Вы находитесь на странице: 1из 4

Developing a Mobile Phone-based Fall Detection

System on Android Platform


Shih-Hau Fang1 , Yi-Chung Liang1 and Kuan-Ming Chiu2
Department of Electrical Engineering, Yuan-Ze University, Taiwan1
Far Eastern Memorial Hospital, Taiwan2

AbstractSince todays smartphones are programmable and


embed various sensors, these phones have the potential to change
the way how healthcare is delivered. Fall detection is definitely
one of the possibilities. Injuries due to falls are dangerous,
especially for elderly people, diminishing the quality of life or
even resulting in death. This study presents the implementation
of a fall detection prototype for the Android-based platform. The
proposed system has three components: sensing the accelerometer
data from the mobile embedded sensors, learning the relationship
between the fall behavior and the collected data, and alerting preconfigured contacts through message while detecting fall. We
adopt different fall detection algorithms and conduct various
experiments to evaluate performance. The results show that the
proposed system can recognize the fall from human activities,
such as sitting, walking and standing, with 72.22% sensitivity and
73.78% specificity. The experiment also investigates the impact
of different locations where the phone attached. In addition,
this study further analyzes the trade-off between sensitivity and
specificity and discusses the additional power consumption of the
devices.
Index Termsfall detection; mobile phones; accelerometer

I. I NTRODUCTION
Todays smartphones or mobile phones serve as the central
computing and communication device in peoples lives. Such
a trend is inevitable in the real world due to the tremendous
growth in recent years of new smart devices, such as iPhones,
iPads, and Android-based platform, which is expected to
continue. These phones embed various sensors [1], including
accelerometer, digital compass, GPS, and camera, enabling
new applications in various domains such as healthcare,
social networks, and environmental monitoring. This study
focuses on healthcare domain. Traditionally, the information
for personal healthcare largely comes from infrequent doctor
consultations. The sensor-enabled phones have the capabilities
to collect data that can dramatically change the way health
are assessed and how care is delivered. Recently, e-health
leverages the ubiquity of mobile technologies to assist individuals in monitoring symptoms and treatment [2]. For example,
the UbiFit Garden project [3] attempts to capture levels of
physical activity and relates this information to personal health
while feedbacking to the user. These types of systems have the
potential to greatly improve peoples health and quality of life,
while simultaneously reducing overall healthcare costs.
The elderly population is increasing; therefore, various
systems are proposed to improve elderly peoples health and
quality of life. One good example of these systems is fall
detection. This is because falls are a major health risk for

978-1-4577-1719-2/12/$26.00 2012
2015 IEEE

elderly people, diminishing the quality of life or even resulting in death. The fall detection system can be classified
into four categories: image-based systems [4][6], databasebased approach [7], context-aware techniques [8], [9], and
acceleration-based detection [10][13]. Each approach has
its limitations and advantages. For example, the advantage
of image-based and context-aware approaches is an accurate
detection rate; however, it requires the previous camera or
environmental setup and the maintenance of whole architecture. The database-based approach stores various sensed user
behavior into a database for various activities. This system
can be used to recognize different user behaviors. However,
collecting sensed data for every possible activity of daily living
is time consuming and unfeasible. The acceleration-based
detection is the most widely used method, as current high-level
devices build in the acceleration sensors. This study focuses
on acceleration-based detection, utilizing mobile phones as the
platform to detect fall.
This study proposes a low-cost fall detection system, using
the existing devices and wireless technology, without the
need for hardware modification, environmental setup, and
wearing external sensors. The proposed system has three
central components: sensing, learning, and alerting. In the
first component, we take the advantages of the database-based
approach to collect realistic fall data. That is, we collect the
real accelerometer data from the mobile embedded sensors
and record the corresponding user behavior to determine the
required parameters. In the second component, the proposed
system learns the relationship between the fall behavior and
the collected data. In this step, we utilize different fall features,
including vertical and total acceleration, to design different fall
detection algorithms [12]. We also measure the performance in
both sensitivity and specificity while considering their tradeoff [14]. In the third component, the mobile phones alert preconfigured emergency contacts through message. We further
design an interface which allows users to manually disable the
alert to avoid false positive and to reduce transmission costs.
The experimental results show that the proposed system
can recognize the fall from human activities of daily living,
such as sitting, walking and standing, with 72.22% sensitivity
and 73.78% specificity. This study also analyzes the tradeoff between sensitivity and specificity by plotting receiver
operating characteristic [13]. This illustrates how to determine
the parameters to strike a graceful balance between sensitivity
and specificity. The experiment further investigates the impact
of different locations where the phone attached, including

