, Fuji E, Endo M (1998). Anti-inflammatory potency of
dehydrocurdione, a zedoary-derived sesquiterpene. Inflamm Res 47:
476-481.
[22] Sturm A, Tang GQ (1999). The sucrose cleaving enzymes of plants are
crucial for development, growth and carbon partitioning. Trends Plant
Sci 4: 401-407.
[23] Syu WJ r, Shen CC, Don MJ , Ou J C, Lee GH, Sun CM (1998).
Cytotoxicity of curcuminoids and some novel compounds from
Curcuma zedoaria. J Nat Prod 61: 1531-1534.
[24] Villarreal ML, Arias C, Feria-Velasco A, Ramirez OT, Quintero R.
(1997). Cell suspension culture of Solanum chrysotrichum (Schldl) - A
plant producing an antifungal spirostanol saponin. Plant Cell Tissue and
Organ Culture 50:39-44.
[25] Wu CH, Dewir YH, Hahn EJ , Paek KY (2006). Optimization of
culturing conditions for the production of biomass and phenolics from
adventitious roots of Echinacea angustifofia. J Plant Biol 49:193-199.
88

Use of BMWP
VIET
and ASPT indices as bioindicators
for testing water quality of rivers in Danang city


Nguyen Van Khanh, Vo Van Minh, Kieu Thi Kinh
Danang University of Education, Danang, Vietnam
vankhanhsk23@gmail.com
Tran Duy Vinh
(1)
, Phan Thi Hien
(2)

(1) Okayama University, Japan
(2) Danang Dept. of Natural Resources and Environment,
Vietnam


Abstract As a consequence of rapid urbanization and
industrialization, pollutants have increased in both quantity and
diversity in the water environment. Monitoring plays a key role
in prevention of negative impacts from potential environmental
risks. Currently, physical and chemical analysis is considered as
the popular methods for monitoring. Yet it is obvious that these
methods require costly expenditure, high sampling frequency
and difficult to evaluate the environmental status as well as to
forecast the movement of chemical concentration. Using macro-
invertebrates as bio-indicators has been evaluated as the
advantageous method for testing water quality since the
processed data can mirror water quality based on the ecological
and biological characteristics. Biological indices named BMWP
(biological monitoring working party) and ASPT (average score
per taxon) have been initialized and adopted in temperature-
countries and have shown remarkable advantages in assessment
of water quality. In this paper, we present the initial results in
applying BMWP and ASPT to monitor water quality of rivers in
Central Region during 5 years from 2007 to 2011.

Keywords: BMWP, ASPT, macroinvertebrate, water quality,
biomonitoring.
I. INTRODUCTION
Human activities have led to series adverse impacts the
environment. It is widely recognized that the increase of
chemical concentration in water resulted by receiving effluents
from industrial production, agricultural fertilizers, and
domestic wastes have significantly contributed to
environmental quality deterioration. Environmental risks
should be prevented in order to avoid negative impacts to
ecological health and human by betimes identifying the state of
environmental quality. In terms of monitoring methods,
currently, popular methods for testing water quality are majorly
based on physical and chemical analyses [7], [9]. Conversely,
these methods require costly expenses and high sampling
frequency. In addition, effects of pollutants on species
composition and community structure in aquatic ecosystem
could not correctly be reflected and it is difficult to forecast
aquatic-environmental changes [1], [2], [7], [9], [10].
One of the most practical methods to assess water quality
that has highly been considered as an advantageous type is
biological indices. Biological index called BMWP (Biological
Monitoring Working Party) has been originated from Great
Britain and adapted in Spain and Brazil, associated with ASPT
(average score per taxon) to evaluate water quality of streams
[10]. The principle of this approach is based on presence or
absence of macroinvertebrate families and their relative
proportion may be indicative of the state of the water bodies
[9], [11]. The mentioned biological indices have shown
significant advantages in water quality evaluation, great
abundance, for example, restricted mobility, predictable
community structure, easy and cost-effective sampling and
identification [9].
Vietnam has a continuous and stable economic growth rate,
the remarkable improvement of social lives and living
standard, however, negative impacts from urbanization on
aquatic life may be inevitable and Danang city is not an
exception. So far, in Vietnam, a practical procedure associated
with data analysis and score system in term of biological
indices has been established that called the BMWP
VIET
[9].
This paper aims at presenting the research results on the
BMWP and ASPT for testing water quality of rivers in the
period of 5 years from 2007 to 2011 and evaluates the
applicability of the BMWP index.
II. METHODOLOGY
In the period of 5 years from 2007 to 2011, researches were
carried out using the BMWP and ASPT for testing water
quality of Phu Loc river, Han river, Cu De river and Cau Do
Tuy Loan river system in Danang city. Macroinvertebrates
were sampled by using dredge bucket (500 nm mesh) and
pond-net (500 nm mesh). Kick sampling technique was utilized
from shallow sites, then the pond net was held vertically on the
river bed against the current to catch macroinvertebrates [9].
From deeper sites, sweep sampling in which pond net and
dredge bucket were used. In some cases, we searched the
habitat of macroinvertebrates when the above basis techniques
could not be done. After caught, samples were classified and
stored in alcohol 70 degrees in the university laboratory. In the
laboratory, macroinvertebrate samples were taxonomically
identified as separated groups or families allocated a score
between 1 and 10 in accordance to the method based on the
BMWP
VIET
system established by Nguyen Xuan Quynh, Clive
Pinder, Steve Tilling and Mai Dinh Yen [8], [9]. The scores for
macroinvertebrate samples were then summed to give the
BMWP score. After that, the Average Score Per Taxon (ASPT)
were so calculated by dividing the BMWP score by the total
number of taxa (families) in the sample [9].

89


Figure 1. Map of research sites
The extent of water pollution was classified in accordance
to water quality identification system based on ASPT index of
Richard Orton, Anne Bebbington, Jonh Bebbington (1995) and
Stephen Eric Mustow (1997) as shown in table I [9].
TABLE I ASPT INDEX AND WATER POLLUTION LEVEL

ASPT index Category
0 0.99 Extremely severe pollution
1.00 2.99 Severe pollution (Oligosaprobic)
3.00 4.99 Moderate pollution ( -mesosaprobic)
5.00 5.99 Moderate pollution (-mesosaprobic)
6.00 7.99 Relatively clean water (Polysaprobic)
8 - 10 Clean water

III. RESULTS
A. Taxon composition in rivers in Danang city.
Evaluation of macroinvertebrate composition provides
significant information in terms of ecological health and
structure of aquatic organism community. Regardless of
macroinvertebrates out of the BMWP
VIET
score system, from
2007 to 2011 in 4 rivers in Danang city, the research has
resulted in 16 families belonging to 6 orders and 1 sub-class in
Phu Loc river (2007 2008) [6]; 24 families belonging to 18
orders in Cu De river (2009 2010)[5]; 20 families belonging
to 14 orders and 1 sub-class in Cau Do Tuy Loan system
river (2009 2010) [4]; 16 families belonging to 11 orders and
1 sub-class in Han rivers (2010 2011)[3]. The taxon
compositions are specified in the following tables (from I to
V).
TABLE II. TAXON COMPOSITION IN PHU LOC RIVER (2007 2008)

No. Order Family
1 Odonata Lestidae
2 Odonata Coenagrionidae
3 Coleoptera Chrysomelidae
4 Coleoptera Dytiscidae
5 Coleoptera Hydrophilidae
6 Coleoptera Hygrobiidae
7 Heteroptera Nepidae
8 Heteroptera Pleidae
9 Heteroptera Belostomatidae
10 Architaenioglossa Viviparidae
11 Basommatophora Lymnaeidae
12 Basommatophora Planorbidae
13 Decapoda Palaemonidae
14 Decapoda Parathelphusidae
15 Neotaenioglossa Thiaridae
16 Diptera Chironomidae
17 Oligochaeta*
*: sub-class

