You are on page 1of 15

sensors

Article
Highly Portable, Sensor-Based System for Human
Fall Monitoring
Aihua Mao 1, *, Xuedong Ma 2 , Yinan He 1 and Jie Luo 3, *
1 School of Computer Science & Engineering, South China University of Technology,
Guangzhou 510006, China; heyinanscut@outlook.com
2 School of Electronic and Information Engineering, South China University of Technology,
Guangzhou 510006, China; eemaxuedong@mail.scut.edu.cn
3 School of Fine Art and Artistic Design, Guangzhou University, Guangzhou 510006, China
* Correspondence: ahmao@scut.edu.cn (A.M.); jieluo@gzhu.edu.cn (J.L.); Tel.: +86-20-39380285 (A.M.)

Received: 12 August 2017; Accepted: 11 September 2017; Published: 13 September 2017

Abstract: Falls are a very dangerous situation especially among elderly people, because they may
lead to fractures, concussion, and other injuries. Without timely rescue, falls may even endanger
their lives. The existing optical sensor-based fall monitoring systems have some disadvantages,
such as limited monitoring range and inconvenience to carry for users. Furthermore, the fall detection
system based only on an accelerometer often mistakenly determines some activities of daily living
(ADL) as falls, leading to low accuracy in fall detection. We propose a human fall monitoring system
consisting of a highly portable sensor unit including a triaxis accelerometer, a triaxis gyroscope,
and a triaxis magnetometer, and a mobile phone. With the data from these sensors, we obtain the
acceleration and Euler angle (yaw, pitch, and roll), which represents the orientation of the user’s
body. Then, a proposed fall detection algorithm was used to detect falls based on the acceleration and
Euler angle. With this monitoring system, we design a series of simulated falls and ADL and conduct
the experiment by placing the sensors on the shoulder, waist, and foot of the subjects. Through the
experiment, we re-identify the threshold of acceleration for accurate fall detection and verify the
best body location to place the sensors by comparing the detection performance on different body
segments. We also compared this monitoring system with other similar works and found that better
fall detection accuracy and portability can be achieved by our system.

Keywords: fall detection; sensors; easy portability; body segment

1. Introduction
The aging society is becoming a serious public health issue in many areas, especially in developed
countries. According to the World Health Organization, falls are the second leading cause of accidental
or unintentional injury deaths worldwide [1], and more than one third of elderly people fall once or
more each year [2]. For this group of people, once they fall, they may suffer serious health problems.
Damage may be greatly reduced if they have access to timely rescue. Thus, a reliable fall monitoring
system has great application value and development prospects.
In the last decade, many methods have been developed for fall detection; most of which are based
on optical sensors [3–6], accelerometers [7–11], and accelerometers combined with gyroscopes [12–14].
The principle of the optical sensor-based method is to capture the image by visual sensors, such as
digital cameras and Kinect, and then distinguishes the human body with other items using digital
image processing algorithms to detect falls. For example, Yang et al. [15] used a Kinect sensor to
capture three-dimensional (3D) depth images and the dense spatio-temporal context (STC). The dense
STC algorithm was applied to track the head position. They calculated the distance between the
head and floor plane. If the distance is lower than an adaptive threshold, then the centroid height of

Sensors 2017, 17, 2096; doi:10.3390/s17092096 www.mdpi.com/journal/sensors


Sensors 2017, 17, 2096 2 of 15

the human was used to determine whether a fall accident occurred. Gasparrini et al. [16] proposed
a fall detection method for indoor environments based on a Kinect sensor. In this method, an Ad hoc
segmentation algorithm was used to analyze the raw depth data from the Kinect and recognize all
the elements captured in the depth scene. The system extracts the elements and classifies blobs in
the scene to recognize human subjects among the blobs, followed by a tracking algorithm between
different frames. Lee et al. [17] used shape feature variation extracted from the captured image to
detect abnormal events, such as fall, and utilized a shadow removal algorithm to improve the accuracy
of object detection and tracking. Anderson et al. [18] recognized the silhouette of individuals from
the image captured by cameras and then extracted the features and trained hidden Markov models to
recognize future performances of these known activities, such as fall. Rougher et al. [19] used a single
camera to obtain images and obtained the 3D trajectory of the human head and then distinguished fall
from activities of daily living (ADL) by the trajectory. Methods based on optical sensors have high
reliability and accuracy. However, their monitoring range is limited. Once the users leave the captured
range of the picture or are blocked by other objects, falls cannot be detected. Furthermore, intensive
computation is required to finish the detection, leading to difficulty in real-time monitoring.
Accelerometer-based methods measure acceleration of the human body by an accelerometer,
and falls are detected based on a threshold value of acceleration. For example, Mario et al. [20] proposed
an algorithm to detect human basic movements, such as fall events from wearable measured acceleration
data. The algorithm was designed to minimize computational requirements while achieving acceptable
accuracy levels based on characterizing some particular points in the temporal series obtained from
a single sensor. Hsieh et al. [21] reported a novel hierarchical fall detection algorithm based on the
measured acceleration data involving threshold-based and knowledge-based approaches to detect a fall
event. Kurniawan et al. [22] used a triaxis accelerometer to obtain the acceleration of the human body
and then determined whether the fall occurs depending on the acceleration. Albert et al. [23] used the
accelerometer of the mobile phone to obtain the acceleration data, which is applied in the machine
learning classifier to detect falls and classify the type of falls. These methods are not limited by the
monitoring space and are not vulnerable to external interference. However, they used only acceleration
data, which often mistakenly determine ADL, such as jumping and running, as a fall event and then
trigger a false fall alarm. Thus, the accuracy of this kind of method may be affected if it is only based on
an accelerometer. Thus, to improve the accuracy, an alternative approach is to combine a gyroscope and
even a magnetometer with the accelerometer to work together. Pierleoni et al. [24] fixed an inertial unit
on the waist of a human body by a belt to obtain the acceleration and the attitude angle and then detect
the falls. However, they can only send the detection result to a nearby device via Bluetooth. If nobody is
nearby, the fall alarm would not be noticed by other people. Wibisono et al. [25] used an accelerometer
and a gyroscope embedded in a mobile phone to obtain the acceleration and angular velocity of the
human body. A threshold-based algorithm was applied to detect falls based on the acceleration and
angular velocity. However, this kind of method based on the mobile phone required the mobile phone
provided with acceleration and gyroscope, which is not available in some mobile phones. Besides that,
the weight of most mobile phones is far greater than that of the sensor devices and may lead to discomfort
when carried for prolonged periods. Furthermore, though the mobile phone now is a necessity in daily
life, it is not practical to fix it on a body location all the time. He et al. [26] introduced a fall detection and
alerting system consisting of a smart phone and a customized vest integrated with a triaxial accelerometer
and a gyroscope. The raw triaxial gyroscope and accelerometer data are preprocessed by a program
running on the smart phone by the Kalman filter to reduce the noise measurement from the sensors.
Then, a Bayes network classifier was used to determine whether the individual falls or not. The system
only considers four typical subcategories of ADLs and two kinds of falls and only tried the sensors
on the vest near to the neck. Ntalampiras et al. [27] proposed a learning mechanism based on hidden
Markov models for automatic recognition of physical activities of humans and assessed the performance
based on previous datasets from an accelerometer and a gyroscope. However, the performance of this
kind of method based on machine learning is dependent on the training dataset.
Sensors 2017, 17, 2096 3 of 15
Sensors 2017, 17, 2096 3 of 15

