IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 15, NO.

1, JANUARY 2019 35

FallDroid: An Automated Smart-Phone-Based

Fall Detection System Using Multiple
Kernel Learning
Ahsan Shahzad , Student Member, IEEE, and Kiseon Kim , Senior Member, IEEE

Abstract—Common fall occurrences in the elderly pop- over 70 years of age [1]. These falls are responsible for majorly
ulation pose dramatic challenges in public healthcare 90% of hip and wrist fractures and 60% of head injuries [2].
domain. Adoption of an efficient and yet highly reliable
Besides these injuries, long-lie situation (i.e., remaining on the
automatic fall detection system may not only mitigate
the adverse effects of falls through immediate medical as- ground for long time) is another outcome of the fall that has se-
sistance, but also profoundly improve the functional ability rious consequences such as dehydration, hypothermia, and even
and confidence level of elder people. This paper presents death. Moreover, frequent incidence of falls in the elderly may
a pervasive fall detection system developed on smart provoke fear of falling, which in turn, deteriorates their con-
phones, namely, FallDroid that exploits a two-step algorithm fidence in living independently and being socially active [3].
proposed to monitor and detect fall events using the embed-
ded accelerometer signals. Comprising of the threshold- Thus, the development of automated, reliable, and prompt fall
based method and multiple kernel learning support vector detection systems is vital to guarantee immediate assistance in
machine, the proposed algorithm uses novel techniques to case of falls, especially those involving long lies, and minimize
effectively identify fall-like events (such as lying on a bed severe health complications [4].
or sudden stop after running) and reduce false alarms. In Contemporary techniques employed for automatic detection
addition to user convenience and low power consumption,
experimental results reveal that the system detects falls of imminent real-life falls can be broadly classified into two
with high accuracy (97.8% and 91.7%), sensitivity (99.5% categories: 1) context-aware systems and 2) wearable systems
and 95.8%), and specificity (95.2% and 88.0%) when placed [5]. The former category concerns the deployment of sensory
around the waist and thigh, respectively. The system also gadgets such as cameras, microphones, infrared, and pressure
achieves the lowest false alarm rate of 1 alarm per 59 h of sensors to track the movement of people in limited environ-
usage, which is best till date.
ments. The main strength of these systems lies in usability
Index Terms—Accelerometer, Android app, fall detection, among the elderly as no dedicated device is needed to be worn.
multiple kernel learning (MKL), power consumption, smart Nonetheless, such systems are vulnerable to issues such as lim-
phone (SP)-based application.
ited coverage, high installation cost, high false alarms due to
I. INTRODUCTION other mobile entities, and privacy (especially in video-based
systems). Fall detection methods based on wearable motion
XPONENTIAL increase in falls are recognized as a major
E factor causing physical, psychological, and economical
concerns within the growing elderly population worldwide. Sta-
sensors that rely on kinematic signals, like triaxial accelerom-
eters and gyroscopes, fall under the latter category. While such
body-worn systems offer several advantages over video-based
tistical reports from the World Health Organization (WHO) in- systems, the bearer is still required to carry at least one device
dicate that 28%–35% of seniors over 64 are subject to a fall that may be intrusive and raise usability concerns. Moreover, the
event each year, which further elevates to 32%–42% for those cost incurred by wearable commercial/customized fall detection
devices such as LifeCall [6] and LPFD [7] is another issue of
Manuscript received January 11, 2017; revised January 8, 2018 and fundamental importance.
March 20, 2018; accepted May 10, 2018. Date of publication May 23,
2018; date of current version January 3, 2019. This work was supported Momentum in the advancement of microelectromechanical
in part by the GIST Research Institute, and in part by the Brain Research systems (MEMS) technology has, however, enabled the devel-
Program through the National Research Foundation of Korea funded by opment of very light, compact, low power, and inexpensive
the Ministry of Science and ICT 2016M3C7A1905477. Paper no. TII-17-
0035. (Corresponding author: Kiseon Kim.) wireless inertial sensors that eliminate concerns regarding their
The authors are with the School of Electrical Engineering and portability and user inconvenience [8]. Consequently, smart
Computer Science, Gwangju Institute of Science and Technology, phones (SPs) integrated with dedicated detection hardware serve
Gwangju 500-712, South Korea (e-mail: [email protected];
[email protected]). as potential candidates that have widely been accepted as a daily
This paper has supplementary downloadable material available at life commodity [9]. The remarkable penetration of SPs as a mo-
http://ieeexplore.ieee.org. bility and safety tool among the elderly has, therefore, led the
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. surge for cost-effective, efficient, and commercially viable fall
Digital Object Identifier 10.1109/TII.2018.2839749 detection applications [10].

1551-3203 © 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
36 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 15, NO. 1, JANUARY 2019

