Proceedings of the 2011 IEEE

International Conference on Robotics and Biomimetics

December 7-11, 2011, Phuket, Thailand

A Wearable Wireless Fall Detection System with Accelerators

Diansheng Chen, Wei Feng, Yu Zhang , Xiyu Li, and Tianmiao Wang

g

Abstract— Falls in elderly is a major health problem and a with the ground and the posture of lying on a horizontal
cost burden to social services. Thus automatic fall detectors are floor.
needed to support the independence and security of the elderly. Because of the severe injuries caused by fall events and
The goal of this research is to develop a real-time portable the following long time lying, a fall detection system is
wireless fall detection system, which is capable of essential for timely salvation. A number of different
automatically discriminating between falls and Activities of
Daily Life (ADL). The fall detection system contains a portable
automatic fall detectors have been developed, using
fall-detection terminal and a monitoring centre, both of which body-attached accelerators or gyroscopes or even both.
communicate with ZigBee protocol. To extract the features of Williams et al. [4] first introduced an automatic fall detector
falls, falls data and ADL data obtained from young subjects are in 1998, which included a piezoelectric sensor to detect the
analyzed. Based on the characteristics of falls, an effective fall impact shock as well as mercury tilt switch to detect the
detection algorithm using tri-axis accelerometers is user’s body position. This research was continued by
introduced, and the results show that falls can be distinguished
from ADL with a sensitivity over 95% and a specificity of
Doughty et al. [5], and the company Tunstall (Whitley
100%, for a total set of 270 movements. lodge, Yorkshire, England) developed a commercial vision
base on this system. The system claimed a fall-detection
I. INTRODUCTION accuracy of 90% however fall testing was performed using a
wooden mannequin and the system has not been tested on
F alls are one of the major health risks among elderly
community by result in increase in mortality, morbidity,
disability, and frailty [1]. 50% of people over 80 years old
elderly people just surveyed. Bourke et al. designed a
threshold-based tri-axial accelerometer fall detection
algorithm, which achieved sensitively of more than 90% and
have experienced falls each year, which means falls affect
a specificity of more than 99%, using young healthy subjects
tens of millions of the elderly people throughout the world
performing falls and normal activities of activities of daily
[2]. 40% of hospital admissions of older adults were the
living (ADL) [6]. Also, Bourke et al. developed a custom
result of fall-related injuries, leading to an average length of
vest to detect fall events and organized a long-term fall
stay of 11.6 days [1].
detection test, over 833 hours of monitoring was performed
A fall is defined as “unintentionally coming to ground, or
some lower level not as a consequence of sustaining a from elderly subjects, during normal daily activities [7].
violent blow, loss of consciousness, sudden onset of However, during the test, 42 fall-alerts were detected but no
paralysis as in stroke or an epileptic seizure” [3]. With the actual falls were reported, and only 9 were received at the
development of research, the definition extended to include care centre. Thus, further research is required to eliminate
falls result from dizziness and syncope, and other the number of false-positives and improve the
consequences of an epileptic fit or cardiovascular collapse communication quality. Kangas et al. [8] evaluated three
such as postural hypotension and transient ischemic attacks, different fall detection algorithms, using tri-axial
all of which may have resulted in a fall and its potential accelerators attached at the waist, wrist, and head. The result
consequences [2]. Therefore, the key criterions of falls are showed that fall detection using a tri-axial accelerator worn
unintentional movements toward to the ground, an impact at waist or head with a threshold-based algorithm can
achieve high sensitivity and specificity, over 90%
respectively.
Manuscript received July 19, 2011. This work was supported by the
Generally speaking, different fall detection systems detect
National High Technology Research and Development Program of China different phases of fall events as their algorithms. Chosen
(863 Program) (No.2008AA040207). the impact with the ground as the cut-off point, the fall
Diansheng Chen is with Robotics Institute of Beihang University,
detection systems could be divided into three types: (1)
Beijing, China (phone: 86-010-82339089; e-mail: chends@[Link]).
Wei Feng is with Robotics Institute of Beihang University, Beijing, Early detection before the impact, based on velocity profiles
China (e-mail: fngwii@[Link]). [9] or angular characteristics [10]; (2) Direct detection of the
Yu Zhang is with Robotics Institute of Beihang University, Beijing, impact, based on high acceleration [11] [12] or a rapid
China (e-mail: yhhyzhyu@[Link]).
Xiyu Li is with Robotics Institute of Beihang University, Beijing, China change in acceleration [13]; (3) Body status detection after
(e-mail: lixiyu20@[Link]). the impact, based on body posture or active level [14] [15].
Tianmiao Wang is with Robotics Institute of Beihang University, Most of the reported fall detection applications were
Beijing, China (e-mail: itm@[Link]).