TABLE III. TAXON COMPOSTION IN CU DE RIVER (2009 2010)

No. Order Family
1 Odonata Amphipterygidae
2 Odonata Petaluridae
3 Decapoda Potamidae
4 Decapoda Palaemonidae
5 Decapoda Atyidae
6 Mesogastropoda Fairbankiidae
7 Mesogastropoda Fluminicolidae
8 Coleoptera Ptilodactiliidae
9 Coleoptera Psephenidae
10 Hypsogastropoda Assimineidae
11 Hypsogastropoda Bithynidae
12 Mytiloida Mytilidae
13 Aciculata Nereidae
14 Veneroida Corbiculidae
15 Unionoida Unionidae
16 Amphipoda Gammaridae
17 Basommatophora Ancylidae
18 Neotaenioglossa Thiaridae
19 Isopoda Corallanidae
20 Sorbeoconcha Pachychilidae
21 Rhynchobdellida Piscicolidae
22 Plecoptera Perlodidae
23 Diptera Ptychopteridae
24 Hemiptera Aphelocheiridae

90

TABLE IV. TAXON COMPOSITION IN CAU DO TUY LOAN RIVER
SYSTEM (2010 2011)
No. Order Family
1 Basommatophora Ancylidae
2 Basommatophora Lymnaeidae
3 Basommatophora Planorbidae
4 Neotaenioglossa Bithyniidae
5 Neotaenioglossa Thiaridae
6 Decapoda Palaemonidae
7 Decapoda Potamidae
8 Mesogastropoda Fluminicolidae
9 Mesogastropoda Trichotropidae
10 Coleoptera Psephenidae
11 Trichoptera Philopotamidae
12 Mytiloida Mytilidae
13 Heteroptera Hydrometridae
14 Veneroida Corbiculidae
15 Odonata Platycnemiidae
16 Diptera Ptychoteridae
17 Hypsogastropoda Assimineidae
18 Aciculata Nereidae
19 Sorbeoconcha Pachychilidae
20 Polychaeta**
*: sub-class

TABLE V. TAXON COMPOSITION IN HAN RIVER (2010 2011)

No. Order Family
1 Decapoda Potamidae
2 Decapoda Palaemonidae
3 Decapoda Parathelphusidae
4 Basommatophora Ancylidae
5 Basommatophora Lymnaeidae
6 Hemiptera Hydrometridae
7 Hemiptera Pleidae
8 Mesogastropoda Pilidae
9 Unionoida Unionidae
10 Odonata Lestidae
11 Architaenioglossa Viviparidae
12 Heteroptera Belostomatidae
13 Coleoptera Chrysomelidae
14 Basommatophora Planorbidae
15 Neotaenioglossa Thiaridae
16 Polychaeta**
*: sub-class
B. Results of testing water quality of rivers in Danang city.
Macroinvertebrate community structure contains data
significant for reflecting the state of water body, the presence
of pollution-intolerant families within high BMWP score
suggest great freshwater biodiversity and good water quality,
whilst low BMWP score families as pollution-tolerant families
provide signs of environmental contamination.
TABLE VI. WATER QUALITY OF PHU LOC RIVER BASED ON ASPT
INDEX
Research
site
ASPT Rank ASPT Rank
June 2007 September 2007
PL1 3.73 -mesosaprobe 4.09 -mesosaprobe
PL2 3.83 -mesosaprobe 3.98 -mesosaprobe
PL3 4.33 -mesosaprobe 4.08 -mesosaprobe
PL4 3.72 -mesosaprobe 4.61 -mesosaprobe
December 2008 March 2008
PL1 3.67 -mesosaprobe 3.60 -mesosaprobe
PL2 3.67 -mesosaprobe 3.77 -mesosaprobe
PL3 3.76 -mesosaprobe 3.78 -mesosaprobe
PL4 4.40 -mesosaprobe 3.89 -mesosaprobe

Table VI described water quality in Phu Loc river in 4
sampling times (2007 2008) within 4 seasons. The samples
were taken to investigate APST index for testing water quality.
The analytical results showed that APTS index ranged from
3.60 to 4.40 (Table VI). This indicated that Phu Loc river had
signs of contamination, specifically, the water quality of Phu
Loc river was identified as moderate pollution levels (-
mesosaprobe). Also, the research at Phu Loc indicated that no
significant differences of seasonal variation were found, in
other words, changes of the water quality investigated in Phu
Loc river throughout 4 seasons (2007 2008) was
inconsiderable [6].
TABLE VII.
WATER
June 2009 January 2010
ASPT Rank ASPT Rank
CD1 3.25 -mesosaprobe 5.25 -mesosaprobe
CD2 5.50 -mesosaprobe 5.25 -mesosaprobe
CD3 5.00 -mesosaprobe 5.50 -mesosaprobe
CD4 6.67 Polysaprobe 7.00 Polysaprobe
CD5 7.00 Polysaprobe 6.40 Polysaprobe
CD6 7.00 Polysaprobe 5.67 -mesosaprobe

It can be seen that the water biological indices for each site
can be seen. ASPT scores ranged from 3.25 to 7.00 recorded at
Cu De river in 2 occasions (June 2009 and January 2010). The
water quality of Cu De river was ranged within levels from
polysaprobe to -mesosaprobe, especially, some research sites
had high ASPT score (up to 7.00) indicated relatively good
quality (Table VII) [5].
TABLE VIII. WATER QUALITY OF CAU DO TUY LOAN RIVER
SYSTEM BASED ON ASPT INDEX
Research
site
August 2009 February 2010
ASPT Rank ASPT Rank
TL1 3.00 -mesosaprobe 4.33 -mesosaprobe
TL2 3.00 -mesosaprobe 3.50 -mesosaprobe
TL3 4.00 -mesosaprobe 4.75 -mesosaprobe
TL4 3.00 -mesosaprobe 3.83 -mesosaprobe
TL 5 5.00 -mesosaprobe 3.83 -mesosaprobe
TL6 3.88 -mesosaprobe 3.00 -mesosaprobe

In Cau Do Tuy Loan river system (2009 - 2010), ASPT
scores ranged from 3.00 to 5.00 at research sites. According to
classification system for freshwater quality [9], the quality of
Cau Do Tuy Loan river system was classified from -
mesosaprobic to -mesosaprobe, in other words, the river water
quality had signs of deterioration. Also, it implied that aquatic
ecosystem at Cau Do Tuy Loan river system was suffered
negative impacts from pollution [4]. Specifically, ASPT scores
for each site can be seen at the Table VIII.
TABLE IX. WATER QUALITY OF HAN RIVER BASED ON ASPT
INDEX
Research site ASPT Rank
November 2010
H1 3.25 - mesosaprobe
H2 3.50 - mesosaprobe
H3 4.00 - mesosaprobe
H4 3.25 - mesosaprobe
H5 3.50 - mesosaprobe
H6 4.00 - mesosaprobe
January 2011
H1 4.40 - mesosaprobe
H2 4.40 - mesosaprobe
H3 4.40 - mesosaprobe
H4 4.57 - mesosaprobe
91

H5 4.10 - mesosaprobe
H6 4.67 - mesosaprobe
March 2011
H1 4.00 - mesosaprobe
H2 4.71 - mesosaprobe
H3 3.67 - mesosaprobe
H4 4.00 - mesosaprobe
H5 4.17 - mesosaprobe
H6 5.60 - mesosaprobe