Standalone body-worn or SP-based fall detectors have been the placement of the SP is provided. Furthermore, more battery
extensively surveyed in [5] and [11]–[13]. Earlier studies were power is drained due to the continuous use of the ANN.
predominantly grounded on threshold-based methods (TBM) Successful deployment of a fall detection system among
that rely on certain decision thresholds applied on features ex- elderly population depends on various factors: usability, bat-
tracted from the inertial sensor signals [14], [15]. These detec- tery lifetime, privacy issues, cost, and reliability. The System
tion algorithms are easy to implement, offer less computation usability hugely depends on the number of sensors used, their
cost, and are power efficient. The authors of [12] compared four prefix orientation, locations, battery life, and the like. The sys-
famous TBM algorithms with a commercial device and showed tem should be reliable, i.e., it should detect almost all the fall
that the TBM techniques fail to avoid false negatives (falls that events with 100% sensitivity while keeping the false alarm rate
remain undetected) and false positives (activities of daily living at its minimum. High false alarm rates can be very annoying and
(ADL) classified as falls) simultaneously. Since fine tuning the ultimately reduce user compliance with the system. To decrease
thresholds invokes a tradeoff between number of false negatives false alarms, some researchers utilize postural information by
and false positives, it is difficult to attain optimal thresholds either employing multiple sensors or by using a single triaxial
ensuring performance consistency for everybody. sensor with prefix orientation. As older people tend to easily
More recent studies reported in the literature use sophisti- forget, neither of the two options serve convenient in terms of
cated machine learning (ML) techniques such as support vector usability. As the existing literature relating to these issues is
machines (SVMs) [16], artificial neural networks (ANNs) [9], lacking, there is no widely accepted system among elderly until
[17], and k-nearest neighbor (k-NN) [18] to classify falls from now. Hence, it is of great significance to develop a reliable fall
ADL. A comprehensive MATLAB-based comparison of the of- detection system to fill this gap while accounting for the existing
fline classification performance of six ML approaches is given practical issues.
in [19], wherein 1404 features extracted from accelerometer, Accordingly, we present an automated high performance SP-
gyroscope, and magnetometer sensing were used by the au- based fall detection system focusing on practical issues such as
thors to obtain an average accuracy of around 99%. In spite of user convenience and power consumption. The proposed stan-
the profound results, usage of multiple sensors, especially gyro- dalone fall detector is developed as an Android app, namely
scope, along with the high computational cost of ML techniques FallDroid, which uses the accelerometer sensor embedded in
exploiting large feature sets demands high power consumption, SPs. The designed application provides an elder-friendly GUI
thus, making them unfeasible for SPs. and supports the two most convenient SP carrying locations:
With over 86% of the market share in the third quarter of waist (belt/pouch) and thigh (pant pocket). In comparison with
2016 [20] and due to its open source approach, Android OS ML techniques, the proposed two-step algorithm is shown to be
stands out as the most widely used programming environment more power efficient. In the first step, a low computational cost
for SP-based fall detection solutions [21]. Though many SP- approach based on the TBM is used, followed by the pattern
based fall detection algorithms have been reported, yet only recognition technique, multiple kernel learning support vector
few have actually been tested in real life. Moreover, to our best machine (MKL-SVM) in the second step which is rarely in-
knowledge, smart fall detector (SFD) [17] and iFall [14] are voked. The battery consumption was analyzed and reported for
the only two applications that have been released for public different scenarios. The recorded datasets were acquired from
use. Other fall detection Android apps available on Google Play human trials conducted systematically in both, laboratory and
Store include Fade: Fall detector, Emergency Fall detector, and free living environments. Finally, we report the offline and on-
T3Lab [22]. Nonetheless, information on neither the underlying line classification results on fall-like ADLs such as lying on the
algorithm nor their performance exists. floor, sudden stop after walking, accidentally hitting the sen-
Abbate et al. proposed an SP-based fall detector that uses a sor, etc., to demonstrate the better performance of the presented
combination of TBM and ANN [9]. In spite of the reported 100% system.
classification performance in offline analysis, the dataset used The remainder of this paper is structured as follows:
for training and testing the ANN was very small (86 samples Section II presents the design, implementation, and algorithm
in total). Also, the performance of the application (false alarms used in the proposed system. The experimental settings and cri-
rate and battery consumption) in real-life scenarios, commonly teria for various scenarios are detailed in Section III, followed
known as online analysis, was not presented. by results and discussions in Sections IV and V, respectively. Fi-
More recently, Kerdegari et al. developed an Android nally, Section VI concludes this paper by pointing out the impor-
application, SFD, built on the notion of the multilayer tant observations and guidelines for potential future directions.
perceptron (MLP) neural network for fall event detection [17].
During offline analysis, their algorithm achieves 92.03% sensi-
tivity, 91.07% specificity, and 91.06% accuracy on data recorded II. FALLDROID SYSTEM DESIGN
around the waist. However, when applied for online analysis, FallDroid is designed as a standalone and user-independent
the system performance slightly degrades in terms of specificity fall detection system that actively runs in the background and
(93.18% sensitivity, 88.88% specificity, and 91.25% accuracy). uses a two-step algorithm (described in Section II-B) to analyze
Additional to the long system decision time (at least 30 s for subject movement. Upon fall detection, the application triggers
algorithm decision + 60 s default time for alert cancellation), the SP to vibrate and an alert cancellation page appears on the
no specific information on the false alarm rate with respect to screen. Unless canceled within a specified time period (default
SHAHZAD AND KIM: FALLDROID: AN AUTOMATED SMART-PHONE-BASED FALL DETECTION SYSTEM USING MULTIPLE KERNEL LEARNING 37

