Increased Fall Detection Accuracy in an Accelerometer-based Algorithm

Considering Residual Movement

Panagiotis Kostopoulos1, Tiago Nunes1, Kevin Salvi1 , Michel Deriaz1 and Julien Torrent2
1 Institute of Services Science, University of Geneva, Geneva, Switzerland
2 Fondation Suisse pour les Téléthèses, Neuchâtel, Switzerland

Keywords: Fall Detection, Smartwatch, Sensors, Residual Movement, Accelerometer, Alarm.

Abstract: Every year over 11 million falls are registered. Falls play a critical role in the deterioration of the health of the
elderly and the subsequent need of care. This paper presents a fall detection system running on a smartwatch
(F2D). Data from the accelerometer is collected, passing through an adaptive threshold-based algorithm which
detects patterns corresponding to a fall. A decision module takes into account the residual movement of
the user, matching a detected fall pattern to an actual fall. Unlike traditional systems which require a base
station and an alarm central, F2D works completely independently. To the best of our knowledge, this is the
first fall detection system which works on a smartwatch, being less stigmatizing for the end user. The fall
detection algorithm has been tested by Fondation Suisse pour les Téléthèses (FST), the project partner for the
commercialization of our system. Taking advantage of their experience with the end users, we are confident
that F2D meets the demands of a reliable and easily extensible system. This paper highlights the innovative
algorithm which takes into account residual movement to increase the fall detection accuracy and summarizes
the architecture and the implementation of the fall detection system.

1 INTRODUCTION quences of falling, various fall detection systems

were developed over the last decade. These systems
Unintentional falls are frequent and quite dangerous are mainly based on video-cameras (Nait-Charif and
for elderly people. Yearly, more than 11 million falls McKenna, 2004; Lee and Lee, 2009; Huang et al.,
are registered (Brown, 2005), leading to a wide spec- 2008), acoustic (Alwan et al., 2006; Litvak et al.,
trum of injuries for this age group. Aside from caus- 2008) or inertial sensors (Hwang et al., 2004) and
ing physical injuries, falls can also have dramatic psy- mobile phone technology (Sposaro and Tyson, 2009;
chological consequences that reduce elderly people’s Dai et al., 2010; Dumitrache and Pasca, 2013; Aguiar
independence (Ryynanen et al., 1992). It has been et al., 2014).
found that after falling, 48% of older people report Common fall detection systems are based on
a fear of falling and 25% report curtailing activities. a sensor detecting a strong vertical acceleration,
Moreover, falls can also lead to disability and de- launching an alarm when a fall event is recognized.
creased mobility which often results in increased de- More recent systems usually take into account other
pendency on others and, hence, an increased need of sensors able to detect the device’s orientation in or-
being admitted to an institution. Finally, one other se- der to determine whether the user is lying or standing.
rious consequence of falling is the ”long-lie” condi- In (Dai et al., 2010) the authors present PerFallD, a
tion where a falling person remains on the ground or system which combines the detection and the com-
floor for more than an hour after a fall. The ”long-lie” munication components using mobile phones. They
is a sign of weakness, illness and social isolation and compare it with existing academic and commercial
is associated with high mortality rates among the el- solutions and conclude that their system is quite ac-
derly. Time spent on the floor can be associated with curate despite the fact that they are using only low-
a fear of falling, muscle damage, pneumonia, pres- cost sensors (the ones from an Android G1 phone). In
sure sores, dehydration and hypothermia (Lord et al., (Sposaro and Tyson, 2009) the authors present iFall,
2001; Nevit et al., 1989). another Android application tested on a G1 phone,
In an attempt to minimize these serious conse- which is focused on the communication between the

30 Kostopoulos P., Nunes T., Salvi K., Deriaz M. and Torrent J..
Increased Fall Detection Accuracy in an Accelerometer-based Algorithm Considering Residual Movement.
DOI: 10.5220/0005179100300036
In Proceedings of the International Conference on Pattern Recognition Applications and Methods (ICPRAM-2015), pages 30-36
ISBN: 978-989-758-077-2
Copyright
c 2015 SCITEPRESS (Science and Technology Publications, Lda.)
 IncreasedFallDetectionAccuracyinanAccelerometer-basedAlgorithmConsideringResidualMovement