The research conducted in Han river (2010 2011) showed
that ASPT indices ranged from 3.25 to 5.60. Hence, the water
quality of Han river was classified from -mesosaprobe to -
mesosaprobe. Similar to state of water body of Cau Do Tuy
Loan river, Han river had signs of contamination by receiving
polluted effluents [3]. ASPT scores for each site can be seen at
the table 9.
IV. DISCUSSION
In the life circle, aquatic organisms frequently interact with
their habitat. Their numbers and compositions contain data
being useful to determine environmental quality based on
ecological features (e.g. chemical, physical, and biological).
Based on that, broad measurements of the combined impacts of
contaminations could be identified. Specifically, the
ecological-based score of each family of aquatic species
reflects their susceptibility to pollution, which is based on the
principle that aquatic macroinvertebrates have different
tolerance to pollutants, ecological health and state of water
body hence they change in the way, which could be
recognized.
Biomonitoring has not been accepted and utilized for
governed monitoring programs by authorities in Vietnam. The
results of this long-term research have enhanced and provided
scientific data that it is applicable to adopt the BMWP
VIET
and
ASPT for freshwater quality monitoring. The results have
proved the advantages of biomonitoring which is a cost-
effective and easy applicable method with easy identification
and simple practical procedure. Particularly, it can rapidly
indicate of the state of the water body and combined impacts of
pollutants on community structure and ecological health which
physical and chemical analyses implied limitation. Thus, the
BMWP
VIET
and ASPT indices should be considered as an
typical indicator for river quality monitoring.
Nguyen Xuan Quynh et al (2002) proposed a practical
procedure with sampling which should be conducted 4 times in
4 seasons per year in the North of Vietnam. In the South of
Vietnam, sampling may be implemented in 2 times
(representing wet season and dry season) per year or over.
Concerning to aquatic life, macroinvertebrates may be
impacted by benthic structure, flow rate, chemical and physical
parameters of the water body, specially, sudden weather
changes. From the features of climatic condition in central
Vietnam which could be divided to 2 seasons, dry season and
wet season and the results of Phu Loc river, we propose that
sampling could be taken in 2 times.
However, the adaption of BMWP score system to natural
conditions in Danang city is required in order to enhance
applicability of the BMWP
VIET
and ASPT. In fact, natural
condition in Danang is different with natural condition in the
North of Vietnam with features that prolonged rainy season
and high frequency of storm and other inordinate weather
events, leading to reduce sudden disturbance of environmental
parameters. From empirical results of the research, we propose
that the best time for sampling per year is the end of wet season
(from February to March) and mid dry season (from July to
August) that are the moments when rainfall and river flows are
relatively stable. Benthic structure of rivers, a factor
influencing on aquatic ecosystems, such as Cu De river and
Tuy Loan Cau Do river system are complicated, therefore, to
accurately evaluate the state of river water quality, distance
between sampling sites should be from 1000 to 3000 m,
depending on the length of each river. We identified and
sampled in 4 rivers, the research results showed that the
mentioned distance was appropriate.
A suitable adjustment associated with proper practical
procedure demonstrated that it is significantly applicable to
adopt BMWP and ASPT score system water quality testing in
central Vietnam. It can be used as an independent method for
rapid assessment of water quality, alternatively, as a
advantageous tool combinative with physical and chemical
analyses to elevate efficiency of environmental monitoring
activities.
V. CONSLUSION
The research has found 16 families and sub-class belonging
to 10 orders in Phu Loc river (2007 2008); 24 families
belonging to 18 orders and 1 sub-class in Cu De river; 20
families belonging to 16 orders and sub-class in Cau Do Tuy
Loan system river; 16 families belonging to 11 orders in Han
rivers in the BMWP
VIET
system.
The analytical analyses of ASPT index showed the water
quality was moderately polluted (-mesosaprobe) in Phu Loc
river (2007 2008); classified from relatively clean water to
moderately polluted water in Cu De river (2009 2010);
moderately polluted (from -mesosaprobic to -mesosaprobe)
in the Cau Do Tuy Loan river (2009 2010) and in Han river
(2010 - 2011). Biological indices reflected the state of water
body of rivers in Danang city and provided broad measure of
combined impacts of pollutants on ecosystem and organism
life. The initial results also indicates that BMWP
VIET
and ASPT
indices can undoubtedly be applied to monitor rivers in Danang
city in particular and in central Vietnam in general.
REFERENCES
[1]. Truong Thanh Canh and Ngo Thi Tram Anh, "Application of
macroinvertebrates for assessment of water quality in 4 main canal
systems in Hochiminh city", J ournal of Science and Technology
Development, Vietnam10, 25 - 31,2007.
[2]. L. Cota, M. Goulart, P. Moreno, and M. Callisto,"Rapid assessment
of river water quality using a adapted BMWP index: a practical tool
to evaluate ecosystemhealth", Verh. Internat. Verein. Limnol. 28, 1 -
4,2002.
[3]. Phan Thi Hien, Nguyen Dinh Anh, Vo Van Minh, and Nguyen Van
Khanh, "Use of macroinvertebrates for water quality monitoring in
Han river, Danang city", J ournal of Science and Education, Danang
University of Education 3 (02), 25 - 30,2012.
[4]. Nguyen Van Khanh, Pham Thi Hong Ha, and Dam Minh
Anh,"Assessment of water quality of Cau Do - Tuy Loan river
system in Danang city using BMWPVIET", J ournal of Science and
Technology, Danang University Vol 1. No. 5(40),2010.
92

[5]. Nguyen Van Khanh, Vo Van Minh, and Vo Huy Cam, "Assessment
of water quality of Cu De river, Danang city using BMWP
VIET
score
system" in National Scientific Conference for Union Officials. 2012:
Hanoi University of Education.
[6]. Nguyen Van Khanh, Tran Duy Vinh, Duong Cong Vinh, and Ung
Van Thach,"Use of macroinvertebrates for assessment of water
quality in Phu Loc river", J ournal of Science and Technology,
Danang University 2 (37),2010.
[7]. Vo Van Phu, Hoang Dinh Trung, and Le Mai Hoang Thy,"Using
large-size invertebrates to estimate water quality in Bo river, Thua
Thien Hue province", J ournal of Science, Hue University 57, 129 -
139,2010.
[8]. Nguyen Xuan Quynh, Clive Pinder, and Steve Tilling, Identification
of commonly encountered macroinvertebrates in freshwater in
Vietnam. 2001: VietnamNational University, Hanoi.
[9]. Nguyen Xuan Quynh, Clive Pinder, Steve Tilling, and Mai Dinh
Yen, Freshwater environment biomonitoring using
macroinvertebrates. 2002: VietnamNational University, Hanoi.
[10]. Kennedy Francis Roche, Ednilson Paulino Queiroz, and Karina
Ocampo Righi Glaucio Mendes de Souza,"Use of the BMWP and
ASPT indexes for monitoring environmental quality in a neotropical
stream", Acta Limnologics Brasiliensia vol. 22, no. 1, p.105-
108,2010.
[11]. Caroline L. Wenn,"Do freshwater macroinvertebrates reflet water
quality improvement following the removal of point source polution
fromSpen Beck, West Yorkshire? " Earth and Envionment Vol. 3,
369 - 406,2008.
93

WHOLE CELL IMMOBILISATION OF BACILLUS
SUBTILIS ON CELLULOSE CARRIERS AND
WASTEWATER TREATMEMT APPLICATION

TRAN, Thi Xo
Dept. of Biotechnology, Faculty of Chemistry
Engineering, Danang University of Technology.
NGUYEN, Thi Diem Quynh
Dept. of Medical and Pharmaceutical Science,
Duy Tan University,
Da Nang
Danang city, Vietnam


