30th Annual International IEEE EMBS Conference

Vancouver, British Columbia, Canada, August 20-24, 2008

A performance comparison of accelerometry-

based step detection algorithms on a large, non-
laboratory sample of healthy and mobility-
impaired persons
Michael Marschollek, Mehmet Goevercin, Klaus-Hendrik Wolf, Bianying Song, Matthias Gietzelt,
Reinhold Haux, and Elisabeth Steinhagen-Thiessen

 for more advanced algorithms for gait analysis. Step counters

Abstract ² Accelerometers are frequently used for activity
assessment and as reference devices for counting steps. Their
SHUIRUPDQFH RQ KHDOWK\ VXEMHFWV¶ GDWD LV JRRG EXW WKHUH DUH
doubts as to their applicability on elderly and mobility-impaired
subjects. Furthermore, only few step detection algorithms have
been published so far, and their performance has not been
evaluated on a large, non-laboratory sample. The aim of this
paper is to compare the performance of four freely accessible
are frequently used to promote a healthier lifestyle by
providing a feedback measure of individual activity [2].
Despite the number of commercial products that are
already available, only few algorithms for step detection
have been published so far (e.g. [3]). Furthermore, despite
the use of accelerometry as a reference method for step
counting or activity assessment for medical purposes (e.g.
accelerometry-based step detection algorithms in a non-
laboratory setting. Two samples of healthy persons (n=140) and
mobility-impaired, geriatric in-patients (n=10) wore a single
triaxial accelerometer on a waist-belt during unconstrained
walking. The relative error rate of the four algorithms on the
two samples was compared with reference video recordings. All
four algorithms show a fairly poor performance on healthy
VXEMHFWV¶-30.8% relative error rate) and especially geriatric
patients¶ GDWD -62.1%). Among the tested ones, a simple
autocorrelation algorithm works best on both data sets
together. More complex algorithms might work better, and
more research is needed to evaluate the accuracy of step
detection methods on mobility-impaired subjects.
Index Terms ² wearable sensors, step detection, elderly
people, home monitoring, triaxial accelerometry
I. INTRODUCTION
T HE assessment of activity parameters with sensory
measurements is a major focus of research.
Accelerometry is frequently used as a simple and
inexpensive, yet effective method [1]. A reliable detection of
DVXEMHFW¶VVWHSVLQDQDFFHOHURPHWHUVLJQDOLVLPSRUWDQWQRW
only for counting steps or measuring activity levels
respectively energy expenditure, but also as a precondition
Manuscript submitted April 1st, 2008.
M. Marschollek, K.-H. Wolf, M. Gietzelt, B. Song and R. Haux are with
the Peter L. Reichertz Institute for Medical Informatics of the University of
Braunschweig ± Institute of Technology and Medical School Hannover,
Muehlenpfordtstrasse 23, 38106 Braunschweig, Germany (phone: +49-531-
3919504; fax: +49-531-3919502; e-mail: michael.marschollek@plri.de).
M. Goevercin and and E. Steinhagen-Thiessen are with the Department
for Geriatric Medicine at the Charité, Berlin (Germany).
978-1-4244-1815-2/08/$25.00 ©2008 IEEE.
[4] WR WKH DXWKRUV¶ NQRZOHGJH WKHUH DUH QR ODUJH-scale
studies that compare the performance of step detection
algorithms in a non-laboratory setting, including frail and
mobility-impaired elderly subjects. This may be of special
importance as commercial devices have proven to work well
RQ KHDOWK\ VXEMHFWV¶ GDWD [5], but sometimes encounter
problems when used with frail and elderly subjects with
walking speed below 0.8 m/s [6].
II. OBJECTIVES
The aim of our research for this paper is to compare the
performance of four freely accessible algorithms for step
detection from accelerometric data on a large, non-laboratory
sample of healthy as well as mobility-impaired subjects.
III. METHODS
A. Study data sets and devices
Two studies have been conducted to obtain the data
necessary for this investigation. The first study was done in
the context of a public event within the City of Science-year
in Braunschweig, Germany, between June 14th to June 17th,
2007. Altogether 211 persons aged between 7 and 88 years
participated and completed a test procedure that was aimed
at assessing the ability to perform daily activities. The
procedure consisted of a 20 m level walk, walking up and
down DIOLJKWRIVL[VWDLUVDµ7LPHG8S *R¶WHVW [7] and
the Eisfeld test for functional balance during daily activity
tasks [8]. The walkway was filmed, and the subjects wore a
1319