‹,((( 2259
prototypes for principle research, so the real-time, (b) (c)
portability and comfort were not considered well in these
studies. In addition, there is a lack of supporting network to
locate the fall events for timely salvation. Thus, this research
aims to develop a real-time portable wireless fall detection
system with a high accuracy to discriminate between falls
and ADL, in order to provide timely help for the elderly by
locating service when fall happens.

II. SYSTEM DESIGN

A. System Overview
The wireless fall detection system contains two main
modules (Fig. 1a), one is the fall-detection terminal (Fig. 1b)
worn at the waist by the elderly, and the other is the remote
Fig. 2. Illustration of the fall detection system: (a) Block diagram of the
monitoring centre. The two main modules communicate by
system composition. (b) Photograph of the fall detection terminal. (c)
the wireless network. The wireless fall-detection terminal is Photograph of the ZigBee gateway.
the size of 56.6mmh89.1mmh22.3mm (lengthhwidthh
thickness), which is small-size and suitable for belt worn. It
consists of an inertial sensor unit, a wireless communication
unit, a microcontroller, and a fall detection algorithm. A
tri-axis accelerometer (MMA7260QT, Freescale
Semiconductor, Inc.) is used as the inertial sensor. A
high-performance, low-power AVR microcontroller
(ATmega16L, Atmel Corporation) is used for data
processing and housing the real-time fall detection
algorithm. A wireless communication SoC (System on Chip) Z
(CC2430, Texas Instruments, Inc.) is used as the wireless
communication unit, which transported kinematic
parameters and fall alert messages in ZigBee protocol. Since X
Y
the fall-detection algorithms are implanted into the
fall-detection terminal, the monitor centre is set up to receive Fig. 1. Worn position of fall detection terminal and sensitivity axes of the
and store kinematic parameters and fall alert messages for accelerometer
long-term analysis and personal analysis, no fall detection
B. The Fall-Detection Algorithm
algorithms are operated in the PC. Thus, without the monitor
centre, the fall-detection terminal can also accomplish the In a broad view, fall detection is a process of extracting
task of fall detection. This structure of system assures the features of falls by signal processing technology to
real-time and robustness of the fall detection system. automatically discriminate between falls and ADL events.
The positions of the inertial sensors are significant for Thus, kinematic parameters should be constructed to study
body attached fall detection system. The worn positions the features of falls.
close to the centre of the gravity reflect the posture of the 1) Parameters: The total sum acceleration STOT
trunk more truly. Considering the accuracy and portability describes the activity level of the body, which contains both
of the system, the waist worn fashion is adopted in this paper the dynamic acceleration and the gravity acceleration. It was
(Fig. 2). Sensitivity axes of the accelerometer are also shown calculated from sampled data as indicated in Eq. (1):
in Fig. 2.
(a) STOT ( Ax )2  ( Ay )2  ( Az )2 (1)

Fall Detection System where Ax, Ay, Az, are the acceleration in the x, y, and z
Fall Detection Terminal
axes, respectively. STOT has a value 1g, which equates with
Tri-axis
Accelerometer
Monitor Centre gravity acceleration, when the body is static.
The tilt angle TA describes the body’s posture, when the
PC
Zigbee
Micro Controller
δData Processing & Real- Zigbee SoC Gateway
(Data Storage &
Long-term Analysis)
body is static. Its value was defined as indicated in Eq. (2),
time Fall Detectionε
(3), and (4):
Alarm Unit