Abstract - In this study, we carried out fixing Bacillus subtilis on
cellulose-rich carriers (filter-paper, coconut fiber). B. subtilis is
capable of producing extra-cellular amylase, cellulase and
protease effectively. After immobilization, bacteria have still
maintained its original properties. Bacteria-fixed materials could
be stored at 4C, for 3 months. Moreover, application of these
carriers in treating wastewater which contains a large amount of
protein, starch, and cellulose showed that COD and BOD5
significantly reduced (12.8% and 10.4%, respectively) in
comparison with treating wastewater by active sludge.
Key words: Bacillus subtilis, immobilization, cellulose carriers,
waste water treatment.
I. INTRODUCTION
Bacillus subtilis is a rod-shaped, Gram-positive bacterium
which can be found in soil, water, air, dried grass, skinThis
anaerobic or aerobic, non-pathogenic strain could prosperously
grow at 5 50
0
C. At higher temperature, B. subtilis is able to
produce an endo-spore that allows it to endure extreme
conditions of heat and desiccation in the environment. Bacillus
subtilis is one of the most widely used bacteria for the
production of enzymes (amylase, glucoamylase, glucanase,
cellulase, dextranase, protease), and special chemical
compounds, such as riboflavin. Therefore, B. subtillis has been
applied in many industrial fields [1].
In 1978, Toshinori Kokubu immobilized Bacillus subtilis
on polyacrylamide gel to produce -Amylase [2]. In 2005,
Kunamneni Adinarayana [3] attached Bacillus subtilis on
others carriers such as calcium alginat, k -Carrageenan,
ployacrylamide, agar-agar and gelatin to evaluate protease
producing ability of B. subtilis.
Cellulosic materials also were choosen to immobilize
bacteria. Aijun A. Wang, Ashok Mulchandani and Wilfred
Chen (2001) [4] who studied on Eshcherichia coli
immobilization on cellulosic carriers, concluded that E. coli
was capable of binding on surface of cellulosic materials and
its binding level depended on surface structure. In 1980,
Mogilevich et al. carried out Bacillus subtilis fixation on
cellulosic carrier. They indicated that bacterial activity was not
changed as fixing on this material [5].
Based on the previous research, we performed
Bacillus subtilis immobilization on cellulosic surface (filter-
paper, coconut fiber) for preservation as well as wastewater
treatment application
II. MATERIALS AND METHODS
A. Materials:
Bacillus subtilis was provided by Biotechnology Lab,
Danang University of Technology. Filter-paper with high
thickness was taken from Advantec company, China. Coconut
fiber bought from supermakets was pretreated before
performing the immobilization.
B. subtilis medium ingredients: Beef extract 3g/L, Yeast
extract 10g/L, Pepton 10g/L, NaCl 5g/L.
B. Methods
1) Method of evaluating extracellular enzyme production (
amylase, cellulose and protease).
Bacteria were cultured on medium containing 1%
substrates (starch, Carboxymethyl cellulose - CMC, and
gelatin). Then, the culture was grown at 30
o
C for 24h. We used
Lugol reagent to determine amylase and cellulases activity and
amido black dye to determine proteases activity.
2) Method of whole cell immobilization
We chose absorption method to fix whole cell on carriers.
Cellulose carriers autoclaved were put into the growth culture
of B.subtilis, then incubated at 30
o
C for 12h with shaking (150
rpm) for attachment between bacteria and cellulosic surface.
After that, these materials were taken out of culture and dried
94

at appropriate temperature. Finally, bacteria-immolized carriers
were stored at 4C.
III. RESULTS AND DISCUSSION
A. Carrier preparation
When soaking coconut fibres into water, water color became
brown-red. Therefore, we pretreated coconut fiber by soaking it
into NaOH 1% for 4h. Then, they were washed many times by
water until pH of water was neutral (Fig. 1).

Figure 1. Water color before (a) and after (b) treatment.
Treated fiber dried to moisture lower than 15% and filter-
paper were cut into 4 cm lines and 1x1.5 cm piece,
respectively. Before fixing bacterial cells on surface, we
evaluted cellulose and lignin concentration of cellulosic
surfaces. The result (table 1) showed that treated fiber consist
of not only cellulose, but also lignin (37.2%).
Table 1. Cellulose and lignin concentration of coconut fibres and
filter-paper.
Sample Fiber Filter-paper
Cellulose (%) 40,0 63,3
Lignin (%) 37,2 -

B. Evaluation of exoenzyme production of Bacillus subtilis
After activating the bacteria in media containing 20 g/l agar,
we checked extracellular enzyme producing ability of the
bacteria by measuring hydrolyzed circle diameter. The results
indicated that circle diameters produced by amylase, cellulase,
and proteases hydrolyzation on appropriate substrates (starch,
cellulose, protein) were 0.9, 2.8 and 1.0 cm, respectively (Fig
2). This result indicated that B. Subtilis is able to synthesis all 3
types of these enzymes.






Figure 2. Evaluation of exoenzyme production of Bacillus subtilis:
amylase (a), cellulase (b), protease (c)
C. Whole cell immobilization of Bacillus subtilis on filter-paper
and coconut fiber
Firstly, the bacteria were grown in broth at 30
o
C, with shaking
(150 rpm) for 24h until the culture reached to log phase. Then,
cellulosic carriers were incubated into the culture for 12h. This
incubation combined with discontinous shaking which
performed shaking for 15 min after 1h stationary incubation.
After that, the carriers were taken out of the culture and dried
at 38.5
o
C by using vacuum drying cabinet. In order to achieve
initial dryness, bacterial cell-immobilized filter paper and
coconut fiber were heated for 32h and 35h, respectively.
After immobilization, bacterial activity was examined at 2
periods: after drying and after 3- month storage. In order to
activate the bacteria, we soaked dried carrier in liquid medium
for 15 min and then asepticly put on agar medium. The plate
was incubated at 30
o
C. After 24h, typical colonies of Bacillus
subtilis were formed (Fig. 3)

Figure 3. Evaluation of growth of B. Subtilis after immobilization on carrier
(a) filter paper (b) Coconut fibre
After 3-month storage, we evaluated extracellular enzyme
production of B. subtilis. The result showed that the bacteria
still maintained original properties (Fig.4).





(a) (b) (c)
Figure 4: Evaluation of exoenzyme production of Bacillus subtilis: amylase
(a), cellulase (b), protease (c) after 3 months storage
Based on these results obtained, we realized that:
- Bacillus subtilis is able to immobilize on cellulose rich
carriers by using absorption method.
- The present of lignin in coconut fiber did not badly affect
the binding of bacterial cell on carrier.
B A
A C B
95

- Immobilization manipulation was performed in aseptic
conditions. Therefore, it is possible to use this method for B.
subtilis storage purpose.
- After 3 months storage at 4
o
C, B. Subtilis still conserves
characteristic properties, such as exoenzyme producing ability
(amylase, cellulase and protease).
D. Application of Bacillus subtilis immobilization on carriers
in wastewater treatment.
The bacteria immobilized on cellulosis carrier is used as a
potential source for wastewater treatment. Wastewater is a
nutrient-rich medium that is suitable for bacterial growth.
After treatment, bacterial biomass in the wastewater is fixed on
carrier. By that way, it is completely possible to use bacteria-
immobilized carriers for many times.
We attached the bacteria on long coconut fibres with 8 cm
length, 5 cm width and 1.5 cm thickness. Wastewater samples
were taken from wastewater pre-biotreating area of Dien Nam-
Dien Ngoc Industrial Park, Quang Nam province. This
Industrial Park hosts seafood factories, breweries, paper-mills,
so wastewater from these factories facilitate the bacterias
growth to produce enzymes for wastewater treatment.
Then we carried two procedures for the treatment of that
wastewater. The first procedure is aerobic treatment with
actived sludge for 12h. The effuluent sample was considerd as
a control. In the second procedure, the wastewater was passed
through bacteria-immobilized coconut fibre, then the output
wastewater was aerobically treated with actived sludge for 12h
as in procedure 1. BOD
5
and COD of pretreated and treated
wastewater were evaluated. The results (Table 2) showed that
after passing wastewater through bacteria immobilized in the
carrier, COD and BOD
5
significantly reduced. BOD
5
and COD
was 10.4% and 12.8% lower than controls, respectively.
Table 2: COD and BOD5 value of pretreated and treated
wastewater
Sample COD (mg/l) BOD
5
(mg/l)
Pretreated 126.70 74.53
Treated (Procedure
1)
83.30 47.60
Treated (Procedure
2)
67.10 38.90
This indicated that it is possible to fix non-pathogenic
bacteria on natural carriers for environmental treatment.
4. Conclusion
Bacillus subtilis, non-pathogenic, enzymes producing
baterium, could be fixed on cellulosis carriers, such as filter-
paper, coconut fibre. After immobilization, bacteria have still
maintained its original properties. Bacteria-immobilized
carriers could be stored 4C for 3 months. Application of these
carries fixed with bacteria in protein, starch and cellolose rich
wastewater gave lower COD v BOD
5
value than treatment
wastewater with active sludge.