In this datasets
previous study, we frompropose a highly portable,
an accelerometer sensor-based
and a gyroscope. However, system for monitoring
the performance of thishumankind falls.
of
A small
method self-packaged
based on machinesensorlearning
unit andisadependent
mobile phone on thecomposed
training this system. The sensor unit includes
dataset.
a triaxisInaccelerometer,
this study, weapropose a highly portable,
triaxis gyroscope, sensor-based
and a triaxis system for
magnetometer, monitoring
which
p can be human
easily portedfalls. Aby
thesmall
user self-packaged
to obtain the root sensormeanunitsquare
and a mobile
(RMS)phone composed(RMS
of acceleration this system.
= ( Ax The2 + Ay2 +
sensor unit 2 ), A , A ,
Azincludes x y
anda Atriaxis
z are accelerometer,
the x-axis, a
y-axis, triaxis
and gyroscope,
z-axis and
acceleration) a triaxis
and magnetometer,
orientation of thewhich
user. can
All be
the easily
sensor ported
data are
sentbyto the
theuser to obtain
mobile phone the via
rootBluetooth
mean square (RMS) of acceleration
communication ( RMS  processed
and is further Az 2a) ,fall
( Ax 2  Ay 2 by Ax, detection
Ay, and
algorithm.
Az are the If falls
x-axis,arey-axis,
detectedandand recognized,
z-axis the mobile
acceleration) phone will
and orientation issue
of the an All
user. alarmtheandsensor automatically
data are
sent to the mobile phone via Bluetooth communication and is further
call the emergency contact for timely rescue. To determine the threshold of acceleration for accurate processed by a fall detection
fallalgorithm.
detection and If falls
verifyarethedetected
best body andlocation
recognized,
to place thethemobile
sensors, phone will issue
we designed an alarm
seven and
fall activities
and automatically
nine ADLs and callconduct
the emergency contact for
the fall detection timely rescue.
experiment by placing To the
determine
sensor unit the on threshold of
the subjects’
acceleration for accurate fall detection and verify the best body location
shoulder, waist, and foot, respectively. We also compare this system with other similar works in terms to place the sensors, we
designed portability,
of accuracy, seven fall activities and nine
and comfort. OurADLssystem andhas conduct the fall detection
good accuracy experiment
and is highly portable bydue
placing
to the
the sensor unit on the subjects’ shoulder, waist, and foot, respectively.
components of a small sensor unit and a mobile phone. Furthermore, our system is comfortable We also compare this system
to the
with other similar works
user even if used for a long time. in terms of accuracy, portability, and comfort. Our system has good
accuracy and is highly portable due to the components of a small sensor unit and a mobile phone.
The main contributions of this study include the following:
Furthermore, our system is comfortable to the user even if used for a long time.
1. WeThe mainacontributions
design highly portable of this
fallstudy include system,
monitoring the following:
which only consists a small self-packaged
sensor unit and a mobile phone to detect
1. We design a highly portable fall monitoring system, which falls, and a fall alarm onlycanconsists
be automatically issued to the
a small self-packaged
emergency
sensor unit contact
and a to seek timely
mobile phone to rescue.
detectFurthermore,
falls, and a fall this system
alarm can has great detection
be automatically accuracy
issued to
forthe
both falling and
emergency ADL.to seek timely rescue. Furthermore, this system has great detection
contact
2. Weaccuracy
re-identify an acceleration
for both falling and ADL. threshold of 2.3 g by experimental observation for accurate fall
2. detection,
We re-identify
which an acceleration
is effective threshold
to avoid of 2.3 g ADL
determining by experimental
as falls, andobservation
thus improving for accurate fall
the accuracy
ofdetection, whichBenefitting
fall detection. is effectivefrom to avoid
this newdetermining
threshold,ADL as falls,can
our system andachieve
thus improving
better accuracy the
accuracy of fall detection.
compared with other similar works. Benefitting from this new threshold, our system can achieve better
3. Weaccuracy
verify thecompared with other
best location similar
of the humanworks.body to wear the sensor by conducting a series of
3. experiments
We verify the best location of the
and comparison of the performancehuman body of to fall
wear the sensor
detection whenby the
conducting
sensor unit a series of
is placed
experiments and comparison of the performance of fall detection when the sensor unit is placed
on different body segments. The waist is the best one for sensor placement where the detection
on different body segments. The waist is the best one for sensor placement where the detection
accuracy is 100% without any misjudgment, whereas wearing the sensor in the shoulder and foot
accuracy is 100% without any misjudgment, whereas wearing the sensor in the shoulder and
may still mistakenly identify ADL as falls.
foot may still mistakenly identify ADL as falls.
2. Fall Monitoring System
2. Fall Monitoring System
The system consists of a small self-packaged sensor unit and a mobile phone with an Android
The system consists of a small self-packaged sensor unit and a mobile phone with an Android
operating system (Figure 1). The sensor unit is easily carried by the user, containing a microelectrometrical
operating system (Figure 1). The sensor unit is easily carried by the user, containing a
system (MEMS) sensor, a kernel processer, and a Bluetooth module. The sensors’ data are transmitted
microelectrometrical system (MEMS) sensor, a kernel processer, and a Bluetooth module. The
to the Android
sensors’ mobile
data are phone via
transmitted Bluetooth.
to the Android A developed
mobile phone application (APP)
via Bluetooth. runs on application
A developed the Android
system
(APP) runs on the Android system to receive sensor data from the sensor device and processes thethe
to receive sensor data from the sensor device and processes the raw data to calculate
orientation
raw data and acceleration.
to calculate Based onand
the orientation the acceleration.
orientation and BasedRMS, the orientation
on the APP automatically
and RMS,detects falls
the APP
based on a proposed
automatically fallfalls
detects detection
based algorithm. Once
on a proposed the
fall fall is detected,
detection an Once
algorithm. alarmthe
is issued by the mobile
fall is detected, an
phone
alarmand a call will
is issued automatically
by the mobile phone connect to the
and a call willemergency contact.
automatically connect to the emergency contact.

Figure
Figure 1. 1. Highlyportable,
Highly portable,sensor-based
sensor-based fall
fall monitoring
monitoringsystem.
system.MEMS:
MEMS:microelectrometrical system.
microelectrometrical system.
Sensors 2017, 17, 2096 4 of 15

Sensors 2017, 17, 2096 4 of 15


Sensors
Figure2017,
2 17, 2096
shows the internal details of the self-packaged sensor unit, wherein the MEMS 4 ofsensor
15

module Figure 2 shows


integrates the internal
a triaxis details ofat
accelerometer the self-packaged
the range of ±16 sensor
g, a unit, wherein
triaxis the MEMS
gyroscope at thesensor
range of
◦ Figure 2 shows the internal details of the self-packaged sensor unit, wherein the MEMS sensor
±2000 , and a triaxis magnetometer. The Kernel processor can reduce the measurement noise of
module integrates a triaxis accelerometer at the range of ±16 g, a triaxis gyroscope at the range from
module integrates a triaxis accelerometer at the range of ±16 g, a triaxis gyroscope at the range of
±2000°,
the ±2000°, and
by aaafiltering
sensor and triaxis magnetometer.
algorithm The in
setup Kernel
the processor can
processor. The reduce the measurement
Bluetooth module noise afrom
includes BC417
triaxis magnetometer. The Kernel processor can reduce the measurement noise from
the sensor
Bluetooth bybased
chipby a filtering algorithm setup
on Bluetooth inthethe processor. The Bluetooth module includes a sampling
BC417
the sensor a filtering algorithm2.0, andin
setup therange of theThe
processor. Bluetooth
Bluetooth module
module is includes
15 m. Thea BC417
Bluetooth chip based on Bluetooth 2.0, and the range of the Bluetooth module is 15 m. The sampling
frequency
Bluetoothof the
chipsensors
based onfrom the human
Bluetooth activities
2.0, and the rangeis set
of to
the100 Hz. After
Bluetooth noiseisreduction,
module 15 m. The the sampling
sampling
frequency of the sensors from the human activities is set to 100 Hz. After noise reduction, the
datafrequency
from the of sensors are transmitted to the mobile phone via Bluetooth communication.
the sensors from the human activities is set to 100 Hz. After noise reduction, the The power
sampling data from the sensors are transmitted to the mobile phone via Bluetooth communication.
supply module
sampling contains
data from thea power
sensorssupply chip and to
are transmitted a battery.
the mobileWith the power
phone supply communication.
via Bluetooth module, the device
The power supply module contains a power supply chip and a battery. With the power supply
canThe
work power
withoutsupplyany module
additionalcontains
power aafter
power supply the
charging chip and a Thus,
battery.
module, the device can work without any additional power after charging the battery.
battery.
theWith
devicetheispower supply
veryThus,
convenient
the
module,
to be carried the
by device
the can work without any additional power after charging the battery. Thus, the
users.
device is very convenient to be carried by the users.
device is very convenient to be carried by the users.

Figure2.2.Small,
Figure Small, self-packaged
self-packaged sensor
sensorunit.
unit.
Figure 2. Small, self-packaged sensor unit.
Figure
Figure 3 shows
3 shows thethemain
maininterfaces
interfaces of
of the
the Android
Android APP.
APP. First,
First,the
theAPP
APPsearches
searches forfor
nearby
nearby
Figure 3 shows the main interfaces of the Android APP. First, the APP searches for nearby
Bluetooth
Bluetooth devices and connects to the sensor unit. After establishing the connection and receiving
Bluetooth devices and connects to the sensor unit. After establishing the connection and receiving the
devices and connects to the sensor unit. After establishing the connection and receiving
the sensorthedata,
sensor the APP calculates the Euler angle and then visualizes the Euler angle and
the data, APPthe
sensor data, calculates the Euler
APP calculates theangle
Euler and thenand
angle visualizes the Eulerthe
then visualizes angle andangle
Euler acceleration
and
acceleration by data charts. Meanwhile, a fall detection algorithm was used to detect falls in real time.
by data charts.by
acceleration Meanwhile,
data charts.aMeanwhile,
fall detection algorithm
a fall detection was used was
algorithm to detect falls
used to in real
detect fallstime.
in realIftime.
a fall is
If a fall is detected, the APP will make an alarm sound and issue a call to the emergency contact.
detected,
If a fallthe APP will
is detected, themake
APP an
willalarm
make sound
an alarm and issue
sound a call
and toathe
issue callemergency contact.
to the emergency contact.