fall detection system and the alarm notification mod- 2 SYSTEM DESIGN
ule, instead of the sensors only.
Most of the current fall detection systems require Our fall detection system is an Android application
a base central. In this paper we propose a fall de- running on an AW-420.RX smartwatch of Simvalley
tection system (F2D) which works on a smartwatch, Mobile. We have chosen the Android based solution
therefore completely independent from a base station. because it is an open source framework designed for
Using such a device is less stigmatizing for the user. mobile devices. The Android SDK provides the API
In addition, it can be offered for less than half of the libraries and developer tools necessary to build, test
cost of existing systems on the market. Our system and debug applications for Android. We implemented
meets the requirements of reliability, ease of installa- the prototype in Java using the Android SDK API 19.
tion and restriction of false positives (Doughty et al., The fall detection algorithm, which is explained be-
2000) which are essential for a properly built fall de- low, is implemented as a background service. When a
tection system. fall is detected the service informs the main applica-
F2D works on a smartwatch and therefore fixed tion, which notifies the caretakers (family or friends).
on the wrist of the person. We have avoided the dis- There are several notification channels the user can
advantages of (He et al., 2012) where the solution of choose from: call, SMS or email.
the waist-mounted smartphone the authors provide is
not feasible for two reasons: 1) Normally people do 2.1 System Overview
not wear their phones on the waist but in their pockets.
2) The system will be working only when the smart- We collect data from the accelerometer sensor of the
phone is mounted on the waist and not at other times. smartwatch. Then we apply filters to detect patterns
Other problems such as the usage of intrusive devices corresponding to a fall. We use a threshold based
exist in (Hou et al., 2012) and (Li et al., 2012), where algorithm for the fall detection which takes into ac-
the accelerometer and Bluetooth unit are bounded as count the residual movement of the user after the fall.
a wearable unit and placed on the subject’s waist or The thresholds were selected based on experiments
chest. with different profiles of users (age, weight, height
Nowadays, simple smartwatches are very power- are factors that were taken into account). The deci-
ful and have a set of sensors that can be used and sion module combines different data coming from the
diverted from their original intent. More computing filters as depicted in Figure 1, in order to make a deci-
power and storage on these devices offer greater op- sion whether a possible fall corresponds to a real fall.
portunities. Using a single smartwatch as a device for If it is the case, the information is transmitted to the
running the F2D application satisfies the condition of alarm module. When a potentially critical situation is
ease of installation of the fall detection system. detected, the smartwatch uses different communica-
In general the fall detection algorithms wait for tion means (call, SMS, email) to inform the caretak-
a response of the user after a possible fall in order ers. Another important difference with the traditional
to make a decision if the situation is critical or not. systems is that the smartwatch communicates directly
But this is not always applicable and convenient for with the caretakers with no involvement of a base sta-
the end user, especially for elderly people who are the tion and a centralized alarm.
target group of our research. In F2D we propose an
accelerometer-based algorithm considering the resid-
ual movement after the fall. This analysis is per-
formed in the decision module of the F2D application
which is responsible for the classification of a possi-
ble fall pattern being a real fall event or not. Figure 1: System architecture.
The rest of this paper is organized as follows. In
Section 2 our designed fall detection system is de-
scribed in detail emphasizing the innovative fall de- 2.2 Fall Detection Algorithm
tection algorithm. Experimental results are reported
and discussed in Section 3. Future improvements of The fall detection algorithm is implemented in a back-
our work to make F2D more robust are presented in ground service and is running continuously. The user
Section 4. Finally, a brief conclusion is given in Sec- can operate his smartwatch as usual. F2D does not
tion 5. cause any interference with the normal usage of in-
stalled applications. The algorithm is threshold based
like (Dumitrache and Pasca, 2013), relying on the
captured data of the accelerometer of the smartwatch.

