See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/228702476

Fall Detection and Activity Recognition Methods for the Confidence Project: A
Survey

Article · January 2010

CITATIONS READS

14 1,893

2 authors:

Bostjan Kaluza Mitja Lustrek

Jožef Stefan Institute Jožef Stefan Institute
40 PUBLICATIONS 622 CITATIONS 249 PUBLICATIONS 4,275 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Bostjan Kaluza on 31 May 2014.

The user has requested enhancement of the downloaded file.

 FALL DETECTION AND ACTIVITY RECOGNITION
METHODS FOR THE CONFIDENCE PROJECT: A SURVEY
Boštjan Kaluža, Mitja Luštrek
Department of Intelligent Systems
Jožef Stefan Institute
Jamova cesta 39, 1000 Ljubljana, Slovenia
Tel: +386 1 4773944; fax: +386 1 4251038
E-mail: [email protected]

ABSTRACT This paper surveys fall detection and activity recognition

The Confidence project is designed to support independent methods that could be used in the project. They are mostly
living of the elderly. The main goal of the project is based on the data from accelerometers, gyroscopes or
reconstructing the user’s posture to detect falls in real time cameras, which will not be employed in Confidence, so
and to detect other abnormalities in behavior. This paper they cannot be used directly. Nevertheless, some may be
presents a survey of methods for fall detection and activity adapted to the system to be developed in the project.
recognition and discusses how suitable they are for the Section 2 presents the hardware architecture of the
Confidence project. The presented methods use Confidence system and the basic ideas of the project.
accelerometers, velocity profiles and visual markers. Section 3 is dedicated to fall detection using acceleration-
based methods, velocity profiles and computer vision. In
Section 4 are presented some systems and methods for
1 INTRODUCTION activity and posture recognition. Section 5 summarizes the
survey.
While the expected lifetime in Europe is increasing, the
population growth is negative. The over-65 population is
anticipated to rise from 17.9 % in 2007 to 53.5 % by 2060 2 CONFIDENCE SYSTEM DESCRIPTION
[5] as shown in Figure 1. Thus an effort needs to be made
to ensure that elderly people could live longer The user will wear small low-cost wireless tags placed on
independently with minimal support of working-age the significant places on the body such as wrists, elbows,
population. shoulders, ankles, knees and hips. The precise number and
placement of tags will be defined during development. The
tags may be sewn into the clothes. The location of the tags
60 will be detected by a base station located in the apartment
or a portable device carried outside. This will make it
50 possible to reconstruct the user’s posture as shown in
40 Figure 2.
population %

Some tags will be placed in the user’s environment at

30 specific positions, such as bed, chair, sofa, table, etc. These
20 tags will enable the recognition of situations such as the
user lying in the bed or sitting in a chair.
10 Intelligent modules in the Confidence system will process
0 and analyze the data and raise the alarm if the user’s
2010 2020 2030 2040 2050 2060 behavior indicates a hazardous situation such as a fall,
stroke, epileptic attack etc. First, the system will make a
Figure 1: Projected dependency ratio of persons aged 65 phone call to the user and verify the user’s state by
and over in EU (27 countries). requiring the user to press a sequence of buttons or say a
certain word. If the user does not pick up the phone, the
The Confidence project will develop a ubiquitous care system will make a phone call to relatives, friends and even
system to support independent living of the elderly. The the emergency services if nobody answers.
system will be able to reconstruct the user’s posture and
detect abnormal situations, such as falls, loss of
consciousness or behaviors indicating a disease.
 simple threshold-based algorithms [3, 7]. Such algorithms
simply raise the alarm when the threshold value of
acceleration is reached. There are several sensors with
hardware built-in fall detection [1, 4, 9].
D. J. Willis [12] developed a more complex fall detection
system based on dynamic belief network models, which can
be used to model and produce conclusions about the state of
complex temporal environments. He used pressure
transducers besides accelerometers.
T. Zhang et al. [15] designed fall detector based on support
vector machine algorithm. Their method detected falls with
96.7 % success rate. Researches embedded an
accelerometer in a cell phone [14] and detected falls with
the proposed method. The cell phone was put in the pocket
of clothes or hanged around the neck. The system correctly
raised the alarm in 93.3 % of the cases.
Researchers using accelerometers give a lot of attention to
the optimal sensor placement on the body [3, 7]. A head-
worn accelerometer provides excellent impact detection
Figure 2: 3D locations of body tags obtained by motion sensitivity, but its limitations are usability and user
capture system. acceptance. A better option is a waist-worn accelerometer.
The wrist did not appear to be an optimal site for fall
detection. Some researchers made a step further and used
3 FALL DETECTION accelerometers for trying to recognize the impact and the
Falls among the elderly are the leading cause of injury, posture after the fall [8].
even death, and the loss of independent living. Detection In Confidence we will be able to derive accelerations from
and prevention of falls is consequently an important issue in the movement of the tags and use one of the threshold
the Confidence project. Figure 3 shows simulation of a fall algorithms if the localization precision is sufficient. Using
during the capturing of training data in the laboratory. more complex algorithms such are dynamic belief networks
and support vector machine may be considered if threshold
algorithms do not achieve the desired accuracy. Posture
after the fall can be obtained from the locations of the body
tags rather than from accelerations. Studies of the sensor
placement may be valuable for deciding where to place tags
in Confidence.

