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

net/publication/220141331

Mobile Phone-Based Pervasive Fall Detection

Article in Personal and Ubiquitous Computing · October 2010

DOI: 10.1007/s00779-010-0292-x · Source: DBLP

CITATIONS READS

167 2,959

5 authors, including:

Jiangpeng Dai Zhimin Yang

GEIRINA The Ohio State University
20 PUBLICATIONS 1,006 CITATIONS 16 PUBLICATIONS 716 CITATIONS

SEE PROFILE SEE PROFILE

Zhaohui Shen
Zhejiang University of Technology
14 PUBLICATIONS 822 CITATIONS

SEE PROFILE

All content following this page was uploaded by Zhimin Yang on 05 May 2014.

The user has requested enhancement of the downloaded file.

Pers Ubiquit Comput
DOI 10.1007/s00779-010-0292-x

ORIGINAL ARTICLE

Mobile phone-based pervasive fall detection

Jiangpeng Dai • Xiaole Bai • Zhimin Yang •

Zhaohui Shen • Dong Xuan

Received: 21 January 2010 / Accepted: 2 March 2010

Ó Springer-Verlag London Limited 2010

Abstract Falls are a major health risk that diminishes the achieves superior detection performance and power
quality of life among the elderly people. The importance of efficiency.
fall detection increases as the elderly population surges,
especially with aging ‘‘baby boomers’’. However, existing Keywords Pervasive fall detection
Mobile phones

commercial products and academic solutions all fall short Context information
Accelerometer

of pervasive fall detection. In this paper, we propose uti- Magnetic field sensor
Accessory
lizing mobile phones as a platform for developing perva-
sive fall detection system. To our knowledge, we are the
first to do so. We propose PerFallD, a pervasive fall 1 Introduction
detection system tailored for mobile phones. We design
two different detection algorithms based on the mobile 1.1 Motivation
phone platforms for scenarios with and without simple
accessories. We implement a prototype system on the Falls are a major health hazard for elderly people [1] and
Android G1 phone and conduct extensive experiments to also a major obstacle to their independent living [2, 3]. The
evaluate our system. In particular, we compare PerFallD’s estimated fall incidence every year for both institutiona-
performance with that of existing work and a commercial lized and independent living people over 75 is at least 30%
product. The experimental results show that PerFallD [4, 5]. The frequency of falling is considerably higher
among more dependent elderly. Researchers estimate that
up to 50% of nursing home residents fall each year and
J. Dai (&) more than 40% of them might fall more than once [6]. Falls
Key Laboratory of Computer Network and Information
not only cause physical injuries such as disabling fractures
Integration, Ministry of Education, Southeast University,
Nanjing 210096, China [7], but also have dramatic psychological consequences
e-mail: jpdai@seu.edu.cn; jpdai@cse.ohio-state.edu that reduce elderly people’s independence [8, 9]. This situ-
ation deteriorates as the elderly population surges.
J. Dai
X. Bai
Z. Yang
D. Xuan
According to a report from the U.S. Census Bureau, there
Department of Computer Science and Engineering,
The Ohio State University, Columbus, OH 43210, USA will be a 210% increase of the population aged 65 and over
e-mail: baixia@cse.ohio-state.edu within the next 50 years, in part due to aging ‘‘baby
Z. Yang boomers’’ [10].
e-mail: yangz@cse.ohio-state.edu The considerable risk of falls and the substantial
D. Xuan increase of the elderly population stimulate both com-
e-mail: xuan@cse.ohio-state.edu mercial product development and academic research on fall
detection. A typical fall detection system has two major
Z. Shen
functional components: the detection component and the
Division of Physical Therapy, The Ohio State University,
Columbus, OH 43210, USA communication component. As their names imply, the
e-mail: shen.119@osu.edu detection component detects falls and the communication

123
 Pers Ubiquit Comput

component communicates with emergency contacts after a communication services have vast coverage. One might
fall is detected. Figure 1 depicts a typical commercial fall argue that elderly people may not accept such mobile
detection system, which is provided by Brickhouse [11]. phones. However, we argue that elderly people may
The system consists of a portable sensor (the detection prefer to have a single phone with self-contained fall
component on the left side of the figure) and a tele-assist detection functionality than to carry a separate fall
base (the communication component on the right side of detection device on their bodies. In addition, more recent
the figure). data illustrate the increasing popularity of these phones.
The major problem with existing commercial products is The minimum requirement for such a mobile phone
that they cannot provide pervasive fall detection. Consider platform is the presence of a simple sensor, e.g., an
the aforementioned product as an example. The base must accelerometer. Currently, many phones, especially
be installed somewhere indoors and the portable sensor smartphones, contain multiple types of sensors, including
must be attached to a belt around the waist. Once the base accelerometers. Such phones are popular and thoroughly
receives the signal from the sensor indicating a fall, it can accepted in society. As shown in Fig. 2, over 120 million
automatically communicate with a preset emergency con- smartphones were sold in 2008 [12], and their popularity
tact using a fixed phone. However, the maximum distance is projected to continuously increase in the near future
between the sensor and the base is limited. Fall detection due to decreasing price. Recently, several leading tele-
can only be conducted within a small indoor environment communication companies, e.g., AT&T, have made
and elderly people need to bring the sensor everywhere. available affordable smartphones [13, 14], whose features
They may easily forget to do so, as it is an extra device that are similar to those of high-end models, in addition to
they seldom use in daily life. Furthermore, these products cheaper service plans [15].
are expensive. The aforementioned system costs $199.95 We summarize our contributions in this paper as
for the devices and $419.40 per year for monitoring service follows.
[11].
– We propose utilizing mobile phones as the platform
Besides the commercial products, existing academic
for developing pervasive fall detection system. To our
research also has deficiencies that hinder pervasive fall
best knowledge, no existing commercial products
detection. More detailed information on the state of the art
and academic work use mobile phones to inte-
is provided in Sect. 2.
grate comprehensive fall detection and emergency
communication.
1.2 Our contributions
– We design two algorithms for fall detection systems
using mobile phones. The first is an acceleration-based
In this paper, we propose utilizing mobile phones as the
detection approach. The second algorithm is designed
platform for developing pervasive fall detection system, as
for the case when a certain accessory can be used to
they naturally combine the detection and communication
further capture human behavior information. This
components. To the best of our knowledge, we are the first
algorithm is based on shape context and Hausdorff
to do so.
distance.
As self-contained devices, mobile phones present a
mature hardware and software environment for develop-
ing pervasive fall detection system. Mobile phone-based
fall detection system can function almost everywhere,
since mobile phones are highly portable, all necessary
components are already integrated therein, and their