143

Start

Load
Personal
Profile

Load
Threshold

Parameter
Setup

Emergency Notification
Processes

Send Message to the


emergency contact

Pre-configured

NO
Yes

Real Time
Monitoring

Stop Alarm
in Time

NO
NO

Exceeds
Trigging
Threshold

Yes

Is Fall or Not

Yes

Alarm (Record start


time of alarm)

Fall Detection Processes

Fig. 1.

The architecture of the mobile-phone based fall detection system.

chest, waist, and thigh. The result shows that the performance
of chest is the best. This is because the more high location in
body, the more large change of acceleration will be detected
when falling happened. Finally, we conduct extensive experiments to evaluate the extra power consumption resulting from
the fall detection software.
II. M OBILE PHONE - BASED FALL DETECTION SYSTEM
This section presents the proposed mobile phone-based fall
detection system design. Figure 1 shows the system architecture. The proposed system has three major components: Fall
detection process, system pre-configuration, and emergency
notification, as indicated by different colors in Fig. 1.
In the fall detection process, the mobile device continuously
monitor the sensing data from the embedded sensors in real
time. Here we consider two sensors, tri-axis acceleration
sensor and orientation sensor. The former reports three axis
acceleration values denoted as Ax , Ay and Az , while the latter
provide the value of yaw, pitch and roll orientations, denoted
as x , y and z . In this step, we calculate two parameters
to identify the fall behavior. The first parameter is the overall
acceleration value |AT |, computed as
q
2
2
2
|AT | = |Ax | + |Ay | + |Az |
(1)

time after detecting fall. When users receives the message,


they can determine whether healthcare assistance is required
or not. In this step, we further design an interface which
allows users to manually disable the alarm to avoid false
positive and to reduce transmission costs. If the users stops
the alarm in time, then the algorithm backs to the monitoring
stage. If not, the mobile phones alert pre-configured emergency
contacts through message. We assume that the mobile phones
contained the SIM card and can send message through the
telecommunication company.
In the system pre-configuration process, the mobile devices
have to load the thresholds that are determined during the
offline stage. In this system, the threshold of the vertical
acceleration is 1G, representing the free fall during falling.
The threshold of overall acceleration is obtained from training
data. We find the best threshold that can provide the highest
accuracy in the validation test. Note that some works used
only one parameters and some literatures suggested using both.
This study adopts both parameters and performs a performance
comparison.
The system compares performance in terms of the sensitivity and specificity. Sensitivity and specificity are statistical
measures of the performance of a binary classification test,
resulting from the signal detection theory. Generally, there are
four conditions in fall detection. True positive (TP): a fall
occurs, the algorithm correctly detected. False positive (FP): an
alternative state , but the algorithm detects fall. True negative
(TN): an alternative state, and the algorithm correctly detected.
False negative (FN): a fall occurred, but the algorithm misses.
Sensitivity relates to the systems ability to identify fall. It
measures the proportion of actual falls which are correctly
detected as such. This can be written as
Sensitivity =

(2)

Then, we define two thresholds for these two parameters, respectively. The algorithm is based on these thresholds and the
online-calculated parameters. If the difference value between
maximum |AT |max and minimum |AT |min is larger than the
set |AT |th during a short-period, a fall is considered very likely
happen. Then, we observe the change in vertical acceleration
as a verified procedure. If the vertical acceleration is smaller
than |AV |th during that period, a detection of fall is reported
and the process moves to the next component, emergency
notification.
In the emergency notification process, the mobile phone
starts an alarm message on the screen and records the alarm

(3)

Clearly, a high sensitivity results in a low false negative errors.