2260
 TAx arccos( Ax / G) u (180 / pi) (2)
5.5

5
TAy arccos( Ay / G) u (180 / pi)
(3)
TAz arccos( Az / G) u (180 / pi)

Accleration (g)
4
(4)
where Ax, Ay, Az, are the acceleration in the x, y, and z 3
axes, respectively. G is gravity acceleration (=1g). As the
sensor position described above, the value of TAx, TAy, and 2
TAz are 90°, 90°, and 0°, respectively, when the user is
standing. 1
Fig. 3 shows the variations of STOT and Az during Fall to left Fall to right Backward fall Forward fall Lying Sitting Walking Running

different falls and ADL events simulated by subjects.

Fig. 4. Distribution of STOT maximum in falls and ADL
During a classical fall (Fig.3a), STOT mildly increases in the
falling phase, and soars when the body bearing the impact,
Fig.4 illustrates the distribution of STOT maximum in
then maintains about 0g in the lying phase. In the meantime,
falls and ADL. Data indicates that the peak accelerations of
Az rises from -1g to 0g in the falling phase, and keeps
falls are higher than those of ADL in most cases, except
around 0g in the lying phase, which indicates the posture
compared with running.
change of the body. The time trajectories of different ADL
Through analyzing a mass of data of experiments, two
events (Fig.3b) are diverse but different from those of falls.
main characteristics of falls events are abstracted. First,
STOT sharply increased, when the body was falling and
(a)
3 Impact phase
bore the impact with the ground. Second, after the impact,
Falling phase
STOT
Az
Ay
the value of Az remained around 0g, caused by the posture
2 change after the falls. The former is a transient variation of
Lying phase
sum acceleration, while the latter is a steady-state variation.
Acceleration (g)

1 Since seldom ADL events have both these variations in

kinetics characteristics, high specificity and sensitivity could
0
be achieved by using the above features.
2) Algorithms for Fall Detection: The algorithm is based
-1
on the detection of STOT and TA value in different phases
of fall events. Acceleration samples are calculated for STOT
-2.3
0 2 4 6 8
Time (sec)
10 12 14 16 and TA, which demonstrates the level of the body motions
and the posture of the body. Six key parameters include
(b) three thresholds and three delay times are introduced in this
Sitting down & standing up Running
4 4 algorithm. The three thresholds are HAT (High Acceleration
2
Threshold), LAT (Low Acceleration Threshold) and AT
Acceleration (g)

2
(Angle Threshold). The process flow of detection algorithm
0 0 in microcontroller unit is shown in Fig. 5. Follows are three
-2
logical judgments of fall detection algorithm:
-2
a). If STOT value is greater than HAT, the body may be in
-4
0 4 8 12 16
-4
0 2 4 6 8 10
fierce motions or bearing impact.
2
Walking
2
Lying down b). If STOT value is lower than LAT, the body may be in
the static status after the fall impact.
Acceleration (g)

1 1
c). TAz describes the tilt angle of the trunk, so when the
0 0
value of TAz is greater than AT, the subject may be lying.
-1 -1 The three delay times describe the time length of different
-2 -2
phases and intervals between them in fall events. Delay time
0
Time5(sec) 10 0 5
Time (sec)
10 15
1 is the time length of the fall impact, Delay time 2 is the
interval between the impact and the static status when the
subject lying on the ground and Delay time 3 presents the
Fig. 3. Variation diagram of STOT and Az in falls and ADL: (a) Time
trajectory of STOT and Az during a forward fall. (b) Time trajectory of
interval time between the first and the second detection of
STOT and Az during different ADL events the truck angle to confirm the body posture.