Figure 3. Main interfaces of the Android application: (a) device search; (b) data visualization and
Figure 3. Main interfaces of the Android application: (a) device search; (b) data visualization and
fall monitor;
Figure (c) makingof
3. Main interfaces alarm and automatic
the Android calling. (a) device search; (b) data visualization and fall
application:
fall monitor; (c) making alarm and automatic calling.
monitor; (c) making alarm and automatic calling.
Sensors 2017, 17, 2096 5 of 15
Sensors 2017, 17, 2096 5 of 15

3. 3D Orientation Measurement of the Human Body


To
To describe
describe the posture of the human body, we use the Euler angle, which is described described by by yaw,
yaw,
pitch, and roll angles to represent the body’s spatial orientation
angles to represent the body’s spatial orientation as shown in Figure
shown in Figure 4. 4. In the ideal
case, aa single
single3-axis
3-axisgyroscope
gyroscopeplaced
placedonon a rigid
a rigid body
body cancan provide
provide the current
the current orientation
orientation of theofbody
the
body by performing
by performing integration
integration calculation
calculation on gyroscope
on gyroscope data.data. However,
However, in practice,
in practice, integration
integration will
will accumulate
accumulate the white
the white noisenoise
fromfrom the reading
the reading effects,
effects, resulting
resulting in drift
in drift fromfrom
the the expected
expected values.
values. To
To complement
complement thethe drift
drift gyroscope
gyroscope data,sophisticated
data, sophisticatedalgorithms
algorithmsoften
oftenfuse
fuse the
the accelerometer
accelerometer and
magnetometer
magnetometer data datadue
dueto totheir
theircompensable
compensablecharacteristics.
characteristics.The
Theperformance
performance ofof
sensor
sensordata
datafusion is
fusion
often determined by the corrector input for the accelerometer and magnetometer.
is often determined by the corrector input for the accelerometer and magnetometer.

Figure
Figure 4.
4. Pitch,
Pitch, roll,
roll, and
and yaw
yaw angle
angle of the human
of the human body.
body.

As shown in Figure 5, the Euler angle of a human segment is estimated based on the gyroscope
As shown in Figure 5, the Euler angle of a human segment is estimated based on the gyroscope
data, accelerometer data, and magnetometer data. The estimation can be achieved based on Yean’s
data, accelerometer data, and magnetometer data. The estimation can be achieved based on Yean’s
model [28]. A Kalman filter is designed to fuse the sensor data mainly through two stages, namely,
model [28]. A Kalman filter is designed to fuse the sensor data mainly through two stages, namely,
applying the filter to the gyroscope data to predict the variables describing the next state and then
applying the filter to the gyroscope data to predict the variables describing the next state and then
estimating the orientation by using the accelerometer and magnetometer data to correct the
estimating the orientation by using the accelerometer and magnetometer data to correct the prediction
prediction from the gyroscope. In the estimation stage, extracting gravity from the accelerometer
from the gyroscope. In the estimation stage, extracting gravity from the accelerometer doesn’t suffer
doesn’t suffer from the drift issue, meanwhile, the data from the magnetometer measures the
from the drift issue, meanwhile, the data from the magnetometer measures the calibrated magnetic
calibrated magnetic field values for all three physical axes (x, y, z), especially in the circumstance of
field values for acceleration
the dynamic all three physicalandaxes (x, y, z),movement.
direction especially inThese
the circumstance
facts enable of the
thedynamic acceleration
correction for the
and
predicted state. The output of the Kalman filter is further used for quaternion estimation tooutput
direction movement. These facts enable the correction for the predicted state. The of
calculate
the Kalman
the Euler filter
angle of is further segment
a human used for and
quaternion estimation
the corrector to calculate
input by the Euler and
the magnetometer angle of a human
accelerometer
segment
data. Here, quaternion was selected because of its lower computation cost. The quaternion was
and the corrector input by the magnetometer and accelerometer data. Here, quaternion q is
selected
expressed because
as (qw, ofqx, its
qy, lower
and qzcomputation cost. The quaternion
), where qw represents its scalar, andq isqexpressed as (qw , qx , qy , and qz ),
x, qy, and qz represent its vector.

Based qon
where w represents
the estimated its scalar, and qx ,calculating
quaternion, qy , and qz represent
the Euleritsangles
vector.ofBased on thesegment
the human estimated isquaternion,
easy using
calculating the Euler
the following equation: angles of the human segment is easy using the following equation:

Yaw  atan2(2qxx qyy  2qww qzz , 2qww  2qxx  1)



 Yaw = atan2(2q q − 2q q , 2q2 2 + 2q22− 1)
Pitch = arcsin(2qw qy − 2qx qz )

(1)
  Pitch  arcsin(2qw q y  2qx qz ) (1)
Roll = atan2(2qy qz − 2qw qx , 2q2w2+ 2q2z 2− 1)


 Roll  atan2(2q y qz  2qw qx , 2qw  2qz  1)
Sensors 2017, 17, 2096 6 of 15
Sensors 2017, 17, 2096 6 of 15
Sensors 2017, 17, 2096 6 of 15

Gyroscope Magnetometer Accelerometer


Gyroscope Magnetometer Accelerometer

Kalman Filter Corrector input


Kalman Filter Corrector input

Quaternion estimation
Quaternion estimation

Body segment Euler angle estimation


Body segment Euler angle estimation

Figure 5. Estimation of the Euler


5.5.Estimation angle of aofhuman segment.
Figure
Figure Estimationofofthe
theEuler
Eulerangle
angle a human
of segment.
a human segment.

4. Fall Detection
4.4.Fall
Fall Algorithm
Detection
DetectionAlgorithm
Algorithm
During
During
Duringa fall activity,
a afall
fall thethe
activity,
activity, human
the human
human body body
bodyfirst loses
first
firstloses balance
loses balance
balance and and falls
and fallsin in
falls aincertain
a acertain direction
certain direction
direction
weightlessly,
weightlessly, then impacts against the low objects, such as the ground.
weightlessly, then impacts against the low objects, such as the ground. In this process, thehuman
then impacts against the low objects, such as the ground. In this
In process,
this process, the thehuman human
posture
posture generally
posture generally
generallychanges
changes from
changes from standing
from standing
standing to to lying
tolying or or
lying nearly
ornearly flat.
nearly Thus,
flat.
flat. Thus, thethe
Thus, process
the process
processof of
falls
offallscancan
falls bebebe
can
divided
divided into
dividedinto three
intothreephases
threephases as follows:
phasesasasfollows: (1)
follows:(1) start,
(1)start,the
start,the human
thehuman body
humanbody lost
bodylost balance
lostbalance and
balanceand descents
anddescents toward
descentstoward the
towardthe the
ground;
ground; (2) impact,
(2) impact, thethehuman
human body
body falls
falls totothethe ground
ground and
and comes
comes
ground; (2) impact, the human body falls to the ground and comes in contact with the ground; and in
in contact
contact with
with the
the ground;
ground; and
and (3)
(3)posture,
posture, thehuman
the
(3) posture, human
the human body
body remains
remains
body lying
lying
remains ononthe
lying the ground
onground foratat
for
the ground least
least
for ashort
at aleast short period
period
a short ofoftime.
period time.
of time.
During
During these
Duringthese three
thesethree phases,
threephases, the
phases,the supporting
thesupporting
supportingforceforce for the
forceforforthehuman
thehuman body
humanbody from
bodyfrom the
fromthe ground
theground
ground changes
changes
changes
sharply.
sharply. Thus,
sharply. Thus, the acceleration of the
accelerationofofthe
Thus, the acceleration body
thebody also
body changes
alsoalso sharply.
changes
changes Given
sharply.
sharply. that
Given Giventhe acceleration
that that sensor
the acceleration
the acceleration sensor
is sensor
triaxial, isthe
is triaxial, acceleration
triaxial,
the acceleration of the
the acceleration human
of the of body
human the body (RMS)
human can can
body
(RMS) be(RMS)
calculated can be
be calculated from triaxial
calculated
from acceleration
from
triaxial triaxial
acceleration
components.
acceleration
components. components.
The triaxis accelerometer measures AA x, Ay, and Az based on forces acting on the accelerometer
The
Thetriaxis
triaxis accelerometer
accelerometer measures
measures x ,AA , and
x,y A y, andAz Abased
z based on onforces acting
forces on the
acting on accelerometer
the accelerometer in
in the
the three
in three axes.
axes.axes.
the three Thus,
Thus,Thus, RMS
RMS RMS will reach
will reach 1 g even
1 g even
will reach 1 gifeven if the
the if device
device
the is is stationary
stationary
device due due
is stationary to the toinfluence
due the
to influence of of
of gravity.
the influence
gravity. As shown
Asgravity.
shown in Figure
As shownin Figure
6,inRMS 6,during
Figure RMS
6, RMSduring
the start
during the start
phase
the is phase
nearly
start isclose
phase nearly to 0close
is nearly g and tochanges
close 0 to
g and changes
rapidly
0 g and torapidly
changes therapidly
peak
to due
the
to topeak
thethe due
peak to
supporting the
due to the supporting
force force
from the ground
supporting from
force from the
during ground
the the
ground during
impact the
phase.the
during impact
Then,
impactphase.
RMSphase. Then,
remains RMS
Then,almost
RMS
remains
stable almost
remainswithalmoststable
a slight withwith
a slight
fluctuation
stable fluctuation
during
a slight the posture
fluctuation during the posture
phase.
during the posture phase. phase.
Impact
Impact

Start Posture
Start Posture

Figure 6. Root mean square (RMS) during a fall activity.