Fig. 1 A typical commercial fall detection system provided by

Brickhouse [11] consists of a portable sensor (the detection compo-
nent shown in left) and a tele-assist base (the communication Fig. 2 From shipment data, smartphones are widely accepted and
component shown in right) increasing in popularity in the near future

123
Pers Ubiquit Comput

– We design and implement a pervasive fall detection When acceleration is used, the most widely used
system, PerFallD, on the mobile phone-based platform methods are based on thresholds. In [19], Nyan et al. make
to conduct fall detection with or without other small an accelerometer be settled into garment on the shoulder
accessories. PerFallD has few false negatives and false position. They use a threshold of absolute peak values of
positives. It is available in both indoor and outdoor acceleration to determine falls. Kangas et al. in [20] pro-
environment. Besides being user-friendly, it requires no pose four thresholds for total sum vector, dynamic sum
extra hardware and service cost. It is also lightweight vector, vertical acceleration and difference between the
and power-efficient. maximum and minimum acceleration values. A fall is
– We conduct extensive experiments, with both a considered detected as long as one threshold is exceeded.
mannequin and real persons, to evaluate detection In [21] and [22], information of body orientation and
accuracy. The experimental results show our detection posture are used to complement acceleration-based detec-
system achieves superior performance in terms of low tion. All aforementioned work show that the acceleration
false negative and low false positive in fall detections threshold–based detection works well in practice. How-
with or without accessory. For the purpose of com- ever, the detection devices used in them are specified and
parison, we implement algorithms provided in existing not conveniently portable. The communication component
work and also test a typical commercial fall detection that is also critical in a fall detection system is ignored in
product. PerFallD outperforms existing algorithms and these work. Mover [23] is proposed by Fraunhofer recently
achieves better balance between false negative and as a kind of activity monitor and fall detector for Android
false positive when compared with the commercial system. It uses the accelerometers on phones to collect the
product. accelerations and sum them to measure human activity
level. It may also detect user falls; however, the algorithm
Paper organization The rest of the paper is organized as is still being tested as the developer announced. Now,
follows. Section 2 presents the related work. We introduce Mover is more like a simple interesting APP on Android.
the system design in Sect. 3 and system implementation in Ganti et al. in [24] and Karantonis et al. in [25] propose
Sect. 4. In Sect. 5, we evaluate our system with extensive storing sensed user behavior data into a database for vari-
experiments. Section 6 concludes the paper. ous activities, e.g., fall down, recognition thereafter. These
databases built with the sensed data are very useful. They
can be used to detect various normal or abnormal activities.
2 Related work However, it is not a trivial task to collect enough data for
each individual to build up the database.
There are very few fall detection commercial products. Fall Fu et al. in [26], Sixsmith et al. in [27], Miaou et al. in
detector provided by Brickhouse [11] consists of a tele- [28] and Jansen et al. in [29] propose capturing images of
assist base and a portable sensor. The base device needs to people and then detecting visual falls based on image-
be installed indoors and the signal transmission distance processing techniques. Such approaches have limitations
between the sensor and the base is limited. ITT EasyLifeS on pervasive detection, affordability and acceptability. The
[16] is one kind of cellphone that equipped with balance detection area is limited within the monitoring environ-
sensor. The manufacture claims that the phone will auto- ment, which is costly to build up. The people’s privacy is
matic dial SOS numbers if it is dropped. However, the also compromised.
device is too specific and the triggering condition is too We also notice there are work that propose an integrated
trivial to provide pervasive and comprehensive fall detec- fall detection environment. In [30], a separated fall detector
tion. Philips Lifeline [17] provides medical alert service is connected to a mobile phone, which is used to query the
requiring the elderly people to push a certain button to user about his condition when the fall detector signals it. In
report a fall. It cannot handle certain critical cases, e.g., the [31], the mobile phone acts as a relay between the sensors
elderly people fall down with a sudden faint. Betterbuys and a remote server that runs the fall detection algorithm.
[18] provides Economical Fall Alarm Monitor that requires Paper [32] proposes using a network of fixed motes to
sensors deployed with chair pad, bed pad, floor mat, or provide location information of the victim after a fall has
chair seatbelt. It has more limitations to achieve pervasive been detected. Lieberman-Aiden invents ishoe [33], an
detection. electronic shoe insole, to help users manage their balance
Significance of fall detection also attracts academic before a fall occurs. It uses sensors to track users’ balance
research. Proposed fall detection techniques can be clas- patterns and transmits data to computers for interpretation.
sified into three categories: acceleration-based detection, In these pioneering work, reliability and availability of
databases-based motion type classification and image pro- cooperation between different devices raise concerns. The
cessing–based detection. system cost is also high.