Sensor Event Listener to receive motion data from the

accelerometer at a desired sampling rate. The incoming sen-
sor data values are stored in a linked-list queue following the
first-in-first-out (FIFO) discipline, and maintains data history
of up to 6 s. As further explained in Section II-B, initial steps
of the TBM algorithm use this data to detect fall like events.
Upon detection, the rest of the algorithm is then executed. The
parameters associated with the pretrained MKL-SVM are not
hard coded but instead provided via a file. In this manner, appli-
cation personalization or update is facilitated by replacing the
parameters file with a new one.
3) Notification System: When a fall event is detected by the
Fig. 1. Overview of the fall detection and emergency alert system. MKL-SVM, the SP starts vibrating followed by an automatic
launch of the fall alert cancellation page. The user has a default
setting of 30 s) by the user, a sound alarm is activated followed period of 30 s to cancel the false alert in case an ADL event
by a help text message containing location information being is misclassified as fall. If the user cancels the alert, the appli-
sent to specified emergency contacts. This process is illustrated cation then requests for feedback through the feedback page.
in Fig. 1. The entered data are stored in a file that can later be used for
further analysis to improve the algorithm performance. How-
ever, in case of a real fall event, the sound alarm is triggered
A. FallDroid Implementation once the alert countdown timer expires. The system then re-
FallDroid has been developed using Android Studio IDE with trieves the last known geographical position of the user based on
min API 17 and target API level 23. The GUI of application con- the available location providers (like GPS, Network) and sends
sists of four screens: the main page, settings, fall alert cancella- the most accurate result in the form of an alert text message to
tion, and feedback. The layout is designed to facilitate usability the prespecified emergency contact number. If the last known
by the elderly with an overall focus on reducing the battery location was not too accurate (>100 m) or not recent (>10 min),
power consumed particularly for unnecessary computations. In the application then tries to acquire the latest/current location
what follows, we highlight the main services provided by the prior to sending the alert message.
application.
1) Configuration and Control: The main page of the appli-
B. Fall Detection Algorithm
cation serves three functions: 1) start/stop the fall detection pro-
cess; 2) application configuration settings; and 3) summarized As shown in Fig. 2, FallDroid uses a two-step fall detection
display of critical/crucial settings. Once the user starts the sens- approach composed of TBM and MKL-SVM. The first step of
ing process, a notification icon will appear on the top left corner the algorithm relies on the TBM for determining fall-like events
of the screen in the notification bar. The icon remains visible as and efficiently discarding most fall-like ADLs using inactivity
long as the application is running in the background. The set- and posture tests. As falling down is a single incident that occurs
tings page can be used to customize personal details, location of instantaneously, a fall-like event does not exhibit the traits of
the SP being carried, fall detection service priority, and settings repetitiveness and is basically characterized as an acceleration
related to fall alert notification. Settings for fall alert notifica- peak of magnitude greater than the 3-g threshold, followed by a
tion comprise of alarm sound and duration, countdown timer for period of 3.5 s without further peaks exceeding the threshold [1]
alert cancellation, and entries for emergency contacts. Changes (see Fig. 3). The threshold value 3-g is selected as it is widely
made to any user-specified setting can be seen immediately un- reported in the literature [23] and is small enough to avoid
der the settings summary section on the main page. The settings false negatives, as even low impact falls have peak acceleration
are saved using the Shared Preferences class, which per- greater than 3-g value. When the user is static (i.e., no SP or
mits previous settings to persist over multiple sessions as well sensor movement), the acceleration vector magnitude (AVM)
as after SP reboot. shows a value of 1-g representing gravitational acceleration.
2) Fall Detection Service: The algorithm proposed for fall However, during free fall, this value falls below 1-g and upon
detection is implemented as an Android service, using the having impact with the ground, a high value (>3-g) appears.
Intent Service class, that can run continuously in the The AVM signal used to detect fall-like events is
background irrespective of the application. The intent service 
runs the algorithm in its own separate worker thread with- AVM[i] = (Ax [i])2 + (Ay [i])2 + (Az [i])2 (1)
out blocking the main UI thread that otherwise can make the
application nonresponsive. Once the service is activated, it where Ax , Ay , and Az represent, respectively, the acceleration
instantly acquires the PARTIAL_WAKE_LOCK, which pre- signals along x-, y-, and z-axes of the sensor. When a fall-like
vents the CPU from going into sleep mode when the phone is event is detected, the user is sensed for any activity over a de-
idle. Subsequently, the algorithm checks the device accelerom- fault period of 5 s. The average absolute acceleration magnitude
eter and its sensing capabilities. Once found, it registers the variation (AAMV) measure defined as follows is then compared
38 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 15, NO. 1, JANUARY 2019

Fig. 4. Examples of thigh acceleration patterns during ADL that trig-

gered false alarms on the proposed (TBM + MKL-SVM) algorithm.

Fig. 5. ROC curves of AAMV (inactivity test) and TAVy (posture test)
Fig. 2. Fall detection algorithm (proposed), comprising of two main features using thigh location dataset.
steps: TBM and pattern recognition method, with an optional recovery
detection part.
pattern recognition technique, explained later on in this sub-
section. Some examples of acceleration patterns generated by
thigh-located SP that cause false alarms are shown in Fig. 4. An
optional long-lie or recovery detection step can also be applied
at the end of the algorithm to further alleviate the false alarms.
If the user shows significant movement after the fall event and
regains an upright posture within one minute, then it will be
treated as recovery from fall and no fall alert will be generated.
The extended test serves more efficient in discarding fall-like
ADL events or to detect recovery from (noninjurious) fall as
the user is more likely to show movement. On the other hand
Fig. 3. (a) Fall-like ADL event generated by running and sudden stop however, such prolonged tests would delay decision making and
activity. (b) Forward simulated fall pattern with a 3-s window centered at medical assistance in case of an actual fall.
impact point.
The thresholds are obtained from heuristic analysis using
the training dataset samples. Furthermore, the thresholds are
with predefined threshold to analyse the inactivity event: chosen such that no actual fall event is mistakenly classi-
1    fied and discarded as ADL in the TBM stage. This ensures
AAMV = AVM[i + 1] − AVM[i]. (2)
# of samples that the algorithm can achieve 100% sensitivity depending
i∈win
on the performance of the MKL-SVM. In other words, by
In case of little or no movement after impact, the aforemen- choosing higher threshold values for the AAMV and lower
tioned equation yields values close to zero (AAMV < 0.6). If values for the TAV, more number of fall-like ADL events
the user was inactive, then his/her posture variation is evalu- are likely to be passed to the computationally expensive
ated using pre- and postimpact (peak) signals. If the user shows MKL-SVM. However, owing to the fact that the number of
postural change, as tested by tilt angle variation (TAV > 10◦ ), fall-like events occurring daily are usually very few, the impact
then it is most likely a fall. Hence, decisions on such fall-like on power consumption will be insignificant. Fig. 5 shows the
events that are not discarded by the TBM are made in the sec- receiver operating characteristics (ROC) curves of the AAMV
ond step of the algorithm that exploits the powerful MKL-SVM and TAV measures. It can be seen that each of them can achieve
SHAHZAD AND KIM: FALLDROID: AN AUTOMATED SMART-PHONE-BASED FALL DETECTION SYSTEM USING MULTIPLE KERNEL LEARNING 39

TABLE I defined as the total absolute acceleration along each axis.