On the other hand, specificity relates to the ability of the
system to identify negative results. It measures the proportion
of negatives (not fall activities) which are correctly identified.
This can be written as

The second parameter is vertical acceleration value |AV |,


computed as
|AV | = |Ax sin z + Ay sin y Az cos y cos z |

TP
TP + FN

Specificity =

TN
TN + FP

(4)

Similarly, a high specificity resulting a low false positive


errors. A perfect fall detection system aims to achieve 100%
sensitivity and 100% specificity; however, there is usually a
trade-off between them.
III. E XPERIMENTAL R ESULTS
A. Experimental Setup
We implemented the fall detection system using Androidbased devices, including one HTC Desire and two Tattoo
mobile phones, with version 1.6 of the SDK platform. These
phones were attached to different locations of human body,
including the chest (pocket), waist (leather belt), and thigh
(pocket), as shown in Figure 2. We used these phones to collect
realistic three-axis acceleration and orientation measurements

144

Acceleration (m/s2)

AT (Waist)
25
20

Peak
15
10
5

Walking

Falling

0
0

Sitting
12

Walking

Standing up
16

20

24

28

Time (s)
Fig. 3.

An example of the temporal total acceleration data between different activities during 28 seconds.

Chest

0.9

0.8

0.7

Sensitivity (%)

Waist

Thigh

0.6

0.5

0.4

0.3

Fig. 2. The experimental mobile phones were attached to different locations


of human body, including the chest (pocket), waist (leather belt), and thigh
(pocket),

Chest

0.2

Waist
0.1

Thigh
0
0

for different activities of daily living, including walking,


sitting, standing up, and falling. Each data interval is 200
ms and we set a one-second time window to average these
measurements. Figure 3 shows a typical example of the total
acceleration between different activities during 28 seconds.
This figure demonstrates that a fall behavior is highly related
to the acceleration data, resulting in a dramatically change and
a peak acceleration value. There are four different users, with
different height and ages, to perform the experiments. Each
user continuously changed the activities for 28 seconds in a
fixed area, and repeated for 30 times. The collected data was
manually labeled as falling and other activities and divided
into training and testing sets. The trainings data is used to
find the threshold of the overall acceleration.
B. Experimental Results
Figure 4 shows the sensitivity versus specificity in different
phone-attached locations based on the overall acceleration.
For each location, we adjust the thresholds and compute the
sensitivity and specificity, as indicated by the markers of each
line. This is also known as receiver operating characteristic
curve. This figure clearly shows the tradeoff between sensitivity and specificity. From Fig. 4, a higher sensitivity is always
gained at the expense of a low specificity. That means that
determining the threshold should consider the balance between
sensitivity and specificity. Moreover, this figure shows that the
chest perform the best, achieving a 72.22% sensitivity and a
73.78% specificity at the same time. This is because when the

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Specificity (%)

Fig. 4. The sensitivity versus specificity in different phone-attached locations


based on overall acceleration.
TABLE I
T HE BEST DETECTION RESULTS AT THREE DIFFERENT ATTACHED
LOCATIONS

Chest
P eakmax > |AT |th

Sensitivity (%)

Specificity (%)

72.22

73.78

Waist

56.67

66.39

Thigh

53.33

57.22

mobile phone is attached in a higher location, the change of


acceleration is more significant to detect the fall.
Table I shows the best results from three locations individually. This table shows that when the phone is attached in the
waist, the sensitivity and specificity are reduced from 72.22%
to 56.67% and from 73.78% to 66.39%, respectively. The thigh
shows the worst performance, especially for the performance
of specificity. Next, Figure 5 shows the results using both
overall and vertical accelerations, as indicated by the red
squares. This figure shows a similar performance in these
two conditions. This means that using two parameters may
not provide a significant improvement. This may be explained
by that some correlations exist between these measurements.
Using an overall acceleration is sufficient for learning the fall

145

as sitting, walking and standing, with 72.22% sensitivity and


73.78% specificity. In addition, this study analyzes the tradeoff between sensitivity and specificity by plotting receiver
operating characteristic. The experiment further investigates
the impact of different locations where the phone attached and
also evaluates the extra power consumption on mobile phones.

0.9
0.8

Sensitivity (%)

0.7
0.6
0.5
0.4

ACKNOWLEDGMENT

0.3

The authors would like to thank the financial support


provided by National Science Council (NSC100-2221-E-155057).

0.2

Single Threshold
0.1

Two Threshold
0
0

0.2

0.4

0.6

0.8

R EFERENCES

Specificity (%)

Fig. 5.

Comparing the performance of one threshold and two thresholds.