3.2 Velocity profile

A. K. Bourke and G. M. Lyons [2] introduced a threshold
algorithm to distinguish between normal activities and falls.
The ability to discriminate was achieved using a bi-axial
gyroscope sensor mounted on the trunk, measuring pitch and
roll angular velocities. They constructed a threshold
algorithm based on the investigation of peaks in the angular
velocity signal, angular acceleration and trunk angle change.
System proved 100 % successful in fall detection.
In Confidence we may use the proposed method. Having
sufficiently precise tag localization would enable us to
compute the required signals. Decreased precision may
cause problems when defining the threshold values.
G. Wu [13] studied unique features of the velocity profile
during normal and abnormal (i.e. fall) activities so as to
make the automatic detection of falls during the descending
Figure 3: A simulation of a fall. phase of a fall possible. Normal activities included walking,
rising from a chair and sitting down, descending stairs,
3.1 Accelerometer-based methods picking up an object from the floor, transferring in and out
of a tub and lying down on a bed.
The most common and simple methodology for fall The study provides exhaustive velocity parameters for fall
detection is using a tri-axial accelerometer with quite detection, which could be very useful in Confidence.
Supposing that we have sufficiently precise tag localization, The markers in the proposed system have the same role as
we could simply extract those velocities and detect falls tags in the Confidence system. The presented methods could
extremely quickly. be directly used for posture recognition. The Confidence
system will avoid data cleaning due to wireless transmission
3.3 Computer vision of 3D coordinates instead of detecting them with a camera.
Z. Fu et al. [6] described a vision system designed to detect A lot of researches investigated fall detection and posture
accidental falls in elderly home care applications. The recognition using video surveillance. They extracted features
system raised the alarm when a fall hazard was detected. from image signal and used one of the abovementioned
They used an asynchronous temporal contrast vision sensor, methods. Such feature extraction is not applicable to
which extracts changing pixels from the background and Confidence and consecutively not discussed.
reports temporal contrast. A lightweight algorithm computed
an instantaneous motion vector and reported fall events.
They were able to distinguish fall events from normal 5 CONCLUSION
human behavior, such as walking, crouching down and The presented systems and methods are successful in
sitting down. detecting falls and activities. Falls are mainly detected with
Temporal contrast vision sensor gives a similar human body accelerometers and gyroscopes. In the Confidence system,
contour as the Confidence human body model derived from the derivation of accelerations is questionable due to tag
tag locations. Z. Fu et al. defined a centroid event as the sampling rate. Methods using velocity profiles are more
average of the motion events, which was used for fall appropriate. In combination with furniture tags to detect the
detection. This method could be potentially useful when location of the fall, they may give very reliable results.
localization precision is reduced. Activity recognition is mainly based on processing video
images. System based on markers detected with a camera
give promising results.
4 ACTIVITY AND POSTURE RECOGNITION
E. M. Tapia et al. [11] presented a real-time algorithm for
automatic recognition of not only physical activities, but Acknowledgement
also, in some cases, their intensities, using five wireless The research leading to these results has received funding
accelerometers and a wireless heart rate monitor. Features from the European Community's Framework Programme
were extracted from time and frequency domains using a FP7/2007-2013 under grant agreement nº 214986.
predefined window size on the signal. The classification of
activity was done with C4.5 and Naïve Bayes classifiers.
The algorithm was evaluated using datasets consisting of 30 References
physical gymnasium activities collected from a total of 21