SUMMARY OF THE 15 CANDIDATE FEATURES DERIVED FROM THE   
ACCELEROMETER SIGNALS SAAx/y /z = Ax/y /z [i]. (4)
i∈win

d) Summed magnitude area (SMA): The sum of the abso-

lute values of the x-, y-, and z-axes accelerations over
the window length is represented by this feature and is
computed as
      
SMA = Ax [i] + Ay [i] + Az [i] . (5)
i∈win

2) Proposed MKL-SVM: To recognize the pattern of falls

and ADL events, we used state-of-the-art MKL-SVM. The
MKL-SVM is an effective machine learning tool that allows
the optimal combination of multiple kernels to build a single-
100% sensitivity while discarding almost 60% of the ADL kernel SVM. The MKL problem involves learning both, the
events. coefficients (a∗i and b∗ ) and the kernel weights (dm ) in a single
1) Feature Extraction: A set of 15 features listed in Ta- optimization problem. For kernel algorithms such as an SVM,
ble I were extracted from each simulated fall and fall-like the decision function is of the (dual) form
ADL record. These features can be categorized as postural

N
(13–15), energy-related (9–12), statistical (6–8), and reference- f (x) = a∗i K(x, xi ) + b∗ (6)
based (1–5) features [9]. The statistical and energy-related fea- i=1
tures were extracted from the AVM signal using a 3-s long
where a∗i and b∗ denote the optimized support vector and bias
data window centered at impact point (peak >3-g) as depicted
coefficients that were learned using N-labeled training instances
in Fig. 3. The duration of window was considered to cover (1×15)
{xi , yi }N
i=1 , respectively. In our case, xi ∈ R is the feature
pre- and postimpact signal patterns. The impact’s start and end
vector that represents the ith event and the label yi ∈ {−1, +1}
points were found initially and were used to extract other fea-
refers to ADL and fall classes. In the case of the MKL-SVM,
tures. More details on features 1–5, impact start (IS) and impact
the kernel K(x, xi ) is basically a convex combination of basis
end (IE) points can be found in [9], while the remaining postural
kernels
and energy-related features are described as follows.
a) Tilt angle (TA): This feature approximates the orientation 
M 
M

of the SP along with the body of the user carrying it. K(x, xi ) = dm Km (x, xi ); dm  0, dm = 1 (7)
m =1 m =1
Keeping in mind that gravity always acts vertical to the
ground, the x- and y-axes of the SP sensor are mainly where M symbolizes the total number of kernels, Km are the
responsible for capturing the gravitational acceleration  kernels for m ∈ {1, 2, . . . , M }, and the constraint dm  0,
basis
component (assuming standing position) for the waist m dm = 1 tends to result in a sparse solution of dm , i.e.,
(belt/pouch) and thigh pocket locations, respectively. As forcing some dm to be zero. The widely adopted kernels such
the posture changes from standing position to lying down, as linear, second, and third degree polynomials, and Gaussian
the TA value, calculated as follows, goes from 0◦ to 90◦ : with five different σ 2 values (0.1, 0.5, 1, 2, 3) were selected
as candidates for basis kernels (M = 8). All kernels were build
Alp [i] using the same set of input features and were normalized to unit
T A[i] = arccos (3) trace before training. In this study, we considered the following
AVMlp [i]
formulation of (primal) MKL problem:
where Alp [i] and AVMlp [i] represent, respectively, the 1  1
8 N

low-pass-filtered axis acceleration (x-, y-, or z-axes) and min ||wm ||2 + C ξi
d,w ,ξ,b 2 dm
the resultant acceleration. m =1 i=1

b) Tilt angle variation (TAV): The TA value may vary for  8


the same posture depending on the alignment difference s.t. yi wm , φm (xi ) + b  1 − ξi ∀i
between the axis of the SP and the user vertical axis. m =1 (8)
To overcome this issue, we considered the change in TA ξi  0 ∀i
during a fall-like event rather than using the TA value di-
rectly. Therefore, TAV is defined as the absolute value of 
8

the difference between the average TA values computed dm = 1, dm  0 ∀m

within intervals [IE + 0.5 s : IE + 1 s] and [IS − 1 s : IS m =1

− 0.5 s]. where wm is the mth weight vector, b represents bias, ξi is

c) Summed axis acceleration (SAA): This feature character- the ith slack variable, and hyperparameter C controls the reg-
izes the intensity of activity along a particular axis and is ularization between training errors and an optimal separating
40 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 15, NO. 1, JANUARY 2019

hyperplane. The l1 -norm constraint on the kernel weights dm

will remove redundant kernels by forcing their weights (dm ) to
be 0. The primal MKL problem (8) can also be transformed into
a following min–max optimization problem, which is based on
the standard SVM dual formulation (see [24] for details)


N
1 
N N 
8
min max αi − αi αj yi yj dm km (xi , xj )
d α 2
i i=1 j =1 m =1


N Fig. 6. SP placed around the waist or at thigh location and the SP
accelerometer axes.
s.t. αi yi = 0
i=1
A. Protocols for Fall-Like ADL and Simulated Falls
0  αi  C ∀i
Twenty volunteers (17 male and 3 female, age: 28.45 ± 2.72,