123
 Pers Ubiquit Comput

3 PerFallD design background while other components of the program halt;

(2) the sampling frequency can be adjusted; (3) the pattern
In this section, we present the PerFallD design. We present matching process is launched only after daemon-collected
the system overview followed by the design of the detec- data exceed the preset threshold; and (4) hardware such as
tion algorithms. Note that the design is general—it is not the screen is activated only when necessary.
constrained to a particular brand or type of mobile phone. The modules within the smaller dashed box present the
algorithm design part. We will introduce the detection
3.1 System overview algorithms in the following sections.

PerFallD’s workflow is illustrated in Fig. 3. Right after the 3.2 Algorithm design without accessory
program starts, a user profile will be loaded. A user-
dependent profile contains basic fall detection configura- In this section, we present an acceleration-based detection
tion such as the default sampling frequency, default algorithm designed for mobile phones.
detection algorithm, emergency contact list, etc. In differ- Accelerometers usually provide the acceleration read-
ent scenarios, users’ activity patterns have varying degrees ings in directions of x-, y-, and z-axis. Accelerations in
of rapidity, and it is more efficient to use different sampling these directions are represented by Ax, Ay and Az, respec-
frequencies in different scenarios. After the user profile is tively. For generality, we assume the directions of x-, y-,
loaded, we provide users the chance to adjust the sampling and z-axis are determined by the posture of the phone. As
frequency as interfaces that invoke sensor functions at illustrated in Fig. 4, the x-axis has positive direction toward
different frequencies are provided. Then the main program, the right side of the device, the y-axis has positive direction
working as a background daemon, launches. If information toward the top of the device and the z-axis has positive
collected in real time satisfies a certain preset condition, the direction toward the front of the device. Vector AT repre-
pattern matching process begins to determine if a fall sents the total acceleration of the phone body. Its amplitude
occurs. If no fall is detected, execution immediately returns can be obtained by Eq. 1.
to the daemon. If a fall is detected, the daemon service qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
transmits a signal that triggers an alarm and starts a timer. jAT j ¼ jAx j2 þ jAy j2 þ jAz j2 : ð1Þ
If the user does not manually turn off the alarm within a A mobile phone’s orientation sensor can determine its
certain time period, the system automatically calls contacts yaw, pitch and roll values that are denoted as hx, hy and hz,
stored in the emergency contact list according to their respectively. We can further obtain the amplitude of Av, the
priorities. The phone iteratively calls and texts up to five acceleration in the absolute vertical direction, from Eq. 2.
contacts. jAv j ¼ jAx sin hz þ Ay sin hy  Az cos hy cos hz j: ð2Þ
As presented in Fig. 3, power efficiency is explicitly
considered in the design of the modules illustrated within The fall detection algorithm is based on the values of
the larger dashed box. Four steps are taken to reduce power |AT| and |Av|. If the difference of |AT| within a triggering
consumption: (1) the monitoring daemon runs in the time window wintt exceeds triggering threshold Thtt, the

Fig. 3 Working procedure of the PerFallD system. Power efficiency Fig. 4 a Acceleration readings in directions of x-, y-, and z-axis that
is explicitly considered in design of the modules illustrated within the are associated with and fixed regard to the body of the mobile phone.
bigger dashed box. The modules within the smaller dashed box b Mobile phone orientation can be determined by yaw (hx), pitch (hy)
present the algorithm design part and roll (hz)

123
Pers Ubiquit Comput

pattern recognition is triggered to check the difference

between the maximum value and the minimum value of
|AT| within a checking time window winct following wintt.
If this difference is less than another threshold Thct, a fall is
considered detected. A similar rule applies to |Av|, with
corresponding time windows wintv, wincv and thresholds
Thtv, Thcv. If both the detection conditions about |AT| and
|Av| are satisfied, a detection of fall is reported.
Figure 5 presents the examples of |AT| and |Av| that are
obtained from the integrated accelerometer in a mobile
phone. We show the results when the phone is placed in
different locations: in a pocket of a shirt (chest), on the belt
(waist), and in a pocket of the pants (thigh). Thresholds can
be set according to the training data obtained from extensive
experiments. Based on a set of data, Fig. 6 shows the rela-
tionship between false negative and false positive for dif-
ferent values of Thtt (one threshold is represented as one Fig. 6 The relationship between false negative and false positive for
mark) when the Thct is fixed. We adjust thresholds in order to different threshold Thtt when the Thct is fixed to 80 (indicated as
reduce false negative while simultaneously keep false ‘‘range’’ in the Figure). Data are from lateral falls. The phone is
placed at the position of chest
positive in an acceptable range. More details will be pro-
vided in Sect. 5.
represents the strength of magnetic field (MF) around the
3.3 Algorithm design with a magnetic accessory phone. We then can use this value to infer the relative
position between the phone and the magnetic accessory.
The performance of detection using accelerations alone can Research in biomechanics has found that the movements
be improved by introducing some simple accessories. In of legs have salient features in falls [34]. To catch such
biomechanics, some researchers focus on the mechanics of features, we put a magnetic accessory slightly above the
human falling [34]. Some of them have made attempts to knee of left leg and the mobile phone in the right pocket of
define some common step mechanics during falling from the pants.
the viewpoint of biomechanics [35]. The mobile phones In Fig. 7, the two figures in the left column show typical
together with simple accessories can effectively capture readings of magnetic sensor for falls. The top-right figure
these common steps to help detect a fall. shows the readings at the start of walking. The bottom-right
In this section, we introduce a mobile phone-based figure shows the readings in arbitrary postures that are not
detection method with some simple accessories. In particu- falls. We find that the reading curves for falls have some
lar, we focus on a mobile phone with a magnetic field sensor unique patterns. As illustrated in Fig. 7, before a fall
and a magnetic accessory. The reading value from the sensor happens, the MF values are on a relatively stable high
level. At the beginning of a fall, the MF value decreases
Chest Waist Thigh slowly, but suddenly goes higher, making a salient convex
25 25 25
shape, and then decreases fast. After it reaches the mini-
20 20 20
mum, it turns back to a higher value that is much lower
|A T |