REFERENCES
[1]. Nguyn Ln Dng, Nguyn nh Quyn, Phan Vn Ty
(2002), Vi sinh vt hc, Nh Xut Bn Gio Dc.
[2].Toshinori Kokubu, Isao Karube and Shuichi Suzuki,
(1978), -Amylase production by immobilized whole cells of
Bacillus subtilis, European J. Appl. Microbiol, Volume5, pp.
233-240.
[3].Kunamneni Adinarayana, Bezaada Jyothi,
and Poluri Ellaiah, (2005), Production of Alkaline Protease
With Immobilized Cells of Bacillus subtilis PE-11 in Various
Matrices by Entrapment Technique, AAPS PharmSciTech,
Volume 6, pp. 391-397.
[4]. Aijun A. Wang, Ashok Mulchandani and Wilfred Chen,
(2001), Whole-Cell Immobilization Using Cell Surface-
Exposed Cellulose-Binding Domain, Biotechnol. Prog,
Volume 17, pp.407411.
[5]. Mogilevich NF, Garbara SV, (1980), Effect of electro-
immobilization on Bacillus subtilis, Mikrobiologiia, Volume
49, No 06, pp. 876-879





96

LACTIC ACID FERMENTATION FROM
J ACKFRUIT SEED FLOUR

Trng Thi Minh Hanh
Faculty of Chemistry, Danang University of
Technology, email: tminhhanh2001@yahoo.com


Ho Thi Hao
Agriculture and Forestry Faculty, Tay Nguyen
University, email: hoahaoanh123@gmail.com


Abstract In this paper, we present our research results on lactic
acid fermentation process from jackfruit seed flour. The
hydrolysis process of jackfruit seed starch using two enzymes, -
amylase and glucoamylase with the percentage of 0.10% and 0.18%,
respectively, compared to the substrates, yields the hydrolysis
efficiency of 91.436%. Four factors that affect the lactic acid
fermentation are investigated in this study, including reducing sugar
concentration, Lactobacillus casei bacteria percentage,
temperature, and fermentation duration. Based on experimental
results, we propose parameters for lactic acid fermentation: (i)
initial reducing sugar concentration of 5%, (ii) Lactobacillus casei
bacteria percentage of 5% with cell density of 51.10
3
CFU/ml, (iii)
temperature of 37
O
C, and (iv) fermentation duration of 72 hours.
Under the above conditions, the obtained lactic acid
concentration is thus 18.25g/l we propose parameters for lactic
acid fermentation: (i) initial reducing sugar concentration of 5%,
(ii) Lactobacillus casei bacteria percentage of 5% with cell density
of 51.10
3
CFU/ml, (iii) temperature of 37
O
C, and (iv) fermentation
duration of 72 hours. Under the above conditions, the obtained
lactic acid concentration is thus 18.25g/l .
Keywords- Jackfruit seed flour, agricultural residue,
liquefaction, saccharication, lactic acid fermentation medium
I. INTRODUCTION
Lactic acid is a product of a lactic fermentation process,
which is an anaerobic biological metabolism that transforms
sugar compound mainly into lactic acid and other by-products.
Lactic acid is widely used in the food industry and others such
as cosmetic, chemical, and pharmaceutical industries. In novel
material technologies, lactic acid is a raw material to produce
poly lactic acid (PLA), which is an important biomaterial for
many other industries, through polymerization reactions. Raw
materials for producing lactic acid conventionally come from
carbohydrate compounds such as saccharose, glucose and
lactose as well as starch such as corn, potato, cassava,
sugarcane molasses, starch hydrolyzate and timber hydrolyzate.
Recently, many scientists have increasingly focused on using
agricultural and industrial residues as a new approach for lactic
acid production [3].
Jackfruit, a delicious and nutritious tropical fruit, is an
important material for the production of jackfruit jam, dried
jackfruit, jackfruit juice, jackfruit preserve, candied lanka [5].
Interestingly, Daklak is a province that has relatively high
jackfruit production in Vietnam. In 2010, the provinces total
production of jackfruit is 280667.8 tones, in which Vinamit
Companys need is 132000 tones. According to Daklak
Agricultural Department, the amount of jackfruit seeds released
into the environment is 17160 tones. These seeds will become
pollution unless they are treated effectively. Since jackfruit
seeds contain high concentration of starch and protein, it is
possible to use the jackfruit seeds as a good substrate source for
lactic acid fermentation. In other words, lactic acid
fermentation from jackfruit seed starch not only offers
significant economic value but also contributes to
environmental improvement.
II. MATERIALS AND METHODS
A. Materials
We used the following materials:
Jackfruit seeds were collected from a small jackfruit
factory of Vinamit Company in Buon Ho, Krongbuk,
Daklak Province.
Starch hydrolysis enzymes used in this study were -
amylase and glucoamylase (-amylase), provided by
Novozyme Company (Denmark) and Genencor Bio-
products Co. Ltd (China), respectively.
Lactobacillus casei were provided by Nanogen
Biopharma, Hochiminh City, Vietnam.
B. Methods
1) Analysis methods
After being transformed into reducing sugar by using an
acid solution 2%, starch was quantitatively determined using
the Graxianop method [2]. The reducing sugar concentration in
the obtained hydrolyzate was determined by the colorimetric
method using dinitrosalicylic acid (DNS) [2]. In addition, lactic
acid was determined by the colorimetric method using
paraoxidiphenyl [2]. The activities of -amylase and
glucoamylase were determined by using the Rukhliadeva
method and the V.Y.Rodzevich - O.P.Korenbiakina method,
respectively [2].
2) Technological methods
97

a) Jackfruit processing and jackfruit flour production

b) Starch hydrolysis method

c) Lactic fermentation method:
- Medium preparation: The production medium contained
8.0g/l yeast extract, 0.5g/l dipotassium hydrogen phosphate,
0.5g/l potassium dihydrogen phosphate, 1.0 g/l sodium acetate
3H
2
O, 0.6g/l magnesium sulfate 7H
2
O, 0.03g/l manganese
sulfate 4H
2
O, and a 1000-ml hydrolysate of jackfruit starch
[34]. The pH of the medium was adjusted by H
2
SO
4
20% and
NaOH 5N [34] to pH=6. The medium was sterilized at 121
o
C
for 20 minutes
- Fermentation procedure: Fig. 1 summarizes the
fermentation procedure and the processing of the fermented
solution in order to quantitatively determine lactic acid [1]