Figure 6. Root mean square (RMS) during a fall activity.
Figure 6. Root mean square (RMS) during a fall activity.
Given thatthat
Given the start phase
the start is not
phase is the
not evident
the evidentdistinction between
distinction fallsfalls
between and and
other activities,
other suchsuch
activities,
Given
as diving, the that
fall the start
detection phase is not
algorithm the
is evident
based on distinction
the between
recognition of falls
impact and other
and
as diving, the fall detection algorithm is based on the recognition of impact and posture phases activities,
posture such
phases of asof
diving,
the the the
falls,falls,
which fall detection
can can
which algorithm
be achieved
be achieved is based
by threshold
by threshold on the recognition
comparisons
comparisonsof someof
of someimpact
features.and posture
The The
features. RMSRMS phases
displays of
displays
the falls,change
distinctive
distinctivewhich
changecan
in beimpact
the
in achieved
the phase
impact by threshold
of
phase theof falls. comparisons
Thus,
the falls. RMSRMS
Thus, of can
can some
be be features.
selected as one
selected asThe
of RMS
one the displays
of features
the features
for fall detection. When RMS exceeded the acceleration threshold for impact (A ti), the impact phase
for fall detection. When RMS exceeded the acceleration threshold for impact (Ati), the impact phase
is detected. However,
is detected. However, onlyonly
based on the
based on RMS,
the RMS,the fall detection
the fall algorithm
detection algorithmmistakenly
mistakenly determines
determines
Sensors 2017, 17, 2096 7 of 15

distinctive change in the impact phase of the falls. Thus, RMS can be selected as one of the features
for fall2017,
Sensors detection.
17, 2096 When RMS exceeded the acceleration threshold for impact (Ati ), the impact phase 7 of 15
is detected. However, only based on the RMS, the fall detection algorithm mistakenly determines
some ADLs
ADLs as asfalls
fallseasily.
easily.Some
Some ADL,
ADL, such
suchas as
jumping
jumping andand
crouching, maymay
crouching, also also
obtain a similar
obtain RMS
a similar
distribution with falls, as shown in Figure 7. Thus, we detect the posture
RMS distribution with falls, as shown in Figure 7. Thus, we detect the posture phase by the Eulerphase by the Euler angle to
achieve
angle to precise
achievedetection after the after
precise detection impact thephase
impact is recognized. Given that
phase is recognized. the positive
Given that the and negative
positive and
values of roll and pitch only represent the direction of the human body, the
negative values of roll and pitch only represent the direction of the human body, the absolute value absolute value of roll
and
of pitch
roll andispitch
usedisto represent
used the posture
to represent of the of
the posture human body. body.
the human As shown in Figure
As shown 7a, during
in Figure falls,
7a, during
the posture
falls, of the of
the posture human body changes
the human from standing
body changes to lyingto
from standing (orlying
near (or
lying).
nearTherefore, the absolute
lying). Therefore, the
value of roll (abs (Roll)) or pitch (abs (Pitch)) angles change from nearly 0 ◦ to approximately 90◦ .
absolute value of roll (abs (Roll)) or pitch (abs (Pitch)) angles change from nearly 0° to
Meanwhile, during
approximately 90°. ADLs, the absolute
Meanwhile, during value
ADLs,of roll
theand pitch angles
absolute value have a small
of roll change,
and pitch as shown
angles have in a
Figure 7b. Thus, the roll and pitch can be used for posture phase detection
small change, as shown in Figure 7b. Thus, the roll and pitch can be used for posture phase through the comparison with
orientationthrough
detection threshold theofcomparison
posture (Otpwith). If the absolute threshold
orientation value of rollof or (Otp). If O
pitch exceeds
posture tp within
the absoluteTtpvalue
after
theroll
of impact phaseexceeds
or pitch (the time Otpthreshold
within Tfor posture),
tp after the posture
the impact phasephase is detected.
(the time thresholdAfter forthe recognition
posture), the
of the impact and posture phases, a fall of the human body is detected. Given
posture phase is detected. After the recognition of the impact and posture phases, a fall of the human that the recognition of
impactisand
body posture
detected. phases
Given is based
that on threshold
the recognition comparisons,
of impact the fallphases
and posture detection algorithm
is based displays
on threshold
good performance
comparisons, in detection
the fall real-time algorithm
fall detection. Oncegood
displays a fallperformance
occurs and the RMS and fall
in real-time Euler angle exceed
detection. Once
the thresholds, the fall is detected immediately.
a fall occurs and the RMS and Euler angle exceed the thresholds, the fall is detected immediately.

Impact
Posture

Start

(a) (b)
Figure 7. RMS, roll, and pitch during falling and activities of daily living (ADL): (a) forward fall
Figure 7. RMS, roll, and pitch during falling and activities of daily living (ADL): (a) forward fall ending
ending up lying; (b) jumping.
up lying; (b) jumping.

Meanwhile, the determination of the threshold for recognizing the impact and posture phases
Meanwhile,
is very important. theAs
determination
proposed byofPierleoni
the thresholdet al.for recognizing
[24], the impact thephase
impact is and posture
detected phases
when RMS is
very important.
exceeds As proposed
2.5 g. However, theby Pierleoni et
sensitivity of al.
fall[24], the impact
detection phase
based on isthe
detected
2.5 g iswhennot RMS
goodexceeds
in the
2.5 g. However, the sensitivity of fall detection based on the 2.5 g is not
reference [24]. Thus, we conducted a set of preliminary experiments with 5 subjects to determine good in the reference [24].
Thus, we conducted a set of preliminary experiments with 5 subjects to
the threshold, which is then used for validating the detection algorithm. The preliminary determine the threshold, which is
then used forhas
experiment validating
the same the protocol
detectionwithalgorithm. The preliminary
the validation experiment
experiment. has the samefrom
Our observation protocol
the
with the validation experiment. Our observation from the preliminary
preliminary experimental data found that the peak RMS of the impact phase is more than 2 g. experimental data found thatAtithe
is
peak RMS of the impact phase is more than 2 g. A ti is re-identified by our pre-experiments
re-identified by our pre-experiments and set as 2.3 g in our system. The value of 2.3 g is evaluated and set as
2.3 the
on g in data
our system.
from the The value ofand
shoulder 2.3 gwaist
is evaluated
segments, on the datahas
which from the shoulder
similar and waist
distribution duringsegments,
the fall
which
and ADL activities. For the foot segment, we found that Ati during the range of 2.2 g~4.0 g canfound
has similar distribution during the fall and ADL activities. For the foot segment, we reach
that A ti during the range of 2.2 g~4.0 g can reach the best detection performance.
the best detection performance. To facilitate the experiment implementation, we choose 2.3 g as the To facilitate the
experiment implementation, we choose 2.3 g as the same A ti for the shoulder,
same Ati for the shoulder, waist, and foot segments. The thresholds for posture phase recognition iswaist, and foot segments.
Theas
set thresholds for posture
Otp = 50 within Ttp = phase
2 s by recognition
referring tois[24] as Otp = they
set because 50 within Ttp = observed
have been 2 s by referring
to givetogood
[24]
because they have
performance in ourbeen observed to give good performance in our experiments.
experiments.
5. Fall Detection Experiment
5. Fall Detection Experiment
We design and conduct a set of experiments by employing subjects to perform specific activities
We design and conduct a set of experiments by employing subjects to perform specific
to test the performance of our system and also verify the best human body location to place the sensor
activities to test the performance of our system and also verify the best human body location to
unit. In literature, the previous methods have tried the sensor on different body segments individually,
place the sensor unit. In literature, the previous methods have tried the sensor on different body
segments individually, such as neck [26], waist [24,29,30], chest and thigh [31], and ankle [30].
Kangas et al. [32] compared the fall detection performance on the head, waist, and wrist and found
obvious differences between them. Because the detection performance differs greatly when the
sensor is placed on different body segments, our experiments tested the body segments of shoulder,
waist, and foot, concerning the facts of easy portability and low disturbance during the fall activities.
Sensors 2017, 17, 2096 8 of 15