15 15 15
10 10 10 than the values before falling, and then keeps relatively
5 5 5 stable afterwards. This interesting observation can be
0 1 2 3 4 0 1 2 3 4 0 1 2 3 4
explained from the view of biomechanics. Before and after
20 20 20
a fall, the human postures are significantly different. The
15 15 15
distance between two legs when people are standing is
|A v|

10 10 10
5 5 5 often smaller than that when people are lying on the floor
0 0 0 after a fall. Furthermore, during a fall, the legs bend and
0 1 2 3 4 0 1 2 3 4 0 1 2 3 4
Time(s) Time(s) Time(s) cross slightly, which causes knee movement. This slightly
increases the MF value before it decreases further.
Fig. 5 Examples of the amplitude of AT and Av that are calculated Our algorithm aims to capture the aforementioned
out from the readings of the integrated accelerometer in a mobile
unique characteristics effectively and to distinguish the
phone during a fall. We show the results when the phone is placed in
different locations: in the pocket of a shirt (chest), on the belt (waist), readings that indicate falls from others. We present some
and in a pocket of the pants (thigh) preliminaries before introducing the detection algorithm.

123
 Pers Ubiquit Comput

fall non−fall If the check at the first step returns a true, the operations in
8000 5000
the second step will be triggered at a certain short time
4500
6000 after the end point of the second decrement. In the second
4000
4000
step, the shape contexts of two end points in successive
3500 decrements of the current sensor reading curve will be
2000
3000 calculated and combined. Then we compute Hausdorff
0 2500 distance between this new generated shape context and the
0 1 2 3 4 0 1 2 3 4
seconds seconds known shape context from the sensor readings during fall,
6000 6000
for comparison. If the Hausdorff distance is smaller than a
threshold Thhd, the current reading curve is labeled as fall,
5000
5000 otherwise as non-fall. In the following, we use an example
4000
to further illustrate the algorithm.
3000
4000 We zoom in the top-left figure in Fig. 7 and then get
2000 Fig. 8. We first check whether there exist two successive
1000
0 1 2 3 4
3000
0 1 2 3 4
decrements and their total decrement exceeds the preset
seconds seconds threshold. In Fig. 8, AB and CD are two successive
decrements and it is found that their total is larger than
Fig. 7 The two Figures in the left column show the typical readings
of magnetic sensor for falls. The top-right Figure shows the readings
some threshold. Then the second step of the detection
at the start of walking. The bottom-right Figure shows the readings in algorithm launches at 1.6 s after point D occurs. We take
arbitrary postures that are not falls B as the center of one log-polar coordinate system. In
such a coordinate system, the largest distance is divided
into kd distance ranges uniformly according to the log
Successive decrements Assume the readings show a values denoted as Dr1 ; Dr2 ; . . .; Drkd ; and center angle 2p is
monotonous decrease at [t1, t2] followed by another divided into ka angle ranges uniformly denoted as
monotonous increase at [t2, t3] and then a monotonous Ar1 ; Ar2 ; . . .; Arka : Figure 8 illustrates an example of such
decrease at [t3, t4]. We call these two decrements occurred log-polar coordinate system at its bottom-left corner. Each
at time periods [t1, t2] and [t3, t4] successive. point p excluding B in the time window can be repre-
sented by distance q and angle h as shown in Fig. 8,
Shape context The concept of shape context is proposed falling into a certain distance range and angle range. Then
by Belongie et al. at [36]. For a point pi on a shape, con- we can obtain a kd 9 ka matrix MB wherein the value at
sidering the set of vectors originating from this point pi to position (i, j) denotes the number of points that fall into
all other sample points on the shape, we compute a coarse
histogram hi of the relative coordinates of the remaining
n - 1 points by Eq. 3.
hi ðkÞ ¼ #fq 6¼ pi : ðq  pi Þ 2 binðkÞg: ð3Þ
This histogram is defined to be the shape context of pi. We
use bins that are uniform in log-polar space, making the
descriptor more sensitive to the positions of nearby sample
points than to those of points farther away. Interested
readers may refer to [36] for more details.
Hausdorff distance Hausdorff distance is a classical
correspondence-based measure on how far two subsets of a
metric space are from each other. As pointed out in some
recent work, such as [37], Hausdorff distance can be used
for shape comparison. If X and Y are two non-empty sub-
sets of a metric space (M, d), we define their Hausdorff
distance dH(X, Y) by Eq. 4.
maxfsupx2X inf y2Y dðx; yÞ; supy2Y inf x2X dðx; yÞg: ð4Þ
Fig. 8 Y axis represents the magnetic sensor readings and X axis
Our algorithm has two steps. In the first step, the sum of
represents seconds. AB and CD are two successive decrements with B
every two successive decrements in sensor reading curve and D the end points. An example of a log-polar coordinate system is
will be checked to see whether it exceeds a threshold Thtm. shown at the corner