Figure 1. Fermentation procedure and fermented solution processing
III. RESULTS AND DISCUSSION
A. Investigation of material characteristics before lactic acid
fermentation
Table I showed that the content of jackfruit starch in this
work was lower than that reported in previous study (77.76
0.965%) [6]. It is probably due to partially ripe jackfruit.
Nevertheless, with humidity under a certain threshold (<
13%), this jackfruit flour is still suitable for our research
purpose and even for production. In addition, we observe that
the activities of the enzymes at the experimentation time were
high. However, since the ratio of glucoamylase enzymeto the
substrate had not been determined, we carried out further
study to find the best enzyme concentration for the
saccharication process.
TABLE I. MATERIAL CHARACTERISTICS BEFORE LACTIC ACID
FERMENTATION

Feature Jackfruit
flour
Enzyme
Humidity (%) 11.06
Starch (%) 65.62
Activity of -amylase
enzymeproduct
3623.4 unit
Activity of glucoamylase
enzyme product
4752.0 unit

Note:
Unit of Activity of -amylase enzyme product is the
amount of enzyme (ml) that transforms 1g of dissolved
starch into dextrin with different molecular weights at
30
o
C in 1 hour.
Unit of Activity of glucoamylase enzyme product is
the amount of enzyme (ml) that applies on dissolved
starch solution at pH =4.7 at 30
o
C in 1 hour to release
1mg of glucose.
B. Effects of the glucoamylase enzyme percentage on the
hydrolysis of jackfruit seed starch
Experiments were carried out according to the flow chart of
starch hydrolysis in Section 2.b. We used 0.1 % of -amylase
and 33% of jackfruit seed flour and -amylase varied from 0.12
to 0.20%. The results of Table II presented reducing sugar
concentrations obtained from jackfruit seed hydrolysate under
different glucoamylase amounts.
TABLE II: REDUCING SUGAR CONCENTRATION OBTAINED AFTER
HYDROLYSIS

No.
Supplemented
enzyme
amylase, %
Obtained
reducing sugar,
g/100ml
Hydrolysis
efficiency, %
1 0.12 10.769 49.233
2 0.14 11.667 53.339
3 0.16 14.000 64.005
4 0.18 20.000 91.436
5 0.20 15.556 71.119

The data in Table II showed that elevated amylase
gave the increase of reducing sugar amount in the hydrolyzed
solution. However, when the percentage of amylase was
0.20 %, the concentration of reducing sugar decreased. It is
maybe due to competing effects between the enzyme and the
substrate. Moreover, using 0.18% of amylase offered
highest hydrolysis efficiency (91.436 %). It is in the same
magnitude of results reported by Nguyen Minh Hien and
98

Nguyen Thuy Huong [7]. Therefore, we chose 0.18%
amylasefor next experiments.
C. Investigation of factors affecting lactic acid
fermentation
1) Effect of fermentation temperature
This experiment was carried out with the following
parameters: (i) the initial reducing sugar percentage 4%, (ii)
fermentation duration 72h, (iii) stirring speed 150 rpm, (iv)
cultivated Lactobacillus casei strain percentage that was added
5% (v/v) compared to the fermentation medium, with cell
density 51.10
3
CFU/ml and (v) fermentation temperature 35
O
C
- 40
O
C. The result was shown in Fig. 2.

Figure 2. Effects of temperature on lactic acid fermentation
Accoding to Fig. 2, the Lactobacillus casei bacteria grew
well at the temperature of 37
o
C in the medium containing
hydrolyzed jackfruit starch solution. At this condition, the
concentration of lactic acid achieved the maximal value of
18.20 g/l. Hence, we used 37
O
C as a suitable temperature in
next experiments.
2) Effects of bacteria strain percentage
Lactic acid fermentation process was carried out with the
initial reducing sugar 4% at 37
O
C for 72 hours. The solution
was shaken with 150 rpm. The concentration of bacteria strain
inoculated in medium was changed from 2.0 to 8.0 %. The
results of Fig. 3 showed that the initial bacteria strain
percentage of 5 % was the most suitable for lactic acid
fermentation.

Figure 3. Effects of the bacteria strain percentage on lactic acid fermentation
3) Effects of initial reducing sugar percentage
The lactic acid fermentation was carried out by culturing
bacteria in medium containing 5 % of bacteria strain and 2.0
8.0 % reducing sugar. The solution was shaken with rate of
150 rpm at 37
O
C for 72 hours. The result is shown in Fig. 4.
.

Figure 4. Effects of initial reducing sugar percentage on lactic acid
fermentation
4) Effects of fermentation duration:
Similarly, the lactic acid fermentation was carried out
with bacteria strain of the 5% concentration and initial
reducing sugar of 4%. We used different durations in this
experiment. In particular, the solution was shaked with rate of
150 rpm/min at 37
O
C from 24 to 88 hours. The result is shown
in Fig. 5..
By varying three effecting factors, namely bacteria strain
percentage, initial reducing sugar concentration and duration,
we determined that lactic acid fermentation is best when
Lactobacillus casei was grown in the medium containing 5%
initial reducing sugar concentration at 37
O
C for 72 hours.

Figure 5. Effects of fermentation duration on lactic acid fermentation
IV. CONCLUSION
In summary, we have studied the hydrolysis of jackfruit
seed starch using the -amylase enzyme with percentage of
0.1% and the -amylase enzyme with percentage of 0.18%
99

compared to the substrate. The results showed that we achieved
the hydrolysis efficiency of 91.44%, reducing sugar
concentration of 200 g/l, and the quality of hydrolyzed solution
was good. In addition, we have determined optimal conditions
for lactic acid fermentation as following: temperature of 37
O
C,
Lactobacillus casei inoculum of 5%, initial reducing sugar
concentration of 5%, and fermentation duration of 72h. Under
the above conditions, the lactic acid concentration obtained was
18.25 g/l. These results have shown that utilizing jackfruit
seeds in lactic acid fermentation technologies has promising
economic value and reduces significantly harmful wastes in the
environment.
REFERENCES
[1] Nguyn c Lng, Cng ngh vi sinh vt - tp 2, Trng i hc
K thut, 1996
[2] L Thanh Mai, Cc phng php phn tch ngnh cng ngh ln men,
Nh xut bn Khoa hc v K thut H Ni, 2005.
[3] P.Singh.neeNigam, Production of Organic acids from Agro_Industrial
Residues, Faculty of life and Health Sciences, School of Biomedical
Sciences, University of Ulster, Coleraine BT52 1SA, Northern Ireland,
UK. 2009
[4] Sakhamuri Sivakesava, J oseph Irudayaraj Demivci Ali, Simultaneous
determination of multiple components in lactic acid fermentation using
FT_MiR, NiR and FT_Raman spectroscopic techniques, Process
Biochemistry 2001
[5] Technoguide Series, J ackfruit DA-RFU 8, eastern Visayas Integrated
Agricultural Research Center (EVIARC)
[6] Vanna Tulyathana, Kanitha Tananuwonga, Prapa Songjinda and
Nongnuj J aiboon, Some Physicochemical Properties of J ackfruit
(Artocarpus heterophyllus Lam) Seed Flour and Starch, 2001,
Department of Food Technology, Faculty of Science, Chulalongkorn
University, Thailand 10330.
[7] http://www.cesti.gov.vn/kh-cn-trong-n-c/nghi-n-c-u-qua-trinh-th-y-
phan-tinh-b-t-h-t-mit-b-ng-enzyme-termamyl-120l-va-amg-300l-d-l-n-
men-r-u-ch-ng-c-t.html
100

GOLD NANOPARTICLES BASED LOCALIZED SURFACE PLASMON RESONANCE IN
COMBINATION WITH MICROFLUIDIC SYSTEM FOR BIOMOLECULE
DETERMINATION