such as neck [26], waist [24,29,30], chest and thigh [31], and ankle [30]. Kangas et al. [32] compared
the fall detection performance on the head, waist, and wrist and found obvious differences between
them. Because the detection performance differs greatly when the sensor is placed on different body
segments, our experiments tested the body segments of shoulder, waist, and foot, concerning the facts
of easy portability and low disturbance during the fall activities.
The activities designed in the experiment are based on the reality of the elderly fall in daily life,
which includes two categories, namely, falls in daily life and ADL designed as interference activities
for falls. As shown in Table 1, we divided 16 activities into two categories as follows: 7 falls and
9 ADLs. In daily life, most falls occur during stumbling on things or slipping on a smooth surface,
and then the body falls forward or backward. Alternatively, being hit by something from the left and
right sides then losing balance causes the body to fall sideways. Thus, the seven simulated falls are
grouped into four subcategories, namely, backward fall, forward fall, lateral left fall, and lateral right
fall. Furthermore, sometimes the body would not be lying on the ground horizontally, and they are
further subcategorized by concerning the different ending-up conditions in backward and forward
falls. ADL is designed to verify whether the fall detection system can evoke a fall alarm correctly.
In an ideal case, an ADL should not be detected by the system as a fall. However, in some ADLs,
the RMS and Euler angle of a certain body segment may have the same features as falls and then were
mistaken as falls, decreasing the accuracy of fall detection. Thus, we need to verify the best body
location to place the sensor to avoid misjudgment.

Table 1. Falls and ADL designed in the experiment.

Activities
Ending up lying
Backward fall Ending up falling to the ground with elbows
Ending up sitting after lying for 2 s
Fall Activities(FA) Ending up lying
Forward fall
Falling on the knees, ending up lying
Lateral left fall Ending up lying
Lateral right fall Ending up lying
Standing-sitting and lying on a mattress-standing
Standing-sitting on a chair-standing
Taking stairs
Bending over
Activities of Daily
Walking
Living (ADL)
Stretching shoulder
Staggering forward
Staggering backward
Twisting waist

A positive or a negative outcome is expected in all the activities in the experiment protocol.
If a fall activity evokes the fall alarm, it will be specified as a true positive outcome, and if an ADL
evokes the fall alarm, it will be specified as a false positive outcome. Similarly, it will be specified
as a true negative outcome or a false negative outcome if no alarm occurs for ADL or a fall activity.
An ideal fall detection system should identify all simulated falls as positive outcomes and all ADLs as
negative outcomes.
Conducting the experiment would be too risky for elderly users. Thus, a group of 15 healthy subjects
aged from 21 to 25 were asked to perform the simulated falls and ADLs (Figure 8). When subjects
simulate falls, falling to the ground directly could possibly injure them. Thus, we placed an inflatable mat
(thickness 15 cm) on the ground for cushion. Meanwhile, due to the good elasticity of the mat, the subjects
would uncontrollably bounce up at a short distance when they fall to the mat. Therefore, the jitter of RMS
and Euler angle of the impact are greater than that in the ideal situation.
Sensors 2017, 17, 2096 9 of 15
Sensors 2017, 17, 2096 9 of 15

Figure8.8.Conducting
Figure Conductingfalling
fallingand
andADL.
ADL.

Each
In our subject performed
experiment, all thethe
considering 16easy
activities and repeated
portability, the sensorthem
unitafter
was placing the sensor
respectively placed unit
on theon
the shoulder,
subjects’ shoulder, waist, andand
waist, foot,
footrespectively,
to conduct each and activity
each activity was Scotch
repeatedly. repeated tapethrice for each
was used to fixbody
the
segment.
sensor unitInonthe preparation
different segments stage,of the
the experimental
human body. environment was set
In detail, we fixed theup, andhorizontally
device the sensor unit on
was charged to ensure that the battery was adequate during the experiment.
the shoulder (or the foot) by the tape and fixed the device in the coat pocket for the waist segment. Then, the sensor unit
wasEuler
The fixedangle
to thedisplayed
participant’s body
a small segment,
deviation andthe
from thetheoretical
APP wasvalue readybecause
to be connected to the sensor
of the irregularities of
unit.
the In the
human experiment’s
body parts. conduction stage, the mat was fixed tightly by two persons to ensure
safety,
Eachand the participant
subject performed stood all thein 16the front of
activities andtherepeated
mat andthem thenafter
performed
placing all
thethe simulated
sensor unit onfalls.
the
After thewaist,
shoulder, falls, and
the participant performed
foot, respectively, all the
and each ADLs.
activity wasThe resultsthrice
repeated during forthe
eachexperiment
body segment.were
recorded
In by the APP
the preparation automatically.
stage, the experimental environment was set up, and the sensor unit was charged
We did
to ensure thatnotthe collect
batterythe was data of the same
adequate during activity sample onThen,
the experiment. the three humanunit
the sensor segments dueto
was fixed to
theparticipant’s
the following concerns: (1) the and
body segment, sensor theunits
APP simultaneously
was ready to beplaced on the
connected three
to the human
sensor segments
unit. In the
during activities
experiment’s can easily
conduction havethe
stage, errors
matdue
wastofixed
the issues
tightlyofbystabilization
two persons and
to inter-disturbance;
ensure safety, and(2) theit
is difficult stood
participant to distinguish
in the front which sensor
of the unit then
mat and issues the alarmallwhen
performed more thanfalls.
the simulated one After
occur,theand we
falls,
need
the further analysis
participant performed to verify
all theit. ADLs.
ThoughThe there may during
results be difference between the
the experiment were three repeats
recorded byofthean
activity
APP on the three human segments, it does not influence the detection results since the different
automatically.
is slight
We didwhen notacollect
subjecttherepeats
data ofthe thesame
sameactivity.
activity Furthermore,
sample on thethis threedifference between due
human segments samples
to theis
allowed by our detection algorithm.
following concerns: (1) the sensor units simultaneously placed on the three human segments during
activities can easily have errors due to the issues of stabilization and inter-disturbance; (2) it is difficult
6. distinguish
to Experimental Results
which sensor unit issues the alarm when more than one occur, and we need further
analysis to verify
In the it. Though
experiment, thereroll,
the RMS, mayand be difference
pitch data between
of all the the three repeats
participants on theof three
an activity
human onbody
the
three human segments, it does not influence the detection results since
segments (shoulder, waist, and foot) were captured and recorded, respectively, during all the the different is slight when
aspecified
subject repeats
simulatedthe same activity.
falls and ADLs. Furthermore,
We found that this difference
all the fallbetween
activitiessamples
can be is allowedcorrectly
detected by our
detection
when thealgorithm.
sensor unit was placed on the three human body segments. However, some ADLs were
detected incorrectly when the sensor unit was placed on the shoulder and foot segments.
6. Experimental Results
In the experiment, the RMS, roll, and pitch data of all the participants on the three human body
segments (shoulder, waist, and foot) were captured and recorded, respectively, during all the specified
simulated falls and ADLs. We found that all the fall activities can be detected correctly when the sensor
unit was placed on the three human body segments. However, some ADLs were detected incorrectly
when the sensor unit was placed on the shoulder and foot segments.
Figure 9 shows the RMS and absolute value(a)of roll and pitch distribution on the three body
segments during typical fall activities, including the following: (a) backward fall ending up lying;
(b) forward fall ending up lying; (c) lateral right fall ending up lying; and (d) lateral left fall ending
up lying. The peaks of RMS distribution on all three body segments exceeds 2.3 g, and the peaks of
absolute of roll or pitch exceeded 50◦ during the simulated falls. Thus, these activities were detected
as falls correctly based on the threshold comparison.
(b)
In the experiment, the RMS, roll, and pitch data of all the participants on the three human body
segments (shoulder, waist, and foot) were captured and recorded, respectively, during all the
specified simulated falls and ADLs. We found that all the fall activities can be detected correctly
when the sensor unit was placed on the three human body segments. However, some ADLs were
Sensors 2017, 17, 2096 10 of 15
detected incorrectly when the sensor unit was placed on the shoulder and foot segments.