8
weight: 66.15 ± 10.83, and height: 170.7 ± 7.68) were asked
dm = 1, dm ≥ 0 ∀m . to perform the scripted set of fall-like ADLs and simulated falls
m =1
(9) given in Table II as realistically as possible in our experiment.
Here, {αi }N are the lagrange multipliers. The aforemen- During the trials, the subjects carried two LG G2 SPs: one placed
i=1
tioned min–max problem can be solved using a gradient- in a pouch around the waist and the other at thigh position in
descent-based algorithm named SimpleMKL [24]. By keeping d the trouser pocket as in Fig. 6. The application was configured
fixed, the maximization problem over α is solved via any stan- to sample at the rate of 64 Hz and save only fall-like events in
dard SVM solver. After plugging the α∗ values, the remaining a file. An informed consent was taken before the experiments
minimization problem over d is optimized using the gradient and the protocols were approved by the ethics committee of our
descent method. These two steps keep iterating until the stop- institute.
ping criteria is reached, i.e., either the duality gap (see [24, eq. The volunteers performed four different types of falls com-
15]) reduced to less than 0.01 or a maximum number of 500 mon to elderly people: forward, backward, lateral, syncope, or
iterations is reached. vertical down fall with different body dynamics. The exper-
For model selection, ten repetitions of fivefold cross valida- iments were designed based on the previous simulated falls
tion (CV) was used to evaluate the MKL-SVM models with dif- reported in the literature and the video recordings of real-world
ferent misclassification penalty hyperparameter C values. For falls [2]. The fall events were performed over a soft mat to avoid
each repetition of fivefold CV, the dataset was divided into potential injuries. The subjects were asked to maintain the final
five equal-sized stratified partitions. Each MKL-SVM model posture for at least 10 s after each fall event.
(C) was tested for 50 times and the average performance was With regard to nonfall actions, we mainly focused on fall-like
calculated. With increase in C, the classification performance ADLs that can be easily misinterpreted as falls [9] as listed in
(i.e., accuracy, sensitivity, and specificity) kept improving until Table II. In addition, we also deliberately monitored the ADL
a certain point, beyond which it became constant. However, the of three subjects over a continuous period of five days to cover
number of selected kernels also increases with C resulting in in- other possible fall-like ADLs. As a result, additional fall-like
creased computational cost. Therefore, we chose the minimum events were also recorded (cycling, sitting in car) and included
possible value of C that yields the best performance. Eventu- in our final dataset. As a result, 209 simulated falls and 137 fall-
ally, the parametric values set for the model deployed in SP are like ADL events were recorded from around the waist, while
C = 400 for waist and C = 250 for thigh locations. 175 simulated falls and 196 fall-like ADL events were detected
for the thigh location.

III. EXPERIMENTAL DESIGN B. Protocols for Free-Living Trials

Two user trials, according to Sections III-A and III-B, re- In this setting, four (new) young volunteers (three male and
spectively, were conducted to develop and evaluate FallDroid one female, age: 29.5 ± 3.12, weight: 68 ± 11.6, and height:
with the help of young volunteers. The first systematic trial 169.25 ± 3.4) were asked to carry one SP pouched around the
(see Section III-A) was carried out in a laboratory setting that waist and the other in the pant pocket. They were instructed to
included fall-like ADL events and simulated falls. This data keep the phones with them during the day for an entire week
were mainly used for model selection, classifier (MKL-SVM) and perform their everyday routine. The application was con-
training, threshold, and feature selection in the fall detection figured to record the number of times each step of the proposed
algorithm. The second trial (see Section III-B) involved a week algorithm executed. Additionally, each fall-like event, its corre-
long monitoring of the application in real-world scenarios so sponding inactivity and posture test values, and the fall alarms
as to analyse the false alarm rate. Finally, the power consump- (if generated) during the monitoring period were also logged
tion of the application was also evaluated in multiple ways as by the application. In the absence of any real fall events, all
described in Section III-C. recorded fall alarms are treated as false alarms.
SHAHZAD AND KIM: FALLDROID: AN AUTOMATED SMART-PHONE-BASED FALL DETECTION SYSTEM USING MULTIPLE KERNEL LEARNING 41