31
ICPRAM2015-InternationalConferenceonPatternRecognitionApplicationsandMethods

We decided to use a threshold based algorithm and of data from simulated falls mentioned above. The
not a machine learning approach like (Aguiar et al., main goal of the algorithm is the detection of all falls
2014) as it is less complex and therefore requires the and at the same time the elimination of false positives.
lowest computational power (Habib et al., 2014). In Building and testing our system we concluded that
the typical scenario, the user will use the application less than 6 seconds is not enough for the detection of
on his smartwatch normally during the day without all different types of falls. However, setting the win-
the requirement of charging it much more than usu- dow to higher values creates a bigger occurrence of
ally. Since the fall detection system will run continu- false positives.
ously, we should optimize the battery consumption of
the device. Therefore, only the tri-axial accelerometer 2.2.2 Fall Pattern
signal is used since it is the most informative sensor
regarding the fall detection. The next step of the algorithm is the detection of a
The algorithm distinguishes daily activities from possible fall. In order to consider an activity as a pos-
falls. Activities of Daily Living (ADL) are normal sible fall the two following conditions must be satis-
activities such as walking, standing or running. The fied: 1) If the acceleration exceeds an upper thresh-
pattern of a fall must be different from the patterns old which is set to 18 m/sec2 . 2) If after a given
of these activities. Acceleration data is sampled at 40 time interval the acceleration exceeds a lower thresh-
Hz from the 3-axis accelerometer sensor embedded old which is set to 7 m/sec2 . The two thresholds have
in the Android smartwatch. Specifically, the sensor been selected based on the basic trade-off between de-
which provides acceleration information without the tecting all falls and avoiding false positives. This time
gravity component (linear acceleration) is used. We difference between the two peaks represents the resid-
calculate the norm of the acceleration for each mo- ual movement of the user after a fall. The intensity of
ment as described in Equation 1. this movement depends on the profile of the individ-
p ual user. The three possible values of the time interval
acceleration = x2 + y2 + z2 (1) are given in Table 1.
If the two conditions are satisfied during the time
We have analyzed a set of data with 150 different
window of 6 seconds then a possible fall is suspected.
simulated falls from different people involved in the
We can see in Figures 2 - 4 that this time window
experiments, as reported in Table 2, from our project
is sufficient for the satisfaction of the two conditions
partner FST. This company has a long experience
that should happen in order to detect a fall pattern.
in creating and using innovative products adapted to
people with disabilities. Thanks to this data we have 2.2.3 Decision Module
improved the detection of possible falls. We observed
that all falls follow one of the three patterns which we
have called smooth, strong and sharp. They are given The final step of the fall detection algorithm is the
in Figures 2, 3 and 4 respectively. The main differ- classification of the fall pattern as real. In this deci-
ence between them is the time interval of the residual sion module a counter increases every time that both
movement after the fall Table 1. More specifically, conditions of exceeding the thresholds are satisfied.
when a fall takes place, the peak that exceeds the up- The critical range of the values of the fall counter
per threshold of the acceleration corresponds to the is (1 ≤ counter < 14). If (counter ≥ 14), then it is
hit. After this, the pattern of the fall has a second due to another activity being performed (e.g., run-
peak, lower than the first one and exceeding the lower ning) which gives the difference in the acceleration
threshold. Finally, the acceleration returns to normal values as we can see in Figure 6. On the other hand if
values. This behaviour of the acceleration after the (counter < 1) it means the user at most did a sudden
first peak represents the residual movement that we movement with his wrist and so the threshold condi-
take into account in the decision module for the char- tions were not satisfied (e.g., when a user was going
acterization of a possible fall event as a real fall. down the stairs in Figure 8). The graphical explana-
tion and the structure of the fall detection algorithm is
2.2.1 Time Window given in Figure 5.

The time window is an essential part of the fall de- 2.3 Emergency Actions
tection algorithm. We have defined a time window in
which we are able to recognize a fall pattern. This If the algorithm decides that a fall has happened then
window is set to 6 seconds, a value which has been the background service notifies the main application,
selected after conducting experiments, using the set which in turn sends a message to the caretakers.

32
 IncreasedFallDetectionAccuracyinanAccelerometer-basedAlgorithmConsideringResidualMovement

The smartwatch communicates directly with the

caretakers. In case of an alarm the loudspeaker of
the watch is automatically turned on at a high volume
and calls from caretakers are automatically answered.
This allows the user to communicate even in uncom-
fortable positions that could result after a fall.
Table 1: The three possible time intervals.
Smooth Strong Sharp
100 ms 300 ms 500 ms

Figure 4: Sharp fall.

Figure 2: Smooth fall.

Figure 5: Fall detection algorithm.

Figure 3: Strong fall.

3 EVALUATION
For the evaluation of the reliability of the fall detec-
tion mechanism in F2D we performed a series of ex-
periments. We collected different types of falls’ data
(i.e., forwards, backwards, sideways). We also col-
lected activities of daily living data (e.g., walking, go-
ing down the stairs, going up the stairs) as we can Figure 6: Running activity.
see in Figures 7 - 9. Since it is very difficult to test
a fall detection mechanism with elderly people, as it periments with healthy adults. The test subjects were
can more easily cause injuries, we performed the ex- wearing smartwatches while falling on a mattress.