people. On these activities, they obtained a recognition [1] AlertOne Services, Inc. iLife™ Fall Detection Sensor.
accuracy of 94.6 % using subject-dependent training and http://www.falldetection.com, 2008-07-18.
56.3 % using subject-independent training.
[2] A.K. Bourke, G.M. Lyons. A threshold-based fall-
Although the researchers were using a different monitoring
detection algorithm using a bi-axial gyroscope sensor.
system, the presented system successfully dealt with real-
Medical Engineering & Physics, vol. 30, issue 1, pp.
time activity recognition, which could be useful in
84–90, 2006.
Confidence.
G. Qian et al. [10] introduced a gesture-driven interactive [3] A. K. Bourke, C. N. Scanaill, K. M. Culhane, J. V.
dance system capable of real-time feedback for performing O'Brien, and G. M. Lyons. An optimum accelerometer
arts. They used 41 markers on the body measured by 8 configuration and simple algorithm for accurately
cameras with the frame rate of 120 Hz. They developed data detecting falls. In Proceedings of the 24th IASTED
cleaning algorithm to tackle missing markers and international Conference on Biomedical Engineering,
constructed a human body model. The model was used to pp. 156–160, 2006.
extract features such as torso orientation, joint angles
[4] K. Doughty, R. Lewis and A. McIntosh. The design of
between adjacent body parts etc., which was used to
a practical and reliable fall detector for community
represent different gestures. Each gesture was statistically
and institutional telecare. Journal of Telemedicine and
modeled with a Gaussian random vector. Gesture matching
Telecare, vol. 6, pp. 150–154, 2000.
was performed to determine whether a new pose is inside
the gesture space and if it is, the likelihoods of the feature [5] Eurostat. http://epp.eurostat.ec.europa.eu, 2008-09-09.
vector given different gestures were computed.
Experimental results with two dancers performing 21 [6] Z. Fu, E. Culurciello, P. Lichtsteiner, T. Delbruck. Fall
different gestures achieved gesture recognition rate of detection using an address-event temporal contrast
99.3 %. vision sensor. In Proceedengs of the IEEE International
 Symposium on Circuits and Systems – ISCAS 2008,
pp. 424–427, 2008.
[7] M. Kangas, A. Konttila, P. Lindgren, I. Winblad, T.
Jamsa. Comparison of low-complexity fall detection
algorithms for body attached accelerometers. Gait &
Posture, vol. 28, issue 2, pp. 285–291, 2008.
[8] M. Kangas, A. Konttila, I. Winblad, T. Jamsa.
Determination of simple thresholds for accelerometry-
based parameters for fall detection. In Proceedings of
the 29th Annual International Conference of the IEEE,
Engineering in Medicine and Biology Society, pp.
1367–1370, 2007.
[9] N. Noury, P. Barralon, G. Virone, P. Boissy, M.
Hamel, P. Rumeau. A smart sensor based on rules and
its evaluation in daily routines. In Proceedings of the
25th Annual International Conference of the IEEE,
Engineering in Medicine and Biology Society, pp.
3286–3289, vol. 4, 2003.
[10] G. Qian, F. Guo, T. Ingalls, L. Olson, J. James, T.
Rikakis. A Gesture-Driven Multimodal Interactive
Dance System. In Proceedings of the International
Conference on Multimedia and Expo, Taipei, Taiwan,
China, 2004.
[11] E. M. Tapia, S. S. Intille, W. Haskell, K. Larson, J.
Wright, A. King, R. Friedman. Real-Time Recognition
of Physical Activities and Their Intensities Using
Wireless Accelerometers and a Heart Rate Monitor. In
Proceedengs of the 11th IEEE International
Symposium on Wearable Computers, 2007, pp. 37–40,
2007.
[12] D. J. Willis. Ambulation Monitoring and Fall
Detection System using Dynamic Belief Networks. PhD
Thesis. School of Computer Science and Software
Engineering, Monash University, 2000.
[13] G. Wu. Distinguishing fall activities from normal
activities by velocity characteristics. Journal of
Biomechanics, vol. 33, issue 11, pp. 1497–1500, 2000.
[14] T. Zhang, J. Wang, P. Liu and J. Hou, Fall Detection
by Embedding an Accelerometer in Cellphone and
Using KFD Algorithm. International Journal of
Computer Science and Network Security, vol. 6, no.
10, 2006.
[15] T. Zhang, J. Wang, P. Liu and J. Hou. Fall Detection
by Wearable Sensor and One-Class SVM Algorithm.
Lecture Notes in Control and Information Science,
issue 345, pp. 858–863, 2006.

View publication stats