123
Pers Ubiquit Comput

the distance range Dri and angle range Arj : We take D as fall non−fall
the center of another log-polar coordinate system that is
the same as mentioned earlier. We can obtain another

log R

log R
matrix MD. These two matrices represent the shape con-
texts of two end points in the successive decrements. We
combine MB and MD together into a 2kd 9 ka matrix
θ θ
M. This M is illustrated as an image with gray color level
shown as the top-left image in Fig. 9. A darker color
indicates a higher element value in the matrix. Figure 9

log R

log R
shows images of different combined shape context
matrices obtained from corresponding figures in Fig. 7.
The matrices of falls are similar, while the matrices of
θ θ
non-fall have great differences. That similarity can be
captured by Hausdorff distance. If Hausdorff distance Fig. 9 Image representation of combined shape context matrices (2
between one M and the Ms known from falls is small 9 5 distance ranges and 12 angle ranges) corresponding to Figures in
enough, the M is considered to indicate a fall. Fig. 7. A darker color indicates a higher value in the matrix

4 Implementation

We develop the PerFallD prototype on Android G1 phone.

It features an ARM-based, dual-core CPU capable of up to
4 million triangles/sec, a 98MB RAM and a 70MB of
internal storage [38]. It uses a 1150 mAh rechargeable
lithium ion battery. Besides camera, G1 phone provides
three sensors: an accelerometer, an orientation sensor and a
digital magnetic field sensor. In the following, we describe
the implementation details of the PerFallD prototype.
Fig. 10 User interfaces in PerFallD: (a) bright, large virtual buttons
We implement the prototype in Java, with Eclipse and on operating screen (b) clear alert window (c) simple, non-confusing
Android 1.6 SDK. It consists of 7 class files, which preference screen
includes 4 Activities, 1 View, 1 Service and 1 Resource.
They can be divided into five major components: user Jitterbug1. The user interfaces of PerFallD have the fol-
interface, monitoring daemon, data processing, alert noti- lowing features. Large, lit key buttons make usage easy.
fication and system configuration. After user starts the Bright color screen displays everything with clarity. There
system, the monitoring daemon keeps running in back- is no confusing menus, making accessing all options
ground as a Service in Android, collecting and recording clearly. Figure 10 illustrates the user interfaces of Per-
the readings of sensors. These readings are processed based FallD. Guided by friendly user interfaces, the operation is
on power-aware strategy and used to detect a fall. In data simple and straightforward. In the operating screen, three
processing component, for simplicity, all the time windows buttons are shown: Monitoring, Setting and Quit. The
are set to 4 s. When a fall is detected, the alert notification Monitoring button leads to the daemon. An alert window
component works to sound alarm to notify the attendant will prompt out once a fall is detected. The siren also
nearby and call the emergency contacts. Also, user can sounds. The Setting button leads to the preference screen of
change the configuration settings by invoking the prefer- program.
ence screen. In fall detection method with an accessory, we utilize
We compile and build the system project, create and several magnets and a bandage that are originally for
sign the .apk file in debug mode, then install it onto G1 healthcare use. These magnets’ surface Gauss rating is 350
phone by ADB tool. The size of the .apk application file is Gauss (35 mT), measured by a hand-held digital gauss-
about 216 KB. Ultimately, we may create the .apk file in meter–Model 410 [41]. According to the WHO standard,
release mode, sign it with our release private key and
publish it on Android Market, making it available to users
of Android-powered mobile devices for download. 1
Jitterbug [39] is a cell phone service provider that is known for
We implement the user interfaces of PerFallD toward providing appealing cell phone and phone service to the elderly
the elderly people, following the design ideas from people [40].

123
 Pers Ubiquit Comput

they are harmless to general public who do not use collect the ADL data from the static mannequin, either.
implanted metallic devices [42]. Using the mannequin, we collect data of 600 falls in total.
They cover all the falling direction, environment and the
position of phone attachment (as mentioned before).
5 Evaluation We further conduct the experiments with a group of real
persons. Both falls and ADL are tested. Obviously, we
We evaluate the PerFallD prototype with extensive cannot test falls with real elderly people. We recruit 15
experiments. In this section, we first introduce how the data participants who are graduate students from 20 to 30 years
are collected. Then we present performance of PerFallD old, two of whom are women. Three of them are 161–
and compare it with that of existing algorithms and a 170 cm tall, seven are 171–180 cm tall, and five are 181–
typical commercial product. We also present PerFallD’s 190 cm tall. One person weighs less than 50 kg, two weigh
resource consumption. 51–60 kg, five weigh 61–70 kg, and seven weigh 71–80 kg.
To test detection without accessories, all the participants
5.1 Data collection put the G1 phone in a shirt pocket, on the belt, or in one
pants pocket, respectively. To test detection with accesso-
We collect extensive falls data. We study falls with dif- ries, they put the phone in one pocket of the pants while
ferent directions (forward, lateral and backward), different attaching the magnet accessory to the other leg. Figure 11b
speeds (fast and slow) and in different environment (living illustrates this. In each case of phone attached position,
room, bedroom, kitchen and outdoor garden). We also every participant falls 10 times in different directions and
collect data of activities of daily living (ADL) including environment. In total, we obtain data for 600 falls that
walking, jogging, standing and sitting. We separate all cover all falling directions and environment. We also col-
these collected data into two sets, one for training and the lect ADL data for 30 min from each person.
other for testing.
We first collect falls data using a mannequin equipped 5.2 Detection performance
with PerFallD system. The mannequin is forced to fall in
different directions and in different environment. The We measure the detection performance in terms of false
PerFallD system is attached to the mannequin at one of the negative (FN) and false positive (FP). False negative hap-
three positions for fall detection without accessories: chest, pens when a fall occurs but the device misses it. False
waist, and thigh. Figure 11a illustrates the situation in positive happens when the device alarms a fall but it does
which the phone is on the mannequin’s waist. Since the not occur. In general, the lower the both FN and FP are, the
relative position of two legs on the mannequin does not better the performance is.
change during a fall, we did not conduct the experiments
for fall detection with magnetic accessories. We cannot 5.2.1 PerFallD performance