TABLE II
TYPES OF SIMULATED FALLS AND FALL-LIKE ADLS EVENTS CONSIDERED IN OUR STUDY

C. Protocols for Power Consumption Analysis TABLE III

OFFLINE PERFORMANCE ANALYSIS AND COMPARISON WITH
A crucial factor in SP-based fall detectors is power EXISTING APPLICATIONS
consumption. The continuous use of multiple sensors and
computationally complex functions substantially drain the bat-
tery power thus, making applications practically unfeasible to
use. In general, SP battery consumption largely leans on various
factors including the number of running applications, active
radio services (Wifi, etc.), and user activity. To have a better
estimate of the power consumed by FallDroid, and to avoid any
possible interruptions, the phone was reset to factory settings, The average response time (fall event to alert generation) of FallDroid App is 8.9 s.
all running applications were closed, and put to airplane mode
during Scenarios I and II. The location services were however,
kept active for all scenarios and phone was charged to its
IV. APPLICATION PERFORMANCE EVALUATION
maximum capacity prior to every experiment. The battery state
was monitored and logged via the Battery Log application. In this section, the performance of online and offline clas-
All the experiments were conducted using LG G3 with a sification as well as the power consumption of FallDroid is
3000-mAh battery, Invensense embedded acceleromter sensor, investigated and compared with relevant existing works.
Quad-core 2.7 GHz Krait 450 CPU, Adreno 420 GPU, 3-GB
RAM and Android OS v4.4.2. The scenarios are distinguished
A. Classification Performance Analysis
as follows.
1) Scenario I (Wake-lock and sensor usage): In this scenario, 1) Offline Analysis: The dataset acquired through trials de-
the application is configured to sense the accelerome- tailed in Section III-A was used for offline performance
ter data without processing it. The application holds the analysis of the proposed MKL-SVM classifier in MAT-
partial-wake-lock to prevent the CPU from sleeping and LAB. The final model (C = 400 for waist and C = 250
the fall detection algorithm is not running. for thigh) was evaluated using ten repetitions of fivefold
2) Scenario II (Running algorithm): The user carries the SP CV. At each repetition, the dataset was divided into strat-
with only the proposed algorithm (complete FD App) run- ified partitions. The results for the performance measures
ning on it. Assuming that the SP is not used for any other such as sensitivity, specificity, and accuracy have been
purpose, this scenario includes the battery consumption averaged over 50 iterations and reported in Table III. To
of the algorithm. perform a fair comparison, the algorithms presented in
3) Scenario III (SP usage in real life): This scenario realizes the related works were implemented and tested using the
our everyday routine where the user carries the SP and same dataset. It is apparent that our proposed fall de-
uses it as in real life. The fraction of battery drained tection algorithm outperforms the existing works, yield-
by FallDroid is reported using the Android OS battery ing 99.52% sensitivity, 95.19% specificity, and 97.81%
diagnostic tool. accuracy.
42 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 15, NO. 1, JANUARY 2019

TABLE IV
PERFORMANCE COMPARISON (OFFLINE) OF DIFFERENT MACHINE LEARNING
ALGORITHMS USING WAIST LOCATION DATASET

TABLE V
PERFORMANCE COMPARISON (OFFLINE) OF DIFFERENT MACHINE LEARNING
ALGORITHMS USING THIGH LOCATION DATASET
Fig. 8. Scatter plots of some of the reported features (see Table I)
using thigh location dataset (see Section III-A) to compare ADL and fall
events. Each plot title represents the category of two plotted features.

proposed features SAA and TAV provide good discrimi-

nation of ADL and fall events. By comparing the results
in Tables IV and V, we can conclude that the major per-
formance enhancement is due to the proposed features
set. Whereas, the MKL-SVM classifier further improved
the overall accuracy while outperforming all the other
classifiers tested here.
2) Online Analysis: The performance of FallDroid was also
analyzed in the everyday life scenario in terms of the
false alarm rate. The total monitoring time (SP carried
duration), number of times each step of the algorithm was
used (NSi , see, Fig. 2), and the false alarms triggered dur-
ing the monitoring period for each volunteer is summa-
rized in Table VI. Interestingly, the average time interval
between two consecutive false alarms was calculated to
be 59 h (16 h) for waist (thigh) location.
B. System On-Time and Power Consumption Estimation
To report the battery consumption, two trials of Scenarios I
and II each were conducted. For Scenario II, FallDroid and SFD
Fig. 7. Scatter plots of some of the reported features (Table I) using [17] were both tested. The averaged results for the battery drain
waist location dataset (see Section III-A) to compare ADL and fall events.
Each plot title represents the category of two plotted features. of LG G3 SP over a time period of 24 h are illustrated in Fig. 9.
The add-on power consumption of FallDroid with respect to
the baseline scenario (Scenario I) is observed to be lesser than
In order to find the root cause of high accuracy, i.e.,
that of the SFD. FallDroid consumes almost 24% of the battery
whether it is the classification algorithm (MKL-SVM)
charge in 24.32 h in Scenario II. The estimated battery lifetime
or the proposed features set, we compared the per-
was almost 86 h (24.32-h usage plus 62-h expected residual
formance of several other machine learning techniques
usage time).
with the MKL-SVM using the proposed features set.
The classifiers tested were SVM with single kernel
(RBF, linear, and polynomial), ANN, k-NN, and Naı̈ve V. DISCUSSION
Bayes. The best performing model for each classifier The better performance of our fall detection algorithm is evi-
(parameters optimization) was selected using fivefold CV dent in both, offline and online analysis. In addition, the simple
in WEKA software. The averaged performance using ten GUI of FallDroid and support for multiple carrying locations ul-
repetitions of fivefold CV of each classification algorithm timately improves user convenience. Furthermore, to minimize
on waist and thigh location datasets is presented in Ta- power consumption, the algorithm has been designed to reduce
bles IV and V. Unlike ANN models of [9] and [17] (in the average computational cost. This is achieved by avoiding
Table III), a similar ANN classifier achieved much higher power-intensive signal processing procedures and by largely
performance (accuracy = 96.99%) on proposed fea- exploiting the simple TBM phase of the algorithm most of the
tures set. It can also be seen in Figs. 7 and 8 that the time.
SHAHZAD AND KIM: FALLDROID: AN AUTOMATED SMART-PHONE-BASED FALL DETECTION SYSTEM USING MULTIPLE KERNEL LEARNING 43

TABLE VI
ONLINE PERFORMANCE ANALYSIS; NUMBER OF EXECUTIONS OF EACH STEP (NSi ), TOTAL MONITORING TIME (SP CARRIED DURATION), AND FALSE ALARM
RATE IN FREE-LIVING ADL TRIALS

location, and consequently, yielding a higher false alarm rate.