2261
 female, aged 22-26 years were asked to perform the
intentional falls and ADL indoors and outdoors. The mean
height and mass of volunteers were 172.3 cm and 64.5 kg,
Delay respectively. The test category was set up, referenced the
Data time2
Preparation research of Noury et al. on scenarios for the evaluation of
fall detectors [4] (Table I). There are 5 main kinds of test
N subcategories: forward fall, backward fall, left-side fall,
|TAz|> AT right-side fall and ADL, each subcategory contains 3-6
N events. In the experiments, each event was performed at
STOT>HAT
least three times. A 25-cm high mattress was used for indoor
Y
intentional falls.
Y Delay Since the output of the fall detection is binary, there are
time3 four possible results: true positive (TP), means the device
alerts when a fall occurs; false positive (FP), means the
Delay
device alerts when no fall occurs; true negative (TN), means
time1
N the device does not alert when no fall occurs; false negative
|TAz|> AT (FN), means the device does not alert when a fall occurs.
Two criteria were proposed to evaluate the fall detection
N device: sensitivity is the capacity to detect a fall;
STOT<LAT Y
specificity is the capacity to detect only a fall.
Sensitivity is the capacity to detect a fall (5):
Y Fall Alert TP
sensitivity (5)
TP  FN
Fig. 5. Process flow of detection algorithm Specificity is the capacity to detect only a fall (6):
TN
III. EXPERIMENTS AND RESULTS specifivity (6)
TN  FP
A. Experiment Design B. Experiment Results
TABLE I Through an experiment included 195 simulated falls and
EXPERIMENT CATEGORY 75 ADL, the algorithm could achieved a sensitivity over
Category Instructio Kind 97% and a specificity of 100% (Table 2). All the detected
falls triggered the alarm in 2.5s since the falls happened, and
Backward Fall Ending lying Positive
Ending lateral position Positive
the alert messages were sent to the monitor centre
With recovery Negative immediately. All ADL events were correctly identified as
Forward fall On the knees (only on mattress) Positive non-falls. Only 2 backward falls and 1 left-side fall were
With for forward arm protection Positive undetected.
Ending lying flat Positive
With rotation, ending in the lateral Positive
position TABLE II.
With rotation, ending lying flat Positive EXPERIMENT RESULTS
With recovery Negative Test Alert No alert
Category Accurate rate
Fall to the right Ending lying flat Positive times times times
With recovery Negative ADL 75 0 75 100%
Fall to the left Ending lying flat Positive Forward
90 90 0 100%
With recovery Negative fall
Backward
45 44 2 95.56%
ADL Sitting down on a chair then Negative fall
standing up Fall to the
30 29 1 96.67%
Lying down on the bed then rising Negative left
up Fall to the
30 30 0 100%
Walking 10 meters Negative right
Take the stairs Negative
Bending down, picking something Negative
up on the floor, then rising up
Coughing or sneezing Negative IV. DISCUSSION AND CONCLUSION
This study aims to develop a real-time portable wireless
fall detection system with high accuracy. In this study, A
Five healthy individuals, including four males and one portable fall-detection terminal attached on the belt is