Nguyen Ba Trung, Le Tu Hai,
Danang University of Education, Danang University

Yuzuru Takamura
J apan Advanced Institute of Science and Tecnology



AbstractIn recent years, metallic nanoparticles have been
studied extensively for the nanoelectronics, nanophotonics, nano-
catalyst, and biosensor applications. The unique optical property
of metallic nanoparticles is suitable to be employed as a marker
for the label-free optical detection based on localized surface
plasmon resonance. In this work, we investigate a simple method
for monitoring the interaction of antigen-antibody on the PDMS
sensing surface based on the localized surface plasmon resonance
of immobilized gold nano particles. The interaction between
antigen-antibody was examined by recording the absorbance
intensity, as well as the peak wavelength shift of the LSPR band.
The results showed that our device can be employed to
qualitatively and quantitatively analyze the antigen presenting in
sample. The developed technique is also hopefully expanded the
potential applications of PDMS microfluidic chip on studying
immunoassays and other biochemical analyses
Keywords-component; surface plasmon resonance; biosensing;
gold nano particles; immunoassay; microfluidic
I. INTRODUCTION
In recent years, metallic nanoparticles have been studied
extensively for the nanoelectronics, nanophotonics, and
biosensor applications.
1)
The unique optical property of
metallic nanoparticles is suitable to be employed as a marker
for the label-free optical detection based on localized surface
plasmon resonance (LSPR) which is a collective oscillation
of free electron on metal nanoparticles induced by incident
light. The absorbance peak wavelength and the spectrum
shape of AuNPs are sensitive to the size, shape, and
aggregation of AuNPs, as well as the surrounding medium.
The principle of LSPR sensor is to measure changes in a
characteristic of the LSPR spectrum from alteration in the
effective refractive index of the surrounding medium, leading
to a shift in the absorbance peak wavelength. Nowadays, LSPR
sensors have attracted increasing attention to use for detecting
chemical and biological substances related to medical
diagnostics, environmental monitoring, agriculture pesticide,
antibiotic monitoring, food additive testing, chemical agent
testing, and drug screening.
Micro total analysis system (-TAS) has been
established itself at the forefront of analytical chemistry,
especially biosensing microfluidic devices. Recently, soft
polymer materials are known as the chosen materials for
fabricating multifunctional microfluidic devices instead of
glass, quartz or silicon. Among them, Polydimethyl siloxane
(PDMS) is one of the most widely-used polymer materials due
to its transparency, good biocompatibility, facile bonding
ability, high transparency for UV and fluorescence detection,
and cost-effectiveness for the production.
2)

The combination of microfluidic chip made of PDMS
material and unique property of localized surface plasmon
resonance of noble metal NPs could be efficiently applied for
the label free biosening microfluidic device. In this work,
AuNPs were directly immobilized on the sensing surface of
PDMS microfluidic system for biosensing applications.
Immunno assay for the detection of C-reactive protein (CRP)
antibody-antigen interaction was carried out using the
developed device. The sensitivity is comparable to the reported
method
3)
. This method contributes to the development of label
free biological molecule analyses.
II. EXPERIMENTAL
A. Chemicals
The Sylgard 184 including PDMS monomer and curing
agent were purchased from Dow Corning (Midland, MI, USA).
Aminopropyltriethoxysilane (- APTES) and D-(+) Glucose
were purchased from Sigma-Aldrich. Citrate stabilized AuNPs
100 nm in average diameter distribution was purchased from
British Bio Cell and used as received. 10-carboxyl-1-
decanethiol and 1-ethyl-3-(3-dimethyl aminopropyl)
carbodiimide hydrochloride (EDC) were bought from Dojindo
Laboratories, Japan. N-hydroxysuccinimide (NHS) was
purchased from Wako, Japan. Goat Anti Human C
Reactive Protein (CRP) and C - reactive protein were bought
from Bethyl Laboratories, Inc and Sigma-Aldrich, respectively.
B. Absorbance measurement and surface morphology
characterization
1) Absorbance measurement
Absorbance measurements of AuNPs patterned on PDMS
surface in the visible region were performed using Shimazu
UV-3100. LSPR measurements on the chip were performed
using a set of instruments as shown in Figure 1:
spectrophotometer (USB-4000-UV-Vis, wavelength range: 200
1100 nm), tungsten halogen light source (LS-1, wavelength
range: 360 2000 nm), and optical fiber probe (R-400-7
UV/Vis, fiber core diameter: 400 m, wavelength range: 250
800 nm) with collimating lens that were all purchased from
Ocean Optics. The sample was placed proximately to the
optical fiber probe bundle surface so that the incident light
could be able to pass through the sample. The intensity of the
101

transmitted light was detected by another optical fiber probe,
and then analyzed by UVVis spectrophotometer in a
wavelength range of 400 900 nm at room temperature. The
absorbance peak wavelength of AuNP was determined using
Gaussian curve fitting method.
2) Surface morphology characterization:
The gold colloid distribution on the substrate surface was
characterized by a commercial Atomic force microscope
(Digital Instrument D3100). Measurements done in tapping
mode with a force constant of 18N/m and scan rates in the
range 0.250.50 Hz allowed us to observe the gold colloids on
the substrate surface

Figure 1. Experimental setup for absorbance spectrummeasurements
3) Immobilization of AuNPs on the PDMS substrate
AuNPs were immobilized ion the PDMS sensing surface as
our earlier description. For the immobilization procedure of
AuNPs on the sensing surface of microfluidic chip, mask
tightly sealed with PDMS were used to protect the desired
region on PDMS surface during oxygen plasma treatment. Si
H group on the surface would be consumed by the oxidation
with oxygen plasma and SiOx silica-like layer would be
formed on PDMS surface.
4,5)
The silanisation exclusively
occurred on the oxygen plasma treated area, which allows us to
make spatially selective deposition layer of AuNPs onto a
PDMS surface. Plasma treated PDMS surface was then washed
with ethanol before loading -APTES on the PDMS sensing
surface and kept in the desiccators for 15 minutes. The sample
was then washed again with ethanol and deionized water to
remove any remaining residual -APTES molecules. The
samples were annealed at 120
0
C for two hours. The silanised
samples were immersed into an aqueous solution of citrate
stabilized gold colloids at proper time and thoroughly rinsed
with deionized water. Finally, the samples were then annealed
for about 30 min at 120
0
C
4) Microfluidic LSPR chip for monitoring the real time
interaction between Antigen Antibody
The LSPR PDMS chip is composed of two layers: a flow
layer and a cover layer bonded by oxygen plasma. The
microfluidic device was fabricated using PDMS by standard
soft-lithography techniques
8)
. The flow layer contains a sensing
chamber connecting with inlet and outlet microchannels. The
channels were created with the width of 200 m, the height of
250 m, and 20mm in length. The circular shaped chamber is 5
mm in diameter. The inlet and outlet of the microfluidic device
was formed by connecting with the fluorinated ethylene
propylene (FEP) tube (0.15 0.05mm i.d.) and sealing with
PDMS to prevent leak. Flow through the microfluidic LSPR
chip was performed using micro-syringe pump (KD Scientific).
The flow rate of 60 L/min was set in all experiments
5) Antibody- antigen interaction experiment
The immobilization of CRP antibody on the surface of
AuNPs chip is shown in Figure 2. At first, the formation of
self-assembly monolayer (SAM) was archived by introducing
1mM of 10-carboxyl-1-decanethiol in ethanol onto the surface
of AuNPs layer on chip and keeping overnight at room
temperature. SAM functionalization was carried out with 1:1 in
volume mixture of 400mM of 1-ethyl-3-(3-dimethyl
aminopropyl) carbodiimide hydrochloride (EDC) solution and
100mM of N-hydroxysuccinimide (NHS) solution. The
mixture was added to the SAM functionalized surface for 30
minutes. The 100 g/mL of CRP antibody was immobilized
onto activated SAM modification surface for 30 minutes. After
the immobilization of CRP antibody, the chip surface was
rinsed thoroughly with 20mM phosphate buffered saline (PBS,
pH 7.4), and then blocked with 1% BSA to prevent non-
specific absorption. After rinsing with PBS to eliminate the
unbound molecules, the initial absorbance spectrum was
recorded. Next, CRP antibody immobilization procedure was
performed to monitor the interaction between CRP antibody-
antigen. A desired concentration of CRP antigen solution in
PBS was introduced onto the CRP antibody immobilized Au
NPs surface. The interaction was allowed while continuously
incubating for 30 minutes at room temperature. After a rinsing
and drying up the surface, the changes in the absorption
spectrum caused by the interaction were observed.