Sensors 2017, 17, 2096 10 of 15

(a)
(c)

Sensors 2017, 17, 2096 10 of 15


(b)
(d)
Figure 9. RMS and absolute value of roll and pitch distribution on the shoulder, waist, and foot
during typical fall activities: (a) Backward fall ending up lying; (b) Forward fall ending up lying;
(c) Lateral right fall ending up lying; (d) Lateral left fall ending up lying.
(c)
Figure 9 shows the RMS and absolute value of roll and pitch distribution on the three body
segments during typical fall activities, including the following: (a) backward fall ending up lying; (b)
forward fall ending up lying; (c) lateral right fall ending up lying; and (d) lateral left fall ending up
lying. The peaks of RMS distribution on all three body segments exceeds 2.3 g, and the peaks of
absolute of roll or pitch exceeded 50° during the simulated falls. Thus, these activities were detected
as falls correctly based on the threshold comparison. (d)
Figure 9.10RMSshows
andandRMS and
absolute absolute
value of roll value ofdistribution
roll and on pitch distribution waist,on the three body
Figure
Figure 9. RMS absolute value ofandrollpitch
and pitch distribution the on
shoulder,
the shoulder, and foot
waist, during
and foot
segments
typical during
during some
falltypical
activities:ADLs:
fall (a) (a)
Backward
activities: bending
(a) fall over,
endingfall
Backward (b) staggering
up lying;
ending(b) forward,
upForward
lying; (b) (c)
fallForward walking,
ending up falllying;and
ending (c)up(d) taking
Lateral
lying;
stairs.right
Figure
fall 10a,b
ending are
up detected
lying; (d) as falls
Lateral incorrectly
left fall ending on
up the
lying.shoulder
(c) Lateral right fall ending up lying; (d) Lateral left fall ending up lying. segment because the shoulder
sharply changes from near vertical to near horizontal status when the human body is bending over
Figure
rapidly Figure 10 shows
or staggering
9 shows RMStheand
forward. RMS absolute
In and value
addition,
absolute ofvalue
Figureroll10c,
andofdpitch distribution
are and
roll detected on the
pitchasdistribution
falls threeonbody
incorrectly segments
onthree
the the footbody
during
segment
segmentssome
because ADLs:
during (a)
thetypical bending
movement over,
of the feet
fall activities, (b) staggering
during the
including forward,
walking (c)
or taking
following: walking, and
stairs is large
(a) backward (d) taking
compared
fall ending stairs.(b)
with
up lying;
Figure
that of 10a,b
forward other are
fall detected
human
ending body assegments.
falls(c)
up lying; incorrectly
Theright
lateral on fall
peak the shoulder
of ending
RMS and upsegment
the peak
lying; and because the shoulder
of(d)absolute
lateral value
left fall of sharply
pitchup
ending
changes
exceed 2.3from
g near
and vertical
50°, to near
respectively, horizontal
on the status
shoulder when
and footthe human
body body
segments.
lying. The peaks of RMS distribution on all three body segments exceeds 2.3 g, and the peaks of is bending
Meanwhile, over
in rapidly
Figure
or absolute
staggering
10a–d, during forward.
the same
of roll In
ADLs
or pitch addition,
performed
exceeded Figure on10c,d
50° during other
theare detectedthe
segments,
simulated as falls
RMS
falls. incorrectly
Thus,peaks
these on the2.3
areactivities
below foot segment
g and
were the
detected
because
peaks of the
roll movement
and pitch of
are the feet
below during
50°. The
as falls correctly based on the threshold comparison. walking
ADLs or
were taking
detectedstairs is large
correctly. compared with that of other
human Figure
body segments.
10 showsThe RMS peak
and of absolute
RMS and value the peak of of
rollabsolute
and pitchvalue of pitch exceed
distribution on the 2.3 three 50◦ ,
g and body
respectively,
segments during on the shoulder
some ADLs: and(a)foot body segments.
bending Meanwhile,
over, (b) staggering in Figure
forward, (c)10a–d,
walking, during
and the(d) same
taking
ADLs performed on other segments, the RMS peaks are below 2.3 g and
stairs. Figure 10a,b are detected as falls incorrectly on the shoulder segment because the shoulderthe peaks of roll and pitch are
below 50 ◦ . The ADLs were detected correctly.
sharply changes from near vertical to near horizontal status when the human body is bending over
rapidly or staggering forward. In addition, Figure 10c, d are detected as falls incorrectly on the foot
segment because the movement of the feet during walking or taking stairs is large compared with
that of other human body segments. The peak of RMS and the peak of absolute value of pitch
exceed 2.3 g and 50°, respectively, on the shoulder and foot body segments. Meanwhile, in Figure
10a–d, during the same ADLs performed on other segments, the RMS peaks are below 2.3 g and the
peaks of roll and pitch are below 50°. The ADLs
(a)were detected correctly.

(b)(a)
Figure 10. Cont.

(b)
Sensors
Sensors 2017,
2017, 17, 2096
17, 2096 11 11 of 15
of 15

(c)

(d)
Figure
Figure 10. RMS
10. RMS and and absolute
absolute valuevalue
of rollof roll
and anddistribution
pitch pitch distribution on the shoulder,
on the shoulder, waist, andwaist, and foot
foot during
during ADLs: (a) Bending over; (b) Staggering forward; (c) Walking ; (d)
ADLs: (a) Bending over; (b) Staggering forward; (c) Walking; (d) Taking stairs. Taking stairs.

TheThe correct
correct detection
detection rate rate of these
of these 16 activities
16 activities including
including fallsADL
falls and andonADLthe on thehuman
three three body
human
segments is summarized in Table 2. This rate is obtained through dividing the number of corrected of
body segments is summarized in Table 2. This rate is obtained through dividing the number
corrected
detections by detections by the total
the total detection detection
samples for an samples for an activitybyimplemented
activity implemented by As
all the subjects. all mentioned
the subjects.
As mentioned
in Section in Section
5, the true positive5,orthe
truetrue positive
negative or true isnegative
detection taken intodetection
accountisastaken into account
one result as one
of corrected
detection. The correct detection rate for all the fall activities is 100%, whereas the detection correct ratethe
result of corrected detection. The correct detection rate for all the fall activities is 100%, whereas
detection
in some ADLcorrect
cases arerate in some
variable, ADL cases
namely, are variable,
60% correct rate fornamely,
bending60%overcorrect rate for bending
and staggering forward onover
theand staggering
shoulder forward
segment. on the shoulder
In addition, segment.
6.7% correct In 0%
rate and addition,
correct6.7%
rate correct rate and
are observed for 0% correct
walking and rate
are observed for walking and taking
taking stairs, respectively, on the foot segment.stairs, respectively, on the foot segment.

Table
Table 2. Detection
2. Detection accuracy
accuracy of activities
of 16 16 activities
on on
thethe three
three human
human body
body segments
segments in the
in the experiment.
experiment.
Shoulder Waist Foot
Shoulder Waist Foot
16 Activities
16 Activities Correct Correct Correct
Correct Rate (%) Correct Rate (%) Correct Rate (%)
Rate (%) Rate (%) Rate (%)
Backward fall
Backward
Ending up lying fall 100 100 100
Ending up falling to the ground with elbows 100 100 100
Ending up lying
Ending up siting after lying for 2 s 100
100 100
100 100
100
Falls activities Forward
Ending up falling fallto the ground with
Ending up lying 100 100 100100 100
100
Falling on the kneeselbows
ending up lying 100 100 100
EndingLateral
up siting
left fallafter lying for 2s 100 100 100
Ending up lying 100 100 100
Falls activities Forward fall
Lateral right fall 100 100100 100
100 100
Ending
Ending up lyingup lying
Falling on Walking
the knees ending up lying 100 100 100100 100
6.7
Bending over
Lateral left fall 60 100 100
Taking stairs 100 100 100100 100
0
Ending up lying
Standing-sitting on a chair-standing 100 100 100
ADL activities Standing-sitting and right
Lateral lying on a
fall 100 100 100
mattress-standing 100 100 100
Ending up
Stretching shoulder lying 100 100 100
StaggeringWalking
forward 60 100 100100 6.7
100
Staggering backward 100 100 100
Bending
Twisting waist
over 100
60 100
100 100
100
Taking stairs 100 100 0
Standing-sitting on a chair-standing 100 100 100
We can further analyze the detection performance of all activities (accuracy), simulated falls
Standing-sitting and lying on a
ADL
(sensitivity), and the ADLs (specificity) to demonstrate the best 100
activities body location100 100sensor
for placing the
mattress-standing
unit. The accuracy means the total correct detection rate of all the 16 activities on a human segment,
Stretching shoulder 100 100 100
in which the sensitivity is the detection correct rate of the 7 simulated falls activities, and the specificity
Staggering forward 60 100 100
is the detection correct rate ofStaggering
the 9 ADLs. As shown in Figure 11,
backward 100the sensitivity
100of detection100
method
is 100% for all three body segments, whereas
Twisting waist the specificity of detection
100 method
100 for shoulder,
100waist,
and foot are 91.1%, 100%, and 78.5%, respectively. These values mean that our system has high accuracy
We can further
of fall detection, and theanalyze the detection
waist segment performance
is the best location toofplace
all activities (accuracy),
the sensor simulated
unit, wherein falls
both the
(sensitivity), and the ADLs (specificity) to demonstrate the best body location for placing the sensor
Sensors 2017, 17, 2096 12 of 15
Sensors 2017, 17, 2096 12 of 15

detection
sensor sensitivity
unit, wherein and bothspecificity can reach
the detection 100%. and
sensitivity During the experiment,
specificity can reachall100%.
the subjects
Duringreach
the
an agreement that the waist is also the most comfortable location to place the sensor unit,
experiment, all the subjects reach an agreement that the waist is also the most comfortable location which is
simply
to place placed in the
the sensor jacket
unit, pocket.
which is simply placed in the jacket pocket.