We obtain the thresholds for our detection algorithms from

the ROC curves which are generated from the training data.
ROC curves show the tradeoff between FN and FP (any
increase in FN will be accompanied by a decrease in FP). A
typical ROC curve generated in our testing is shown in
Fig. 6. With different threshold values, we can obtain
different ROC curves. From each ROC curve, we select the
threshold setting that achieves low FN and reasonable FP.
The threshold settings that are reported in the following
achieve the best balance between FN and FP based on our
training data. In detection test without accessories using the
mannequin, we set the Thtt to 150, Thtv to 4, Thct to 70 and
Thcv to 2. While in the tests with real persons, we set Thtt to
120, Thtv to 6, Thct to 50, Thcv to 2; and we set Thtm to 1970
and Thhd to 1 in detection with a magnetic accessory.
Tables 1 and 2 show the experimental results of PerFallD
Fig. 11 a The G1 phone is attached to the waist of the mannequin
with the mannequin and real persons, respectively.
(position denoted by circle A). b The participant puts the G1 phone in
his pocket of the pants (position denoted by circle A) and binds the In detection without accessories, PerFallD has different
magnetic accessory on the leg (position denoted by circle B) performances when the phone is placed at different

123
Pers Ubiquit Comput

Table 1 Performance comparison of detection test with the The basic algorithm uses the simple acceleration
mannequin threshold to determine a fall. The threshold is only based
FN(%) on the value of |AT|. The detection only focuses on one big
acceleration change (regarded as the impact of a fall),
Forward falls Lateral falls Backward falls
ignoring the following acceleration changes. So it will miss
PerFallD without accessory some slow falls and alarm falsely in some ADL. Fall Index
Chest 1.1 2.5 5.5 (FI) is proposed by Yoshida et al. in [43]. For any time i, FI
Waist 2.5 3.4 2.4 can be calculated by Eq. 5.
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Thigh 1.3 9.5 2.6 u
uX X i  2
Basic algorithm 8.2 27.9 5.8 FIi ¼ t ðAk Þi  ðAk Þi1 : ð5Þ
Fall index 4.9 14.1 1.9 k¼x;y;z i19
Commercial product 0 0 28.3
Since FI requires high sampling frequency and fast
No false positive is caused, since the ADL data are not available and
the phone is bound firmly in the test with the mannequin
acceleration changes, it will miss falls that happen slowly.
Its performance decreases in some specific situations.
Figure 12 shows the commercial product of Brickhouse
positions. We discover that the waist is the best position to together with a G1 phone. This commercial product con-
attach the phone. In this case, the average of FN values in sists of one base and a wearable fall detector [11]. The base
the experiments with the mannequin is 2.77%; in the needs to be connected with a phone line to communicate
experiments with real persons, the average FN value is with emergency center. So it has to be fixed somewhere
2.67%, while the FP value is 8.7%. inside home. Due to the constraints of communicating
In detection with a magnetic accessory, we find that it range between base and fall detector, the users must be
can achieve better accuracy. In the experiments with real indoor to be under protection. The algorithm used in this
persons, the average of FN values is 2.13% and the FP commercial product is unknown. Experiments with the
value is 7.7%. Meanwhile, the experimental results show mannequin and real persons both show that this system has
that it performs well in detecting lateral and backward falls high false negative (28.3 and 29.9%) in backward falls.
because the patterns of leg movements in such falls are Meanwhile, the false positive of detection in experiment
apparent. Our experiments also show that fall detection with real persons is also quite high (21.9%).
method with a magnetic accessory works well especially to The results show that no matter testing with the man-
those slow falls, which are common to elderly people. nequin or real persons, PerFallD outperforms existing
algorithms, and achieves better balance between false
5.2.2 Performance comparison negative and false positive compared with the commercial
product.
We compare the performance of PerFallD with two other
existing detection algorithms and one commercial product. 5.2.3 Environment impact
Table 1 shows the results of experiments conducted with
the mannequin, while Table 2 illustrates the experiments We test PerFallD in different locations (in living room,
on real persons. bedroom, kitchen and outdoor garden). In all these