33
ICPRAM2015-InternationalConferenceonPatternRecognitionApplicationsandMethods

partner FST. This set consists of a subset of 384 sim-

ulated falls and 417 ADL. The set of data that we
are using is much larger comparing with other sys-
tems (Dumitrache and Pasca, 2013) where only 34
simulated falls and 200 daily activities simulated by
a single young person were used. Our data has been
collected from 6 people with different profiles as re-
ported in Table 2. The simulated falls took place in a
room falling on a mattress in order to avoid injuries.
Also the ADL were the following: walking, going up
the stairs, going down the stairs, stand up from a chair,
sit down on a chair, running. Based on these facts,
it can be noted that the accuracy of our algorithm is
quite high. We achieved a true positive rate (sensi-
Figure 7: Walking activity.
tivity) of 92.18% for the set of simulated falls and a
true negative rate (specificity) of 87.29% for detecting
the ADL data. The average of sensitivity and speci-
ficity represents the accuracy of the system which is
equal to 89.74%. The analytical results are presented
in Figures 10 and 11.
Table 2: Different profiles.
Age Height (cm) Weight (kg)
22 185 76
26 176 69
27 182 63
29 184 53
30 186 93
Figure 8: Going down the stairs. 40 177 75

Figure 9: Going up the stairs. Figure 10: Accuracy using real data from partner.

For testing the ADL the test subjects wore the We conclude that the F2D system works reliably.
smartwatches for 24 hours. Testers were doing all Some false alarms were detected when the testers per-
their daily life activities. As we can see in the Fig- formed sudden movements with residual activity try-
ures 7 - 9 the acceleration values do not exceed the ing to simulate the same pattern of a fall event.
upper threshold. Hence, a suspected fall is not classi-
fied as a real fall because the counter must exceed 0
in order to detect a real fall. 4 FUTURE WORK
Moreover, we evaluated our fall detection algo-
rithm using a set of simulated falls and another set The first step for further improvement of the robust-
with ADL data, that we received from our project ness of our fall detection system is the reduction of

34
 IncreasedFallDetectionAccuracyinanAccelerometer-basedAlgorithmConsideringResidualMovement

into account the users’ height and weight as parame-

ters for the algorithm. The second manner is to help
the decision module. Once a possible fall is detected
by the algorithm, we will check if it is likely that it is
a real fall according to the current context.
Finally, thanks to the experience of our project
partner with elderly end users, we will be able to test
our fall detection system with real-world fall data in
the future. Testing our algorithm in real-life condi-
tions will give us the opportunity to create a more ro-
bust automated alarm system with higher acceptance.

Figure 11: Decomposition of ADL and respective detection 5 CONCLUSION

rate.
Fall detection is a research field that has a big impact
false positives. We are planning to achieve this goal on the improvement of the daily life of elderly people.
in three steps. The first step is the elimination of mis- In this paper we propose the first fall detection system
leading direction of the movement of the smartwatch. designed to run on an independent smartwatch (F2D).
We cannot expect to detect a fall when the direction There is no base station (which limits the range), no
of smartwatch is going up. Thanks to this recognition central alarm station (which is more difficult to man-
of the movement direction we will avoid some false age) and it works on a standard smartwatch. It implies
positives coming from sudden movements which have that it is less stigmatizing for the end user, quite cheap
similar patterns with a fall but in which the direction comparing to existing systems and it is easily extend-
of the movement is going up. able. F2D uses an innovative fall detection algorithm
The second step for reducing false positives is the which takes into account the residual movement of the
use of adaptive thresholds. With the selected thresh- user in order to match a fall pattern to a real fall.
olds we cover most of the cases of a fall based on the We have conducted a range of experiments in our
experiments carried out in our lab and the data that lab and used real data from our project partner FST
we received from the end users of our project partner. who has relevant experience in the domain involving
With the adaptivity of the upper and lower thresholds real end users. These experiments demonstrated that
we will make our fall detection system more generic the fall detection system is robust.
and therefore allow a tailored response for each spe- Based on the reliability of the fall detection and
cific end user. We expect further improvements to the the restriction of false positives, which are guaran-
fall detection as well as a reduction of the false posi- teed by the fall detection algorithm, we have built a
tives. system which meets the requirements for deployment
The third and last step in decreasing false positives and use.
is taking into account the users’ context. Common
fall detection systems are based on a sensor detect-
ing a strong vertical acceleration. But smoother falls ACKNOWLEDGEMENTS
(e.g., a person grabbing a chair, a table, or any other
object) are not detected while some particular situa- This work was supported by the Swiss Commission
tions (a user that sits abruptly on his sofa) lead to false for Technology and Innovation (CTI grant 15876.2
alarms. To distinguish only the critical situations, we PFES-ES). The authors would like to thank Anastasija
will take into account contextual data. For example, Collen who helped to improve the quality of this pa-
what happened before the fall, after the fall, at which per.
place in the room, at what time of the day, or next to
what object in a room. A user’s profile, continuously
updated with his habits or particular behaviour, will
bring additional data to the user’s context. The cur-
REFERENCES
rent context (e.g., position, time) will be used in two Aguiar, B., Rocha, T., Silva, J., and Sousa, I.
different manners. The first is to define the variables (2014). Accelerometer-based fall detection for smart-
of the filters in order to better differentiate the pos- phones. In Medical Measurements and Applications
sible (expected) patterns. For example, we can take (MeMeA), 2014 IEEE International Symposium on.