Falls detection performance on different human body segments


120
100 100 100 100 100
95 91.1
100 87.9
78.5
80
60
40
20
0
Accuracy(%) Sensitivity(%) Specificity(%)

Shoulder Waist Foot

Figure
Figure11.
11.Fall
Falldetection
detectionperformance
performanceon
ondifferent
differenthuman
humanbody
body segments.
segments.

7.
7.Discussion
Discussionand
andConclusions
Conclusions
In the literature,
In the literature,therethereareare already
already manymanyworksworks
for humanfor human fall detection.
fall detection. To address Totheaddress
advantages the
advantages of our method, we make a comparison with some
of our method, we make a comparison with some previous methods in terms of sensor location, previous methods in terms of sensor
location,
accuracy,accuracy,
portability, portability,
and comfort. and comfort.
Though the Though the experiment
experiment protocolsprotocols of these methods
of these methods are differentare
different and the datasets for detection are based on their experiments,
and the datasets for detection are based on their experiments, all of them considered the distinction all of them considered the
distinction
between simulatedbetween falls simulated
and ADLs. falls InandHeADLs.
et al.’s In He et[26],
method al.’sthemethod [26], the
acceleration andacceleration
angular velocity and
angular
data arevelocity
measured databyare measured
a sensor boardby fixed
a sensor board
in the vestfixed
and in the vest
closed to theandneckclosedandtotransmitted
the neck and to
transmitted to a smart phone. They detect falls based on a Bayes network
a smart phone. They detect falls based on a Bayes network classifier with an accuracy of 95.67%. classifier with an accuracy
of
In 95.67%.
Pierleoni InetPierleoni
al.’s methodet al.’s method
[24], [24], thedevice
the wearable wearable device
is placed onisthe
placed
waiston bythe waist abybelt,
wearing wearing
and the a
belt, and the fall
fall detection detection
is based on theisthreshold
based oncomparison
the threshold comparisonand
of acceleration of Euler
acceleration and Euler
angle, which angle,
is similar to
which is similar to our method. However, our method improves
our method. However, our method improves the detection accuracy on the waist segment to 100%, the detection accuracy on the waist
segment
benefitting to 100%,
from the benefitting
threshold from
of RMSthe threshold
re-identifiedof RMS
in our re-identified
experiment.inIn our experiment.
Erdogan et al.’sInstudy
Erdogan[29],
et al.’s study [29], the data were also measured by wearing a belt
the data were also measured by wearing a belt on the waist, but a computer is needed to receive on the waist, but a computer is
needed to receive and process the data. They used a k-nearest neighbor
and process the data. They used a k-nearest neighbor algorithm to detect falls based on acceleration algorithm to detect falls
based
data atonanacceleration
accuracy ofdata 89.4%.at an
Inaccuracy
Ojetola etofal.’s89.4%.
methodIn Ojetola
[31], theet al.’s method [31],
acceleration the acceleration
and angular velocity
and angular velocity data are also measured by wearing a belt on both
data are also measured by wearing a belt on both the chest and right thigh and are then transmitted the chest and right thigh and
are then transmitted to a computer. The detection accuracy range
to a computer. The detection accuracy range is from 98.91% to 99.45%, which is dependent on the is from 98.91% to 99.45%, which is
dependent
training of on the training
a machine learningof a decision
machinetree. learning decision
In Sorvala tree.method
et al.’s In Sorvala [30],ettheal.’s
datamethod [30], the
is measured by
data is measured by wearing a wireless inertial measurement
wearing a wireless inertial measurement unit (IMU) on the waist and an accelerometer on an ankle.unit (IMU) on the waist and an
accelerometer
The detection accuracyon an ankle. The with
is 98.69% detection
95.6% accuracy
sensitivityisand 98.69%
99.6%with 95.6% Yu
specificity. sensitivity and 99.6%
et al.’s system [33] is
specificity. Yu et al.’s system [33] is vision-based by using a camera
vision-based by using a camera and a personal computer, indicating that the fall detection system is and a personal computer,
indicating
limited in the thatvision
the field
fall detection
of the camera system is limited
and the portabilityin isthenotvision
good. field of the camera
The detection accuracyand the
of their
portability is not good.
method is 98.13%. Thus,The ourdetection
method has accuracy of their
the highest methodaccuracy
detection is 98.13%. Thus, compared
of 100% our method hasother
with the
highest detection accuracy of 100% compared with other methods.
methods. Meanwhile, our system has good portability because the small packaged sensor unit is easily Meanwhile, our system has
good
fixed portability
in a pocket of because
a top. Forthe the
small packaged
wearing sensor
comfort, all unit is easilyaccelerometer
the existing fixed in a pocket of a top.methods
sensor-based For the
wearing
includingcomfort,
[24,26] and all the existing
[29–31] dependaccelerometer
on the wearingsensor-based
time because methods including
long wearing [24,26]
time and or
of a vest [29–31]
a belt
depend on the wearing time because long wearing time of a vest or
on the waist, thigh, and ankle will lead to discomfort for the user, whereas our method is comfortablea belt on the waist, thigh, and
ankle will lead to discomfort for the user, whereas
to the user, even when porting the sensor in the pocket for a long time. our method is comfortable to the user, even
when porting the sensor in the pocket for a long time.
Our method proposes an easy-to-use, portable human fall monitoring system using only a
small self-packaged sensor unit and a mobile phone. We designed 16 simulated falls and ADLs and
Sensors 2017, 17, 2096 13 of 15

Our method proposes an easy-to-use, portable human fall monitoring system using only a small
self-packaged sensor unit and a mobile phone. We designed 16 simulated falls and ADLs and
conducted experiments by placing the sensor units on different body segments, such as the shoulder,
waist, and foot of participants. Although fall activities can be correctly detected in all the three body
segments, ADL is still possibly mistaken as falls when the sensor unit is placed on the shoulder and
foot, but not when placed on the waist. The monitoring system can achieve 100% accuracy, sensitivity,
and specificity on the waist segment, verifying that the waist is the best location for placing the sensor.
The sensor unit can be placed in the pocket of a top at waist-height without needing to be fixed on the
skin. Furthermore, our system is highly portable and comfortable because the sensor unit can be fixed
in the pocket of a top and works only with a mobile phone, which offers great portability and will not
lead to discomfort during usage. Although we designed the experiment based on the reality of the
elderly fall in daily life, the subjects in the experiment are young people to avoid any potential risk of
injury. We still need to consider the possible differences in fall activities between the elderly people
and young people. Evaluation and validation of the practical application of the proposed system for
elderly people is our next objective.