wireless triaxial accelerometer board (Freescale
RD3152MMA7260Q) on a belt around the waist. Due to
technical failures, both in the video and sensory equipment,
71 data sets had to be excluded, so that 140 sets were used in
this study.
The second study was conducted at the Charité
Department for Geriatric Medicine (Berlin, Germany) on
June 20th, 2007. Ten persons with varying degrees of
functional mobility impairment completed a 20 m level walk
DQG D µ7LPHG 8S  *R¶ WHVW $OWRJHWKHU  WHVWV ZHUH
recorded with a video camera, and the subjects also wore the
above-mentioned accelerometer device.
)RUWKLVVWXG\RQO\WKHµZDONLQJµVHFWLRQVRIWKHGDWDVHWV
counted by two of the co-authors in the video recordings.
IV. RESULTS
Table I shows the results for the four implemented step
detection algorithms in terms of the relative error as
compared to the video analysis.
The custom-designed Wolf algorithm shows the best
performance (8.4% relative error) for the sample of healthy
subjects, but works poorly (60.0%) on data of mobility-
impaired elderly persons. The Dual Axis method performs in
much the same way. The Pan-Tompkins algorithm shows
almost equal error rates (30.8% vs. 28.1%) for both groups
of subjects, and the autocorrelation method works slightly
were used. better on healthy VXEMHFWV¶GDWDVHWV
B. Step detection algorithms
TABLE I
The accelerometric data were analyzed using software RELATIVE ERROR RATES OF FOUR STEP DETECTION ALGORITHMS ON
scripts written for the R statistics environment [9] and in ACCELEROMETRIC DATA FROM HEALTHY PERSONS (N = 140) AND MOBILITY-
IMPAIRED GERIATRIC PATIENTS (N=10)
JAVA. Complying with the rules of Good Scientific Practice
(e.g. [10]) all software modules may be requested from the
relative error
authors along with the detailed step count results. Altogether
healthy mobility-
four freely accessible algorithms were implemented: persons impaired
algorithm
x Pan-Tompkins method: This method has originally (n=140) persons (n=10)
been designed to detect QRS complexes in ECGs Pan-Tompkins 30.8% 28.1%
[11], and has recently been proposed for step Dual Axis 10.0% 62.1%
detection [3] using the z-axis (longitudinal). The Wolf 8.4% 60.0%
basic principle is a threshold-based peak detection Autocorrelation 17.5% 29.3%
that follows several filtering, amplification and
signal modulation steps (for details see [3]). V. DISCUSSION
x Dual-Axial method: This method combines the
signals of two accelerometer axes, the z-axis A. Step detection algorithm performance
(longitudinal) and the x-axis (sagittal). During heel- The results show that all four tested, freely accessible
strike both signals show negative peaks at the same algorithms do not perform very well on different data sets of
time. Following several signal preprocessing steps, healthy and mobility-impaired persons. Error rates are
a threshold-based algorithm is used to detect these considerably higher than those reported in literature, e.g. by
signal peaks [3]. Mathie, who achieves an accuracy of up to 100% in a
x Wolf method: A threshold-based peak detection laboratory setting (n=26) using an algorithm based on a
method that was developed by one of the co-authors Fourier transform of the z-axis signal [12]. She notes,
during a step detection competition at the however, that the performance of this approach decreases
University of Braunschweig. The detection with rising step variability which is frequently found in
threshold is adapted with regard to the signal pathological gait [13]. Ryan et al. report an error of less than
amplitudes in sliding windows. The algorithm was 1.2% for detecting steps with a commercial device with a
developed on ten random exemplary data sets proprietary algorithm in a treadmill walking exercise of
recorded during the first study (healthy persons, healthy adults (n=20) [5]. In a recent study Dijsktra et al.
n=140) mentioned above. compare the step detection accuracies of a commercial
x Autocorrelation: A simple autocorrelation analysis of accelerometer and pedometer for older adults (n=20) and
the accelerometric signal which identifies peaks by
using a similarity measure.
C. Performance analysis
For each of the two data sets ± healthy subjects (n=140)
and mobility-impaired geriatric patients (n=10) ± the four
above-mentioned algorithms were used to count steps, and
the results were compared with the real number of steps as
1320
patients ZLWK 3DUNLQVRQ¶V GLVHDVH Q  7KH\ ILQG DQ
overall error rate of 7.4% resp. 6.9% (underestimated) [14].
Table II summarizes the results found in literature.