Table 2 Performance comparison of detection test with real persons

FN(%) FP(%)
Forward falls Lateral falls Backward falls Other activities

PerFallD without accessory

Chest 1.2 2.3 5.0 11.2
Waist 2.6 3.3 2.1 8.7
Thigh 1.0 10.0 2.2 11.0
PerFallD with accessory 3.1 2.2 1.1 7.7
Basic algorithm 8.0 28.3 5.5 14.6
Fall index 5.2 13.9 1.8 7.8
Commercial product 0.8 1.2 29.9 21.9
We call the ADL and movements of phone that may cause the false positive other activities

123
 Pers Ubiquit Comput

100
90
80

Battery Level(%)
70
60
50
40
30
Phone runs without PerFallD
20 Monitoring daemon runs
10 Pattern recognition runs regularly
0
0 1 2 3 4 5 6
Time(H)

Fig. 12 The commercial product of fall detection (white) compared Fig. 14 Power consumption: first curve presents the battery levels
with Android G1 Phone when the phone runs without PerFallD; the second one shows the
battery levels when the monitoring daemon in PerFallD runs; the third
one presents the battery levels when the pattern recognition in
8000 8000 PerFallD runs regularly

7000 7000 total RAM capacity of G1 phone. During the process of fall
6000 6000
pattern recognition, the CPU usage is 8.15%; the usage of
memory increases by about 300KB.
5000 5000

4000 4000
0 1 2 3 4 0 1 2 3 4 5 6 6 Conclusion
Time(s) Time(s)

Fig. 13 (left) The magnetic sensor readings during a fall in kitchen In this paper, we propose utilizing mobile phones as a
are similar to those in other places. (right) The sensor readings when platform for developing pervasive fall detection system, for
passing by an oven in kitchen are clearly different from those in a fall the first time. We design different detection algorithms
based on mobile phone platforms for scenarios with and
without simple accessories. We implement a prototype
locations, PerFallD works well. Especially, the detection system named PerFallD on the Android G1 phone and
with the magnetic accessory is not interrupted badly, even conduct extensive experiments to evaluate our system.
in the kitchen with many ferric stuffs. Figure 13 shows Experimental results show that PerFallD achieves good
some typical readings of magnetic sensor in kitchen. detection performance and power efficiency.
In future, we may further enhance the current system by
5.3 Resource consumption performance integrating extra protection devices, e.g., airbag-based fall
protector proposed by Charpentier [44] and Fukaya [45], to
To test power consumption, we fully charge the G1 phone reduce fall impacts and prevent fall-related injuries.
and then monitor the power states continuously for 6 h in
Acknowledgments This work was supported in part by the US
different scenarios: (1) the G1 phone runs without Per-
National Science Foundation (NSF) under grants No. CNS-0916584,
FallD; (2) the monitoring daemon in PerFallD keeps run- CAREER Award CCF-0546668, and the Army Research Office
ning, sensing and recording sensor reading values, without (ARO) under grant No. AMSRD-ACC-R50521-CI. Any opinions,
pattern recognition; (3) PerFallD not only monitors, but findings, conclusions, and recommendations in this paper are those of
the authors and do not necessarily reflect the views of the funding
also calculates and recognizes fall pattern on the demand of
agencies.
monitoring results. Figure 14 presents the three curves of
battery level states versus time during the time period of
6 h. If PerFallD keeps running normally until the battery References
power is exhausted, it will last about 33.5 h.
Furthermore, we monitor the CPU and memory usage of 1. Stevens J (2003) Falls among older adults: moving from research
G1 phone during the running of monitoring daemon and to practice. In: Proceedings of international conference on aging,
disability and independence, 2003
after the fall pattern recognition is invoked. When only the
2. Fulks JS, Fallon F, King W, Shields G, Beaumont N, Ward-
monitoring daemon is running, the average CPU usage is Lonergan J (2002) Accidents and falls in later life. Generations
6.68%; the memory usage is about 350KB, 0.35% of the Rev 12(3):2–3