Acknowledgments: The work of this paper is financially supported by The National Natural Science Foundations
of China (No. 61502112), NSF of Guangdong Province (No. 2016A030313499), The Science and Technology
Planning Project of Guangdong Province (No. 2015A030401030, No.2015A020219014), Guangzhou Municipal
University Research Projects (No. 1201630279), and the Fundamental Research Funds for the Central Universities
(No. 2017ZD054), The Guangdong Province Universities and Colleges Excellent Youth Teacher Training Program
and the Student Research Project (SRP) of South China University of Technology.
Author Contributions: Xuedong Ma contributes on realizing the fall monitor system, collecting and interpreting
the experimental data; Aihua Mao contributes on conception and design of the work, writing and revising the
paper; Yinan He contributes on design and conducting the experiment; Jie Luo contributes on conception and
design of the work, writing and revising the paper.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. World Health Organization (WHO). Available online: http://www.who.int/mediacentre/factsheets/fs344/
en/ (accessed on 12 September 2017).
2. Heinrich, S.; Rapp, K.; Rissmann, U.; Becker, C.; König, H.H. Cost of falls in old age: A systematic review.
Osteoporos. Int. 2009, 21, 891–902. [CrossRef] [PubMed]
3. Tran, T.H.; Le, T.L.; Hoang, V.N.; Vu, H. Continuous detection of human fall using multimodal features
from Kinect sensors in scalable environment. Comput. Methods Progr. Biomed. 2017, 146, 151–165. [CrossRef]
[PubMed]
4. Bian, Z.P.; Hou, J.; Chau, L.P.; Magnenat-Thalmann, N. Fall detection based on body part tracking using
a depth camera. IEEE J. Biomed. Health Inform. 2015, 19, 430–439. [CrossRef] [PubMed]
5. Tao, S.; Kudo, M.; Nonaka, H. Privacy-preserved behavior analysis and fall detection by an infrared ceiling
sensor network. Sensors 2012, 12, 16920–16936. [CrossRef] [PubMed]
6. Stone, E.E.; Skubic, M. Fall detection in homes of older adults using the Microsoft Kinect. IEEE J. Biomed.
Health Inform. 2015, 19, 290–301. [CrossRef] [PubMed]
7. Liu, S.H.; Cheng, W.C. Fall detection with the support vector machine during scripted and continuous
unscripted activities. Sensors 2012, 12, 12301–12316. [CrossRef] [PubMed]
8. Karantonis, D.M.; Narayanan, M.R.; Mathie, M.; Lovell, N.H.; Celler, B.G. Implementation of a real-time
human movement classifier using a triaxial accelerometer for ambulatory monitoring. IEEE Trans. Inform.
Technol. Biomed. 2006, 10, 156–167. [CrossRef]
9. Lai, C.F.; Chang, S.Y.; Chao, H.C.; Huang, Y.M. Detection of cognitive injured body region using multiple
triaxial accelerometers for elderly falling. IEEE Sens. J. 2011, 11, 763–770. [CrossRef]
10. Kangas, M.; Konttila, A.; Winblad, I.; Jamsa, T. Determination of Simple Thresholds for Accelerometry-Based
Parameters for Fall Detection. In Proceedings of the EMBS 2007 29th Annual International Conference of the
IEEE Engineering in Medicine and Biology Society, Lyon, France, 22–26 August 2007; pp. 1367–1370.
Sensors 2017, 17, 2096 14 of 15

11. Luque, R.; Casilari, E.; Morón, M.J.; Redondo, G. Comparison and characterization of android-based fall
detection systems. Sensors 2014, 14, 18543–18574. [CrossRef] [PubMed]
12. Li, Q.; Stankovic, J.A.; Hanson, M.A.; Barth, A.T.; Lach, J.; Zhou, G. Accurate, Fast Fall Detection
Using Gyroscopes and Accelerometer-Derived Posture Information. In Proceedings of the BSN 2009
Sixth International Workshop on Wearable and Implantable Body Sensor Networks, Berkeley, CA, USA,
3–5 June 2009; pp. 138–143.
13. Nari, M.I.; Suprapto, S.S.; Kusumah, I.H.; Adiprawita, W. A Simple Design of Wearable Device for Fall
Detection with Accelerometer and Gyroscope. In Proceedings of the 2016 International Symposium on
Electronics and Smart Devices (ISESD), Bandung, Indonesia, 29–30 November 2016; pp. 88–91.
14. Andò, B.; Baglio, S.; Lombardo, C.O.; Marletta, V. A multisensor data-fusion approach for ADL and fall
classification. IEEE Trans. Instrum. Meas. 2016, 65, 1960–1967. [CrossRef]
15. Yang, L.; Ren, Y.; Hu, H.; Tian, B. New fast fall detection method based on spatio-temporal context tracking
of head by using depth images. Sensors 2015, 15, 23004–23019. [CrossRef] [PubMed]
16. Gasparrini, S.; Cippitelli, E.; Spinsante, S.; Gambi, E. A depth-based fall detection system using
a Kinect®sensor. Sensors 2014, 14, 2756–2775. [CrossRef] [PubMed]
17. Lee, Y.S.; Chung, W.Y. Visual sensor based abnormal event detection with moving shadow removal in home
healthcare applications. Sensors 2012, 12, 573–584. [CrossRef] [PubMed]
18. Anderson, D.; Keller, J.M.; Skubic, M.; Chen, X.; He, Z. Recognizing Falls from Silhouettes. In Proceedings of
the 2006 International Conference of the IEEE Engineering in Medicine and Biology Society, New York, NY,
USA, 30 August–3 September 2006; pp. 6388–6391.
19. Rougier, C.; Meunier, J.; St-Arnaud, A.; Rousseau, J. Monocular 3D Head Tracking to Detect Falls of Elderly
People. In Proceedings of the 2006 International Conference of the IEEE Engineering in Medicine and Biology
Society, New York, NY, USA, 30 August–3 September 2006; pp. 6384–6387.
20. Munoz-Organero, M.; Lotfi, A. Human movement recognition based on the stochastic characterization of
acceleration data. Sensors 2016, 16, 1464. [CrossRef] [PubMed]
21. Hsieh, C.Y.; Liu, K.C.; Huang, C.N.; Chu, W.C.; Chan, C.T. Novel hierarchical fall detection algorithm using
a multiphase fall model. Sensors 2017, 17, 307. [CrossRef] [PubMed]
22. Kurniawan, A.; Hermawan, A.R.; Purnama, I.K. E. A Wearable Device for Fall Detection Elderly People
Using Tri Dimensional Accelerometer. In Proceedings of the 2016 International Seminar on Intelligent
Technology and Its Applications (ISITIA), Lombok, Indonesia, 28–30 July 2016; pp. 671–674.
23. Albert, M.V.; Kording, K.; Herrmann, M.; Jayaraman, A. Fall classification by machine learning using mobile
phones. PLoS ONE 2012, 7, e36556. [CrossRef] [PubMed]
24. Pierleoni, P.; Belli, A.; Palma, L.; Pellegrini, M.; Pernini, L.; Valenti, S. A high reliability wearable device for
elderly fall detection. IEEE Sens. J. 2015, 15, 4544–4553. [CrossRef]
25. Wibisono, W.; Arifin, D.N.; Pratomo, B.A.; Ahmad, T.; Ijtihadie, R.M. Falls Detection and Notification
System Using Tri-axial Accelerometer and Gyroscope Sensors of a Smartphone. In Proceedings of the
2013 Conference on Technologies and Applications of Artificial Intelligence (TAAI), Taipei, Taiwan,
6–8 December 2014; pp. 382–385.
26. He, J.; Bai, S.; Wang, X. An unobtrusive fall detection and alerting system based on Kalman filter and Bayes
network classifier. Sensors 2017, 17, 1393. [CrossRef] [PubMed]
27. Ntalampiras, S.; Roveri, M. An Incremental Learning Mechanism for Human Activity Recognition.
In Proceedings of the 2016 IEEE Symposium Series on Computational Intelligence (SSCI), Athens, Greece,
6–9 December 2016; pp. 1–6.
28. Yean, S.; Lee, B.S.; Yeo, C.K.; Vun, C.H. Algorithm for 3D Orientation Estimation Based on Kalman Filter and
Gradient Descent. In Proceedings of the 2016 IEEE 7th Annual Information Technology on Electronics and
Mobile Communication Conference (IEMCON), Vancouver, BC, Canada, 13–15 October 2016; pp. 1–6.
29. Erdogan, S.Z.; Bilgin, T.T. A data mining approach for fall detection by using k-nearest neighbour algorithm
on wireless sensor network data. IET Commun. 2012, 6, 3281–3287. [CrossRef]
30. Sorvala, A.; Alasaarela, E.; Sorvoja, H.; Myllyla, R. A Two-Threshold Fall Detection Algorithm for
Reducing False Alarms. In Proceedings of the 6th International Symposium on Medical Information
and Communication Technology (ISMICT), La Jolla, CA, USA, 25–29 March 2012; pp. 1–4.
Sensors 2017, 17, 2096 15 of 15

31. Ojetola, O.; Gaura, E.I.; Brusey, J. Fall Detection with Wearable Sensors-SAFE (Smart Fall Detection).
In Proceedings of the 7th International Conference on Intelligent Environments, Nottingham, UK,
25–28 July 2011; pp. 318–321.
32. Kangas, M.; Konttila, A.; Lindgren, P.; Winblad, I.; Jämsä, T. Comparison of low-complexity fall detection
algorithms for body attached accelerometers. Gait Posture 2008, 28, 285–291. [CrossRef] [PubMed]
33. Yu, M.; Rhuma, A.; Naqvi, S.M.; Wang, L.; Chambers, J. A posture recognition-based fall detection system
for monitoring an elderly person in a smart home environment. IEEE Trans. Inf. Technol. Biomed. 2012, 16,
1274–1286. [PubMed]

© 2017 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).