It has to be noted that ± among the algorithms tested in
this study ± those that are less adapted to the specific
problem of step detection perform better on the two different
sets (healthy and mobility-impaired). The autocorrelation
algorithm, e.g., works fairly well on both data sets, although
it contains no a-priori knowledge of accelerometric signal
patterns. The disadvantage of this algorithm is, of course,
that the data need to be preprocessed so that they only
contain walking data, because otherwise any periodic signal
pattern might be detected. Both the Dual Axis and the Wolf
methods are adapted to specific characteristics of heel strike
events, and it can be concluded that they do not work well
for elderly persons with musculoskeletal impairments
because the mechanics of walking, especially the kinetic
energy values, change with advancing age [13].
TABLE II
SUMMARY OF STEP DETECTION ACCURACIES FOR ACCELEROMETRIC DEVICES
FOUND IN LITERATURE
Ref. # algorithm n type of subjects accuracy
[12] proprietory 26 healthy - 100.0%
[5] proprietory 20 healthy > 98.8%
[14] proprietory 20 older adults 92.6%
[14] proprietory 32 3DUNLQVRQ¶VGLVHDVH 93.1%
B. Limitations and future work
There are some limitations to our comparative study that
have to be mentioned with regard to the generalizability of
our results:
1) We have implemented and tested only simple
algorithms that are freely accessible and computationally not
very demanding. We have not included commercial
algorithms or more complex pattern classification methods in
our study, such as the Template method [3], which might
work better.
2) We are aware that the good results of our own
algorithm (Wolf) RQ WKH KHDOWK\ SHUVRQV¶ GDWD can be
attributed to the fact that a random sample (n=10) of these
sets was used for its development. Thus the performance
cannot be generalized to other data.
3) For this study we have only used accelerometric data
recorded with a single accelerometer that was worn on a belt,
FORVH WR WKH ERG\¶V FHQWHU RI JUDYLW\ Apart from practical
simplicity, this location was chosen because of its limited
obtrusiveness for the subject and the possibility to extract
other parameters such as e.g. pelvic sway, which can be used
to assess trunk balance. Data from other locations such as the
distal tibiae, ankles or feet might yield different results for
WKH DOJRULWKPV¶ SHUIRUPDQFH Han et al. use accelerometers
on both ankles of healthy persons (n=5) and patients with
3DUNLQVRQ¶V GLVHDVH (n=5), and report step detection
accuracies of 93% and 94%, respectively [15]. Furthermore,
a combination of accelerometry and other sensory methods,
e.g. measuring the expansion of the gastrocnemius muscle
[16] or using gyroscopes [17], may further improve detection
accuracy.
Our future work will include not only a refinement of step
detection methods, but also further work on more complex
activity pattern classification algorithms [18-20], which may
be even more valuable in terms of individual feedback with
1321
regard to the promotion of healthy lifestyles. Furthermore we
will work on the standards-based representation of sensory
data and their integration in health information system
architectures to safeguard their re-usability and accessibility
[21, 22].
VI. CONCLUSION
We have compared four freely accessible algorithms for
detecting steps in accelerometric data sets with regard to
their performance not only on a non-laboratory sample of
healthy subjects, but also on a sample of geriatric patients
with mobility impairments. The results show that none of the
algorithms works very well, and that the more customized
the algorithms are to specific samples, the worse they work
on a sample with different motion characteristics (e.g. very
old persons). More complex pattern classification algorithms
may resolve this problem, and more research in this field is
needed to prove the validity of step detection algorithms for
the elderly.
VII. ACKNOWLEDGEMENT
The authors thank Nina Reichwaldt for her valuable
support in processing the video recordings for this study.
REFERENCES
[1] M. J. Mathie, A. C. Coster, N. H. Lovell, and B. G. Celler,
"Accelerometry: providing an integrated, practical method for long-term,
ambulatory monitoring of human movement," Physiol Meas, vol. 25,
pp. R1-20, 2004.
[2] K. A. Croteau, N. E. Richeson, B. C. Farmer, and D. B. Jones, "Effect of
a pedometer-based intervention on daily step counts of community-
dwelling older adults," Res Q Exerc Sport, vol. 78, pp. 401-6, 2007.
[3] H. Ying, C. Silex, A. Schnitzer, S. Leonhardt, and M. Schiek,