123
Pers Ubiquit Comput

3. Pynoos J, Sabata D (2003) Home environmental modification: 25. Karantonis DM, Narayanan MR, Mathie M, Lovell NH, Celler
preventing falls at home. In: Proceedings of international con- BG (2006) Implementation of a real-time human movement
ference on aging, disability and independence, 2003 classifier using a triaxial accelerometer for ambulatory monitor-
4. Duthie E (1989) Falls. Med Clin North Am 73:1321–1335 ing. IEEE Trans Inf Technol Biomed 10(1):156–167
5. Hornbrook MC, Stevens VJ, Wingfield DJ, Hollis JF, Greenlick 26. Fu Z, Culurciello E, Lichtsteiner P, Delbruck T (2008) Fall
MR, Ory MG (1994) Preventing falls among community-dwell- detection using an address-event temporal contrast vision sensor.
ing older persons: results from a randomized trial. Gerontologist In: Proceedings of IEEE international symposium on circuits and
34(1):16–23 systems, ISCAS 2008
6. Tideiksaar R (1998) Falling in old age: prevention and manage- 27. Sixsmith A, Johnson N (2004) A smart sensor to detect the falls
ment, 2nd edn. Springer, Berlin of the elderly. Pervasive Comput 3(2):42–47
7. Sadigh S, Reimers A, Andersson R, Laflamme L (2004) Falls and 28. Miaou S, Sung P, Huang C (2006) A customized human fall
fallrelated injuries among the elderly: a survey of residential-care detection system using omni-camera images and personal infor-
facilities in a Swedish municipality. J Community Health mation. In: Proceedings of 1st distributed diagnosis and home
29(2):129–140 healthcare conference, 2006
8. Ryynanen OP, Kivela SL, Honkanen R, Laippala P (1992) Falls 29. Jansen B, Deklerck R (2006) Context aware inactivity recogni-
and lying helpless in the elderly. Z Gerontol 25(4):278–282 tion for visual fall detection. In: Proceedings of pervasive health
9. Kiel DP (1991) Falls. R I Med J 74(2):75–79 conference and workshops, 2006
10. U.S. Census Bureau (2009) Population data public web site. 30. Hansen TR, Eklund JM, Sprinkle J, Bajcsy R, Sastry S (2005)
http://www.census.gov/population/www. Accessed 29 Sept 2009 Using smart sensors and a camera phone to detect and verify the
11. http://www.brickhousealert.com/personal-emergency-medical- fall of elderly persons. In: Proceedings of European medicine,
alarm.htm. Accessed 2 October 2009 biology and engineering conference, EMBEC 2005
12. Resende P (2008) Smartphone competition heats up with many 31. Zhang T, Wang J, Liu P, Hou J (2006) Fall detection by
choices. NewsFactor Network Sept. 2008. http://www.news embedding an accelerometer in cellphone and using KFD algo-
factor.com/story.xhtml?story_id=6167. Accessed 2 October 2009 rithm. Int J Comput Sci Netw Secur 6(10):277–284
13. Boran M (2009) How the mobile social web will make 2009 a 32. Chen J, Kwong K, Chang D, Luk J, Bajcsy R (2005) Wearable
smart phone odyssey. http://www.independent.ie/business/ sensors for reliable fall detection. In: Proceedings of 27th annual
technology/how-the-mobile-social-web-will-make-2009-a-smart- international conference of the engineering in medicine and
phone-odyssey-1609944.html. Accessed 2 October 2009 biology society, IEEE-EMBS 2005
14. Mies G (2009) Six new phones from AT&T: have a look. PC 33. http://www.cnn.com/2010/TECH/03/03/ishoe.mit.award/index.
World, 30 Mar 2009. http://www.pcworld.com/article/162233/ html. Accessed 3 March 2010
six_new_phones_from_atandt_have_a_look.htm. Accessed 2 34. Kumar S (1999) Biomechanics in ergonomics. CRC Press, Boca
October 2009 Raton
15. New smartphone family wireless plan from Cincinnati Bell. 35. Sacher A (1996) The application of forensic biomechanics to the
http://www.wirelessplansinformation.blogspot.com/2008/10/new- resolution of unwitnessed falling accidents. J Forensic Sci
smartphone-family-wireless-plan.html. Accessed 26 Sept 2009 41(5):776–781
16. http://www.ittmonaco.com/en/easy/lifeS.ht. Accessed 20 Dec 36. Belongie S, Malik J and Puzicha J (2002) Shape matching and
2009 object recognition using shape context. IEEE Trans Pattern Anal
17. http://www.lifelinesys.com/index.jsp?campaign=10. Accessed 20 Mach Intell 24(24):509–522
Dec 2009 37. Ding L, Belkin M (2008) Component based shape retrieval using
18. http://www.betterbuys-r-us.com/HomeCareSale.ht. Accessed 20 differential profiles. In: Proceedings of ACM international con-
Dec 2009 ference on multimedia information retrieval
19. Nyan MN, Tay FE, Manimaran M, Seah KH (2006) Garment- 38. http://en.wikipedia.org/wiki/HTC_Dream. Accessed 26 Sept
based detection of falls and activities of daily living using 3-axis 2009
MEMS accelerometer. J Phys Conf Ser 34:1059–1067 39. http://www.jitterbug.com. Accessed 30 Sept 2009
20. Kangas M, Konttila A, Lindgren P, Winblad I, Jamsa T (2008) 40. http://www.shopaservice.com/articles/jitterbug-cell-phones.
Comparison of low-complexity fall detection algorithms for body Accessed 30 Sept 2009
attached accelerometers. Gait Posture 28(2):285–291 41. http://www.lakeshore.com/mag/ga/gm410po.htm. Accessed 23
21. Brown G (2005) An accelerometer based fall detector: develop- July 2009
ment, experimentation, and analysis. SUPERB internship report, 42. World Health Organization (2009) Electromagnetic fields and
University of California, Berkeley public health: static electric and magnetic fields. http://www.
22. Hwang JY, Kang JM, Jang YW, Kim HC (2004) Development of who.int/mediacentre/factsheets/fs299/en/print.html. Accessed 26
novel algorithm and real-time monitoring ambulatory system July 2009
using Bluetooth module for fall detection in the elderly. In: 43. Yoshida T, Mizuno F, Hayasaka T, Tsubota K, Wada S, Yam-
Proceedings of 26th annual international conference of the aguchi T (2005) A wearable computer system for a detection and
engineering in medicine and biology society, IEEE-EMBS 2004 prevention of elderly users from falling. In: Proceedings of the
23. http://mover.projects.fraunhofer.pt/. Accessed 3 March 2010 12th international conference on biomedical engineering, 2005
24. Ganti RK, Jayachandran P, Abdelzaher TF, Stankovic JA (2006) 44. Charpentier PJ (1996) A hip protector based on airbag technol-
SATIRE: a software architecture for smart AtTIRE. In: Pro- ogy. Bone 18(1):S117
ceedings of the 4th international conference on mobile systems, 45. Fukaya K (2002) Fall detection sensor for fall protection airbag.
applications and services, MobiSys, 2006 In: Proceedings of the 41st SICE annual conference

123

View publication stats