35
ICPRAM2015-InternationalConferenceonPatternRecognitionApplicationsandMethods

Alwan, M., Rajendran, P., Kell, S., Mack, D., and Dalal, S. Ryynanen, O. P., Kivela, S. L., Honkanen, R., and Laippala,
(2006). A smart and passive floor-vibration based fall P. (1992). Falls and lying helpless in the elderly. In Z
detector for elderly. In Proc. ICTTA’ 06. Gerontol.
Brown, G. (2005). An accelerometer based fall detec- Sposaro, F. and Tyson, G. (2009). ifall: An android applica-
tor: Development, experimentation, and analysis. In tion for fall monitoring and response. In Engineering
EECS/SUPERB. in Medicine and Biology Society, 2009. EMBC 2009.
Dai, J., Bai, X., Yang, Z., Shen, Z., and Xuan, D. (2010). Annual International Conference of the IEEE.
Perfalld: A pervasive fall detection system using mo-
bile phones. In Pervasive Computing and Communi-
cations Workshops (PERCOM Workshops), 2010 8th
IEEE International Conference on.
Doughty, K., Lewis, R., and McIntosh, A. (2000). The
desigh of a practical and reliable fall detector for
community and institutional telecare. In Journal of
Telemedicine and Telecare.
Dumitrache, M. and Pasca, S. (2013). Fall detection al-
gorithm based on triaxial accelerometer data. In E-
Health and Bioengineering Conference (EHB), 2013.
Habib, M. A., Mohktar, M. S., Kamaruzzaman, S. B.,
Lim, K. S., Pin, T. M., and Ibrahim, F. (2014).
Smartphone-based solutions for fall detection and
prevention: challenges and open issues. In
http://www.mdpi.com/journal/sensors.
He, Y., Li, Y., and Bao, S.-D. (2012). Fall detection by
built-in tri-accelerometer of smartphone. In Biomedi-
cal and Health Informatics (BHI), 2012 IEEE-EMBS
International Conference on.
Hou, Y., Li, N., and Huang, Z. (2012). Triaxial
accelerometer-based real time fall event detection. In
Information Society (i-Society), 2012 International
Conference on.
Huang, B., Tian, G., and Li, X. (2008). A method for fast
fall detection. In Proc 7th World Congr Intell Contr
Autom.
Hwang, J., Kang, J., Jang, Y., and Kim, H. (2004). Devel-
opment of novel algorithm and real-time monitoring
ambulatory system using bluetooth module for fall de-
tection in the elderly. In Proc 26th IEEE EMBS Ann
Int Conf.
Lee, Y. and Lee, H. (2009). Multiple object tracking for
fall detection in real-time surveillance system. In Proc
11th Int Conf Adv Commn Tec.
Li, Y., Chen, G., Shen, Y., Zhu, Y., and Cheng, Z. (2012).
Accelerometer-based fall detection sensor system for
the elderly. In Cloud Computing and Intelligent Sys-
tems (CCIS), 2012 IEEE 2nd International Confer-
ence on.
Litvak, D., Zigel, Y., and Gannot, I. (2008). Fall detection
of elderly through floor vibrations and sound. In Proc
30th IEEE EMBS Ann Int Conf.
Lord, R., Sherrington, C., and Menz, H. (2001). Falls in
older people: risk factors and strategies for prevention.
In Cambridge University Press. Cambridge.
Nait-Charif, H. and McKenna, S. (2004). Activity summari-
sation and fall detection in a supportive home environ-
ment. In Proc 17th Inte Conf Patt Recog. Cambridge
UK.
Nevit, M., Cummings, S., Kidd, S., and Black, D. (1989).
Risk factors for recurrent nonsyncopal fall. a prospec-
tive study. In J Am Med Ass.

36