"Automatic Step Detection in the Accelerometer Signal," Proceedings of
Body Sensor Networks BSN2007, vol. 80-5, 2007.
[4] C. Tudor-Locke, S. B. Sisson, S. M. Lee, C. L. Craig, R. C. Plotnikoff,
and A. Bauman, "Evaluation of quality of commercial pedometers," Can
J Public Health, vol. 97 Suppl 1, pp. S10-5, S10-6, 2006.
[5] C. G. Ryan, P. M. Grant, W. W. Tigbe, and M. H. Granat, "The validity
and reliability of a novel activity monitor as a measure of walking," Br J
Sports Med, vol. 40, pp. 779-84, 2006.
[6] K. L. Storti, K. K. Pettee, J. S. Brach, J. B. Talkowski, C. R.
Richardson, and A. M. Kriska, "Gait speed and step-count monitor
accuracy in community-dwelling older adults," Med Sci Sports Exerc,
vol. 40, pp. 59-64, 2008.
[7] D. Podsiadlo and S. Richardson, "The timed "Up & Go": a test of basic
functional mobility for frail elderly persons," J Am Geriatr Soc, vol. 39,
pp. 142-8, 1991.
[8] K. Eisfeld, Motor comptence and ressources in senior age [Motorische
Kompetenz und motorische Ressourcen im Seniorenalter]: Kovac, 2005.
[9] WU Wien, "The R Project for Statistical Computing," Available from:
http://www.r-project.org, last accessed Jan 10th, 2008.
[10] Deutsche Forschungsgemeinschaft, "Proposals for Safeguarding
Good Scientific Practice 1998," vol. 2006. Weinheim: Wiley-VCH,
1998.
[11] J. Pan and W. Tompkins, " A Real-Time QRS Detection Algorithm,"
IEEE Transactions on Biomedical Engineering, vol. 32, pp. 230-6,
1985.
[12] M. Mathie, "Monitoring and Interpreting Human Movement Patterns
Using a Triaxial Accelerometer," PhD thesis, Faculty of Engineering,
The University of New South Wales (Sydney), 2003.
[13] J. Perry, Gait Analysis - Normal and Pathological Function.
Thorofare, NJ: Slack, 1992.
[14] B. Dijkstra, W. Zijlstra, E. Scherder, and Y. Kamsma, "Detection of
walking periods and number of steps in older adults and patients with
Parkinson's disease: accuracy of a pedometer and an accelerometry-
based method," Age Ageing, 2008.
[15] J. Han, H. S. Jeon, B. S. Jeon, and K. S. Park, "Gait detection from
three dimensional acceleration signals of ankles for the patients with
3DUNLQVRQ¶VGLVHDVHProceedings of the International Conference on
Advanced Information and Telemedicine Technologies for Health
(ITAB), Ioannina (Greece), 2006.
[16] G. Maccioni, V. Macellari, and D. Giansanti, "Design and
construction of step counters for disable people: preliminary experience
at the Italian Institute of Health," Conf Proc IEEE Eng Med Biol Soc,
vol. 2007, pp. 4927-9, 2007.
[17] A. M. Sabatini, C. Martelloni, S. Scapellato, and F. Cavallo,
"Assessment of walking features from foot inertial sensing," IEEE Trans
Biomed Eng, vol. 52, pp. 486-94, 2005.
[18] M. Marschollek, K. H. Wolf, M. Plischke, and R. Haux,
"Classification of activities of daily life from long-term realistic multi-
sensor data," Proceedings of IEEE Health Pervasive Systems HPS06,
pp. 30-4, 2006.
[19] M. Marschollek, W. Ludwig, I. Schapiewski, E. Schriever, R.
Schubert, H. Dybowski, H. Meyer zur Schwabedissen, J. Howe, and R.
Haux, "Multimodal home monitoring of elderly people ± first results
from the LASS study," Proceedings of the 21st IEEE International
Conference on Advanced Information Networking and Applications
(AINA 2007), Niagara Falls, Ontario, Canada, vol. 2, pp. 815-19, IEEE
Computer Society.
[20] K. H. Wolf, A. Lohse, M. Marschollek, and R. Haux, "Development
of a Fall Detector and Classifier based on a Triaxial Accelerometer
Demo Board," Proceedings of UbiComp 2007, Innsbruck, pp. 210-3.
[21] M. Marschollek, K. H. Wolf, O. J. Bott, M. Geisler, M. Plischke, W.
Ludwig, A. Hornberger, and R. Haux, "Sustainable ubiquitous home
health care - Architectural considerations and first practical
experiences," MEDINFO2007 (Brisbane), 2007, p. 8-12.
[22] O. J. Bott, M. Marschollek, K. H. Wolf, and R. Haux, "Towards New
Scopes: Sensor Enhanced Regional Health Information Systems - Part 1:
Architectural Challenges," Methods Inf Med, vol. 46, pp. 476-83, 2007.

Michael Marschollek received his MD and Dr. med. degrees from the
Medical School in Hannover (Germany) in 1999. He received the M.Sc.
degree in computer science from the University of Braunschweig ± Institute
of Technology in 2004. He is currently a senior research assistant at the
Peter L. Reichertz Institute for Medical Informatics at the same university.
His research interests are health-enabling technologies, tele home care,
wearable and pervasive health care, and data mining and pattern
recognition in medicine.

1322