The SP self/independent/free movements inside the pocket, fre-
quent intense movements of thigh, and the difficulty in isolating
sitting from lying by employing TA or TAV features are the main
factors causing higher fall-like event counts. The false alarms
rate of 1 alarm per 16 h of usage (at thigh location) reported
herein is possibly of annoyance to users and their emergency
contacts, but this rate decreases if the user cancels the alert
proactively.
In contrast to many earlier efforts that were evaluated in labo-
ratory environments using scripted simulated ADL movements,
FallDroid was tested with free-living trials as well. This real-life
testing protocol is better than its scripted counterpart as the users
Fig. 9. Battery consumption comparison of FallDroid App (proposed) acts naturally thus, covering their complete everyday actions in-
and SFD App with respect to reference consumption (wake-lock and
sensor usage) as described in Scenarios II and I, respectively, of cluding several unexpected events. Nonetheless, the application
Section III-C. was assessed using young volunteers. But since elderly people
show reduced activities and have a more sedentary lifestyle, it is
Used in the second phase of the algorithm, the MKL-SVM expected that the system would perform even better when used
achieves excellent classification performance as compared to [9] by elderly. The performance can further be enhanced through
and [17], which are both built upon the concept of the two-layer personalization of the App [26].
feed-forward ANN. The performance of the approaches men- For any SP-based monitoring system, power consumption
tioned in the aforementioned references on our dataset however, due to keeping the CPU awake (partial-wake-lock) and sensors
was not appealing. This is mainly because our dataset consists of provides the baseline as represented in Scenario I. The added
merely fall-like events, i.e., confusing signal patterns that look battery drain in Scenario II is due to the active fall detection
much like fall signals, providing a robust performance analysis. algorithm. As visible in Fig. 9, the battery is majorly consumed
Bagalá et al. [25] investigated and compared the performance for keeping the CPU awake and sensor usage. Since FallDroid
of well-adopted fall detection algorithms based on waist or trunk invokes the computationally lower TBM unit most of the time,
accelerometer measurements. Their study showed that the num- 86% of the fall-like events are rejected without involvement of
ber of false alarms (false positives) per day in real-life scenario the MKL-SVM classifier. The overhead battery consumption of
ranged from 3 to 85, depending on the technique applied. Kerde- our algorithm is very less with a reduction in system ON-time
gari et al. [17] evaluated the online performance of the SFD by 6 h.
using four volunteers monitored for a single day and reported a The estimated system ON-time is calculated as the usage time
total of five false alarms per day. The results in Table VI reveal on battery plus the estimated remaining time to use. With the
that our algorithm achieves the lowest false alarm rate as com- CPU kept awake and samples collected by the accelerometer,
pared to the existing works [17], [25] (i.e., 1 alarm per about 59 the device ON-time was obtained to be 92 h. This however,
h of usage at waist location). reduced to 86 h (or 75 h) if we executed FallDroid (or SFD).
With regard to the thigh position however, the performance These values were recorded only when the user did not use
was not as good as waist location. The SP is loosely attached to the SP for any other purpose (scenario II). In real life, these
the user body inside the trouser pocket thus, generating added values depend hugely on the SP usage frequency of the user.
noise due to its own movement. Additionally, the number of The FallDroid App drains 26.52% battery over a period of 25 h
fall-like events generated per day are much higher than the waist 17 min during real-life usage of SP (Scenario III).
44 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 15, NO. 1, JANUARY 2019

VI. CONCLUSION AND FURTHER WORKS [12] R. Luque, E. Casilari, M.-J. Morn, and G. Redondo, “Comparison and
characterization of android-based fall detection systems,” Sensors, vol. 14,
In this paper, we developed FallDroid, an Android-based fall no. 10, pp. 18 543–18 574, 2014.
detection and emergency alert application that supports multiple [13] E. Casilari, R. Luque, and M.-J. Morn, “Analysis of android device-based
solutions for fall detection,” Sensors, vol. 15, no. 8, pp. 17 827–17 894,
carrying locations (waist and thigh) for SPs and uses a simple 2015.
user-friendly GUI for better accessibility. The proposed two- [14] F. Sposaro and G. Tyson, “ifall: An android application for fall monitoring
step fall detection algorithm achieves excellent classification and response,” in Proc. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc., 2009,
pp. 6119–6122.
results on fall-like events. The first stage adopts TBM to effec- [15] J. Dai, X. Bai, Z. Yang, Z. Shen, and D. Xuan, “Perfalld: A pervasive
tively discard most of the ADL data, whereas the MKL-SVM is fall detection system using mobile phones,” in Proc. 8th IEEE Int. Conf.
used in the second step for classifying difficult fall-like events. Pervasive Comput. Commun. Workshops., 2010, pp. 292–297.
[16] M. V. Albert, K. Kording, M. Herrmann, and A. Jayaraman, “Fall classifi-
Such structure helps reducing false alarms while maintaining cation by machine learning using mobile phones,” PloS One, vol. 7, no. 5,
low computation cost on average, resulting in lesser power con- p. e36556, 2012.
sumption. The proposed algorithm was shown to outperform [17] H. Kerdegari, S. Mokaram, K. Samsudin, and A. R. Ramli, “A pervasive
neural network based fall detection system on smart phone,” J. Ambient
the existing applications [9], [17] in terms of both, offline and Intell. Smart Environ., vol. 7, no. 2, pp. 221–230, 2015.
online analysis. [18] C. Medrano, R. Igual, I. Plaza, and M. Castro, “Detecting falls as novelties
As future work, the application can be tested in normal rou- in acceleration patterns acquired with smartphones,” PloS One, vol. 9,
no. 4, p. e94811, 2014.
tine usage involving the elderly population. Such recorded data [19] A. T. Özdemir and B. Barshan, “Detecting falls with wearable sensors
on the cloud storage will serve very useful for further refinement using machine learning techniques,” Sensors, vol. 14, no. 6, pp. 10 691–
of the algorithm. Additionally, by using individual user data to 10 708, 2014.
[20] IDC Worldwide Mobile Phone Tracker. 2016. [Online]. Available:
train the MKL-SVM classifier, a personalized user-dependent http://www.idc.com/
fall detection system can be anticipated, enabling us to compare [21] M. Clark, J. Lim, G. Tewolde, and J. Kwon, “Affordable remote health
the reduction in false alarms. Finally, in order to enhance the monitoring system for the elderly using smart mobile device,” Sensors
Transducers, vol. 184, no. 1, p. 77, 2015.
overall system performance, another potential future work could [22] Google Play Store. Accessed: Jun. 15, 2016. [Online]. Available:
be on the feature engineering/extraction part to find even better https://play.google.com/store
features. Currently, the AAMV, TAV, and SAA metrics seem [23] A. Bourke, J. Obrien, and G. Lyons, “Evaluation of a threshold-based tri-
axial accelerometer fall detection algorithm,” Gait Posture, vol. 26, no. 2,
quite good at fall detection, but perhaps higher order derivatives pp. 194–199, 2007.
or other nonlinear features comprised of complementary fea- [24] A. Rakotomamonjy, F. R. Bach, S. Canu, and Y. Grandvalet, “Simplemkl,”
tures could be considered especially in the TBM part (first stage J. Mach. Learn. Res., vol. 9, pp. 2491–2521, 2008.
[25] F. Bagal et al., “Evaluation of accelerometer-based fall detection algo-
of algorithm) to increase specificity. rithms on real-world falls,” PloS One, vol. 7, no. 5, p. e37062, 2012.
[26] C. Medrano, I. Plaza, R. Igual, Á. Sánchez, and M. Castro, “The effect
of personalization on smartphone-based fall detectors,” Sensors, vol. 16,
REFERENCES no. 1, p. 117, 2016.