Power Comsumption
100
Running Fall system

95

Normal State

Surplus Energy (%)

90

85

80

75

70

65

60
0

0.5

1.5
Time (hour)

2.5

Fig. 6. Compares the power consumption on mobile devices after running


the fall detection software.

behavior. Finally, Figure 6 compares the power consumption


on mobile devices after running the fall detection software.
This figure shows that the fall detection algorithm degrades
around 6% power more than a normal state without running
the detection software.
IV. C ONCLUSION
Falls are a major health risk for elderly people, diminishing
the quality of life or even resulting in death. This study
designs and implements a fall detection prototype for the
Android-based platform. We propose a low-cost fall detection
system, using the existing devices and wireless technology,
without the need for hardware modification, environmental
setup, and wearing external sensors. The proposed system
has three central components: sensing the accelerometer data
from mobile embedded sensors, learning the relationship between the fall behavior and collected data, and alerting preconfigured contacts through message while not receiving users
response. We adopt different algorithms to design fall detection algorithms, and conduct various experiments to evaluate
performance. The results show that the proposed system can
recognize the fall from human activities of daily living, such

[1] N. Lane, E. Miluzzo, H. Lu, D. Peebles, T. Choudhury, and A. Campbell,


A survey of mobile phone sensing, Communications Magazine, IEEE,
vol. 48, no. 9, pp. 140150, 2010.
[2] J. Dai, J. Teng, X. Bai, Z. Shen, and D. Xuan, Mobile phone based
drunk driving detection, in International Conference on Pervasive
Computing Technologies for Healthcare, pp. 18, 2010.
[3] S. Consolvo, D. W. McDonald, T. Toscos, M. Y. Chen, J. Froehlich,
B. Harrison, P. Klasnja, A. LaMarca, L. LeGrand, R. Libby, I. Smith,
and J. A. Landay, Activity sensing in the wild: a field trial of ubifit
garden, in Proceeding of the twenty-sixth annual SIGCHI conference
on Human factors in computing systems, CHI 08, pp. 17971806, 2008.
[4] D. Anderson, J. Keller, M. Skubic, X. Chen, and Z. He, Recognizing
falls from silhouettes, in Engineering in Medicine and Biology Society,
28th Annual International Conference of the IEEE, pp. 63886391,
2006.
[5] W.-Y. Shieh and J.-C. Huang, Speedup the multi-camera videosurveillance system for elder falling detection, in International Conference on Embedded Software and Systems, pp. 350355, 2009.
[6] C.-W. Lin and Z.-H. Ling, Automatic fall incident detection in compressed video for intelligent homecare, in International Conference on
Computer Communications and Networks, pp. 11721177, 2007.
[7] F. Sposaro and G. Tyson, ifall: An android application for fall monitoring and response, in International Conference of the Engineering in
Medicine and Biology Society, pp. 61196122, 2009.
[8] M. Alwan, P. Rajendran, S. Kell, D. Mack, S. Dalal, M. Wolfe, and
R. Felder, A smart and passive floor-vibration based fall detector
for elderly, in Information and Communication Technologies, vol. 1,
pp. 10031007, 2006.
[9] H.-W. Tzeng, M.-Y. Chen, and J.-Y. Chen, Design of fall detection system with floor pressure and infrared image, in International Conference
on System Science and Engineering, pp. 131135, 2010.
[10] D. Karantonis, M. Narayanan, M. Mathie, N. Lovell, and B. Celler,
Implementation of a real-time human movement classifier using a
triaxial accelerometer for ambulatory monitoring, IEEE Transactions
on Information Technology in Biomedicine, vol. 10, no. 1, pp. 156
167, 2006.
[11] C.-F. Lai, Y.-M. Huang, J. H. Park, and H.-C. Chao, Adaptive body posture analysis for elderly-falling detection with multisensors, Intelligent
Systems, IEEE, vol. 25, no. 2, pp. 2030, 2010.
[12] J. Dai, X. Bai, Z. Yang, Z. Shen, and D. Xuan, Perfalld: A pervasive
fall detection system using mobile phones, in IEEE International
Conference on Pervasive Computing and Communications Workshops,
pp. 292297, 2010.
[13] P. Jantaraprim, P. Phukpattaranont, C. Limsakul, and B. Wongkittisuksa,
Improving the accuracy of a fall detection algorithm using free fall
characteristics, in International Conference on Electrical Engineering/Electronics Computer Telecommunications and Information Technology, pp. 501504, 2010.
[14] T. Fawcett, An introduction to ROC analysis, Pattern Recogn. Lett.,
vol. 27, pp. 861874, 2006.

146

Вам также может понравиться