2262
designed, it is small-size and more comfortable. High real [5] Doughty K., Lewis R., and McIntosh A.: The Design of a Practical and
time is achieved by the fall-detection algorithm and the Reliable Fall Detector for Community and Institutional Telecare. J.
structure of the system. All the data processing and the Telemed Telecare, 2000; 6: 150-154 (2000)
[6] Bourke A., O'Brien J., Lyons G.,: Evaluation of a threshold-based
fall-detection algorithm are operated in the microcontroller
tri-axial accelerometer fall detection algorithm. J. Gait & Posture, vol.
in the fall-detection terminal without transmitting data to an
26, issue 2, pp. 194-199 (2007)
upper computer. Therefore, the fall-detection terminal is [7] Bourke A., van de Ven P., Chaya A., OLaighin G., Nelson J.: Testing
capable for independent work, and detected falls timely. of a long-term fall detection system incorporated into a custom vest
The algorithm of fall detection is low-complexity and for the elderly. Conf Proc IEEE Eng Med Biol Soc. 2008;2008:2844-7
effective. The posture detection excluded most ADL events, (2008)
except lying down in horizontal. But the peak acceleration of [8] Kangas M., et al.: Comparison of low-complexity fall detection
lying could not pass the threshold of STOT, thus this algorithms for body attached accelerometers. J. Gait & Posture, vol.
28, issue 2, pp. 285-291 (2008)
algorithm can achieve a high specificity. Some falls were
[9] Wu G., et al.: Portable Preimpact Fall Detector With Inertial Sensors.
undetected, because of some young subjects broke the fall
IEEE Trans Neural Syst Rehabil Eng. 2008 Apr;16(2):178-83 (2008)
instinctively, thus the peak acceleration may lower than the [10] Nyan M. N. et al.: Application of motion analysis system in pre-impact
thresholds. fall detection. J. Journal of Biomechanics Vol. 41, Issue 10, pp.
Considering the design of the experiments, there were two 2297-2304 (2008)
main limitations. The first one was that the falls events were [11] Diaz A., Prado M., Roa L.M., Reina-Tosina J., Sanchez G.:
simulated by young subjects, which may cause trunk Preliminary evaluation of a full-time falling monitor for the elderly.
accelerations and the changes of trunk accelerations higher Proc IEEE EMBS 2004;3:2180-3 (2004)
than that of elderly in reality. The second one was that the [12] Bourke A., O’Brien J., Lyons G.: Evaluation of a threshold-based
tri-axial accelerometer fall detection algorithm. J. Gait & Posture, vol.
experiments site was simplex, indoor, and simulated falls
26, pp. 194̢9 (2007)
were performed on the mattress, which may cause lower [13] Yoshida T., Mizuno F., Hayasaka T., Tsubota K., Wada S.,
accelerations than that in reality. Thus, broader experiment Yamaguchi T.: A wearable computer system for a detection and
objects and different experiment sites would be adopted in prevention of elderly users from falling. Proc ICBM 2005;12:179̢82
further work. (2005)
In conclusion, this study was performed to introduce a [14] Mathie M.J., Basilakis and Celler B.G.: A system for monitoring
real-time portable wireless fall detection system with high posture and physical activity using accelerometers. In: 23rd Annual
detection accuracy. Through analysis of the kinetics International Conference IEEE Engineering in Medicine and Biology
Society, Istanbul 25–28 (2001).
characteristics of fall events, an effective method using a
[15] Zhang T, et al.: Fall detection by embedding an accelerometer in cell
tri-axes accelerometer was proposed for automatic fall
phone and using KFD algorithm. Int J Comp Sci Netw Secur 2006;6:
detection. An experiment was organized with young 277–84 (2006)
subjects included various kinds of ADL and falls events to [16] Bourke A., O’Donovan K., OLaighin G.: The identification of vertical
evaluate the effect of the fall detection system, and a velocity profiles using an inertial sensor to investigate pre-impact
sensitivity over 97% and a specificity of 100% were detection of falls. J. Medical Engineering & Physics, Vol. 30, Issue 7,
achieved. pp. 937-946 (2008)

ACKNOWLEDGMENT
This work was supported by the National High
Technology Research and Development Program of China
(863 Program) under Grant No.2008AA040207.

REFERENCES
[1] Shumway-Cook A., Ciol M.A., Hoffman J., et al.: Falls in the
Medicare population: incidence, associated factors, and impact on
health care. Phys Ther. 2009;89:324–332 (2009)
[2] Noury N., Rumeau P., Bourke A.K., OLaighin G., Lundy J.E.: A
proposal for the classification and evaluation of fall detectors.
(2008) IRBM, 29 (6), pp. 340-349 (2008)
[3] Gibson M.J., et al.: The prevention of falls in later life. A report of the
Kellogg International Work Group on the Prevention of Falls by the
Elderly. Danish Medical Bulletin 1987; 34(Suppl. 4): 1–24 (1987)
[4] Williams G., et al.: A Smart Fall and Activity Monitor for Telecare
Applications. In: 20th international Conference IEEE-EMBS 1998,
Hong Kong: 29 October–1 November 1998, p. 1151–1154 (1998)

2263