[1] C. Todd, D. Skelton, and World Health Organization, What are the Main Ahsan Shahzad (S’14) received the B.Eng. de-
Risk Factors for Falls Amongst Older People and What are the Most gree in electrical engineering with honors from
Effective Interventions to Prevent These Falls? World Health Organization, the University of Engineering and Technology,
Geneva, Switzerland, 2004. Lahore, Pakistan, in 2009, and M.Eng. degree
[2] S. N. Robinovitch et al., “Video capture of the circumstances of falls in in electronics engineering from the Gwangju
elderly people residing in long-term care: An observational study,” Lancet, Institute of Science and Technology, Gwangju,
vol. 381, no. 9860, pp. 47–54, 2013. Korea, in 2014.
[3] S. R. Lord, C. Sherrington, H. B. Menz, and J. C. Close, Falls in Older He is currently a Postgraduate Student with
People: Risk factors and Strategies for Prevention. Cambridge, U.K.: the School of Electrical Engineering and Com-
Cambridge Univ. Press, 2007. puter Science, GIST. His research interests in-
[4] Y. S. Delahoz and M. A. Labrador, “Survey on fall detection and fall clude mhealth systems, biomedical signal pro-
prevention using wearable and external sensors,” Sensors, vol. 14, no. 10, cessing and application design, gait analysis, and pattern recognition.
pp. 19 806–19 842, 2014.
[5] R. Igual, C. Medrano, and I. Plaza, “Challenges, issues and trends in fall
detection systems,” Biomed. Eng. Online, vol. 12, no. 1, p. 66, 2013.
[6] Lifecall Fallalert System. Accessed: Sep. 15, 2016. [Online]. Available: Kiseon Kim (M’84–SM’98) received the B.Eng.
http://lifecall.com/ and M.Eng. degrees in electronics engineering
[7] C. Wang et al., “Low-power fall detector using triaxial accelerometry and from Seoul National University, Seoul, Korea, in
barometric pressure sensing,” IEEE Trans. Ind. Informat., vol. 12, no. 6, 1978 and 1980, respectively, and the Ph.D. de-
pp. 2302–2311, Dec. 2016. gree in electrical engineering systems from the
[8] A. Shahzad, S. Ko, S. Lee, J.-A. Lee, and K. Kim, “Quantitative as- University of Southern California, Los Angeles,
sessment of balance impairment for fall-risk estimation using wearable CA, USA, in 1987.
triaxial accelerometer,” IEEE Sensors J., vol. 17, no. 20, pp. 6743–6751, From 1988 to 1991, he was with Schlum-
Oct. 2017. berger, Houston, TX, USA. From 1991 to 1994,
[9] S. Abbate, M. Avvenuti, F. Bonatesta, G. Cola, P. Corsini, and A. Vecchio, he was with the Superconducting Super Collider
“A smartphone-based fall detection system,” Pervasive Mobile Comput., Lab, TX. In 1994, he joined the Gwangju Insti-
vol. 8, no. 6, pp. 883–899, 2012. tute of Science and Technology, Gwangju, Korea, where he is currently
[10] J. Zhou, P.-L. P. Rau, and G. Salvendy, “Older adults use of smart phones: a Professor. His current research interests include biomedical applica-
An investigation of the factors influencing the acceptance of new func- tions design, wideband digital communications system design, sensor
tions,” Behav. Inf. Technol., vol. 33, no. 6, pp. 552–560, 2014. network design, and analysis and implementation both at the physical
[11] M. A. Habib, M. S. Mohktar, S. B. Kamaruzzaman, K. S. Lim, T. M. layer and at the resource management layer.
Pin, and F. Ibrahim, “Smartphone-based solutions for fall detection and Prof. Kim is the Member of the National Academy of Engineering of
prevention: Challenges and open issues,” Sensors, vol. 14, no. 4, pp. 7181– Korea, a Fellow of the IET, and the Senior Editor of the IEEE SENSORS
7208, 2014. JOURNAL.