Figure 2. The schematic of antibody immobilization process on the sensing
surface and the interaction between the antibody and its relevant antigen
III. RESULTS AND DISCUSSION
A) Immobilization of AnNPs on to PDMS surface
For the introduction of active groups on the inert PDMS
surface, oxygen plasma was applied in order to replace the
silane (Si-Me) groups on the PDMS surface with silanol groups
(Si-OH). After treating with oxygen plasma, PDMS surface can
react with 1% -APTES in ethanol to generate the -NH
2
group
on its surface. The positively charged amino groups can help to
immobilize AuNPs onto silanised substrates by the strong
electrostatic interaction with the negatively charged gold core
surrounded by ionic double layer. The formation of AuNPs
onto the surface of PDMS was monitor using UV-Visible
absorption spectroscopy. Figure 3 shows the surface plasmon
band of AuNPs on the PMDS sensing surface with and without
102

oxygen plasma. From the results, it can be inferred that
treatment of PDMS with oxygen plasma is crucial for the
formation of -APTES film which is necessary for the AuNP
immobilization

Figure 3. (a) UV-Vis spectra of 100nm AuNP patterned on PDMS
substrates, performed on Shimadzu UV-310 0 UV-Vis- NIR recording
spectrophotometer. The silanisation was carried out with 10% - APTES
solution in ethanol for 15 minutes. The incubation time for AuNP deposition
at initial concentration was 4 hours. (b) AFM images of AuNPs on 10% -
APTES modified PDMS surface.
B/ Localized surface plasmon resonance of AuNPs with the
changes in refractive index
400 500 600 700 800 900
0.06
0.08
0.10
0.12
0.14
0.16
0.18
A
b
s
o
r
p
t
i
o
n
Wavelength (nm)
Glucose 10%
Glucose 20%
Glucose 30%
Glucose 40%
Glucose 50%

Figure 4. Absorbance spectrumof a monolayer of immobilized gold colloids
on PDMS in the following glucose concentrations: Glucose 10% (n =1.347);
Glucose 20% (n =1.361); Glucose 30% (n =1.375); Glucose 40% (n =
1.389); Glucose 50% (n =1.403)
We examined the ability of the immobilized monolayer of
AuNPs to transduce changes in the surrounding refractive
index of different glucose concentrations into the absorbance
spectrum. As shown in the Figure 4 there are two plasmon
resonance peaks of AuNPs on the sensing surface. The first
peak around 600 nm in wavelength characterized for mono
AuNPs with high absorbance intensity and another lower
absorbance intensity peak at longer wavelength around 890 nm
characterized for cluster AuNPs. The Plasmon of single AuNPs
is sensitive enough with the changing of RI of surrounding
medium, so that for the discussion on sensing application of
this material, we would only mention the first peak.
In the absorbance spectrum of AuNPs shown in Figure 5, it
exhibited a red shift in the peak wavelength, along with an
increase in the absorbance intensity as increasing of the
concentration of D-(+) glucose solution in the range of 10% to
50%. It was found that a linear relationship between the LSPR
peak (max) and the refractive index of glucose solution. A
least-square gave a sensitivity of 233 nm/refractive index unit
(RIU) with linear regression value R
2
= 0.93 and 0.38
absorbance unit/RIU with R
2
=0.97. The result indicates that
the optical property of an immobilized monolayer of AuNPs is
sensitive to the refractive index of the surrounding bulk
medium. This sensitivity is three times higher compared to the
reported paper (76.4nm/refractive index unit) [7].

Figure 5. The dependences of the absorbance intensity and the wavelength
shift on different refractive indexs of the glucose solutions in water.

C/ Localized surface plasmon resonance of antibody-
antigen interaction
400 500 600 700 800 900
0.08
0.12
0.16
0.20
0.24
A
b
s
o
r
b
a
n
c
e
Wavelength (nm)
SAM
Antibody
BSA
Antigen

Figure 6. The superimposed absorbance spectra curves obtained during
the experiment of CRP antibody-antigen interaction at 1ng/mL of CRP
antigen.
We examined whether the refractive index change at the
surface of AuNPs due to bio-molecular binding events could
also be transduced into an experimentally detectable change in
the absorbance spectrum. CRP antibody and CRP antigen
interaction was used to monitor the characteristic of this
sensing device. Figure 6 shows the superimposed absorbance
spectrum curves obtained during CRP antibody immobilized
process and the interaction of CRP antibody with 10 ng/mL
CRP antigen. The significant increase in absorbance intensity
as well as the red shift in maximum wavelength absorbance
was observed after each step. The dependence of wavelength
shifts on the CRP antigen was also checked. Figure 7 shows
the linear regression line fitted the experimental data with
linear regression value R2 =0.9875 for the concentrations
ranging from 10 ng/mL to 0.1 g/mL. This result, one more
103

time, could point out/ prove the sensing characteristic of
AuNPs patterned on PDMS to the biomolecular binding
events. This is useful information for the application of this
device on label free bio-molecules detection.

Figure7. Therelationshipbetweenthepeak shift and CRP antigenconcentration

IV. RESULTS AND DISCUSSION
We successfully demonstrated the development of PDMS
microfluidic device for bio-analytical applications. The use of
LSPR phenomena of AuNPs in combination with microfluidic
system was successfully monitored the interaction of antigen-
antibody. CRP antigen could be found in a very small
concentration down to 10ng/ml. With our primarily results
on the sensing applicability or bio-molecules, we have
demonstrated that the integration of LSPR phenomena of
AuNPs with the remarkable advantages of microfluidics can
be used to develop a label-free optical sensing device for
biomolecule analysis in medical uses and bioscience
application.
V. REFERENCES
[1] L. He, M. D. Musick, S. R. Nicewarner, F. G. Salinas, S. J . Benkovic,
M. J . Natan, and C. D. Keating: J. Am. Chem. Soc. 122 (2000) 9071
[2] J . C. McDonald, D. C. Duffy, J . R. Anderson, D. T. Chiu, H. Wu, O. J .
A. Schueller, and G. M. Whitesides: Electrophoresis. 21 (2000) 27.
[3] D. K. Kim, K. Kerman, S. Yamamura, Y. S. Kwon, Y. Takamura, and E.
Tamiya: Jpn. J. Appl. Phys, 47 (2008)1351.
[4] Q. Zhang, J . J . Xu, Y. Liu, and H. Y. Chen: Lab Chip. 8 (2008) 352.
[5] Toth, I. Bertoti, M. Blazio, G. Banhegyi, A. Bognar and P. Szaplonczay,
J. Appl. Polym. Sci. 52 (1994) 1293.
[6] D. C. Duffy, J . C. McDonald, O. J . A. Schueller, and G. M. Whitesides:
Anal. Chem. 70 (1998) 4974.
[7] N. Nath and A. Chilkoti: Anal. Chem, 74 (2002) 504.