See discussions, stats, and author profiles for this publication at: [Link]

net/publication/339572383

Sensor Technologies for Fall Detection Systems: A Review

Article in IEEE Sensors Journal · February 2020

DOI: 10.1109/JSEN.2020.2976554

CITATIONS READS
115 1,619

4 authors:

Anuradha Singh Saeed Rehman

Auckland University of Technology Flinders University
5 PUBLICATIONS 221 CITATIONS 81 PUBLICATIONS 1,872 CITATIONS

SEE PROFILE SEE PROFILE

Sira Yongchareon P.H.J. Chong

Auckland University of Technology Auckland University of Technology
83 PUBLICATIONS 1,195 CITATIONS 330 PUBLICATIONS 6,336 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Sira Yongchareon on 09 February 2023.

The user has requested enhancement of the downloaded file.

 IEEE SENSORS JOURNAL, VOL. 20, NO. 13, JULY 1, 2020 6889

Sensor Technologies for Fall Detection

Systems: A Review
Anuradha Singh , Saeed Ur Rehman, Senior Member, IEEE,
Sira Yongchareon, Senior Member, IEEE, and
Peter Han Joo Chong , Member, IEEE

Abstract —The risk of falls in older adults restrict their

social life and independent living. The assisted living devices
help older adults to live independently in their home, giving a
psychological boost, and releasing the burden on the care-
giver and the healthcare providers. A robust and accurate
fall detection system is essential to provide immediate help
and to reduce the severe post-fall consequences, and the
associated medical care cost significantly. This review aims
to provide a comprehensive technical insight into the existing
fall detection system, to classify various approaches and
the challenges encountered during implementation. The fall
detectors are broadly classified into three categories, namely
wearable, ambiance-based, and hybrid sensing detectors,
which are further explored by the sensor technology. This review provides a comprehensive overview of each competing
sensor technology ranging from an accelerometer, pressure sensor, and radar to camera-based and their infusion into a
complete fall detection system. It outlines the strength and limitations of different sensor fall detection systems in terms
of feature extraction, classification, performance, and experimental dataset. The user adaptability, installation complexity,
and power requirement of the systems are the main areas, which are not addressed adequately in the literature. In the
end, the review provides a basic framework in deciding the technology for a specific scenario or location according to
the prerequisites for the deployment.
Index Terms — Assisted living, elderly assisted living, fall detection, smart homes, sensor technology, wearable sensor.

I. I NTRODUCTION in other tissues. Fall-related injuries result in pain, disability,

and sometimes leading to premature death [3]. The average
T HE world health organization reported that approximately
28-35% of the people of age over 65 years had suffered
fall every year, and it increases to 32-42% for older adults
medical cost for injuries related to falling varies widely across
countries. For instance, the average yearly health system
over 70 years of age [1]. It is projected that the worldwide cost per fall for the people aged 65 and above in Finland
population of old aged people over 65 would increase by US$ 3,611, Australia US$ 1,049, Ireland US$ 6,646, and
21.64% by 2050 [2]. The effect and risk of falls increased in the USA was US$ 17,483 for the survey year of 2002
exponentially with age due to weak leg strength, side effects of (as reported in 2007) [4]. Specifically, the total population
prolonged medication, vision problems, and strength reduction of New Zealand is predicted to increase by 25% over the
next 30 years, and 150% for people aged 75 and above. The
average health system cost per annum is expected to rise to
Manuscript received December 3, 2019; revised January 24, 2020;
accepted February 13, 2020. Date of publication February 28, 2020; NZ$ 10 billion from existing NZ$ 7,000 million within the
date of current version June 4, 2020. The associate editor coordi- next 30 years [5]. More than 25% of this is expended on the
nating the review of this article and approving it for publication was people above 70 years of age. Injuries due to fall being one
Dr. Marco J. Da Silva. (Corresponding author: Anuradha Singh.)
Anuradha Singh and Peter Han Joo Chong are with the of the most significant causes of this medical care expense.
Department of Electrical and Electronic Engineering, Auckland From 2011 to 2016 the fall-related injuries increased from
University of Technology, Auckland 1142, New Zealand (e-mail: 170,000 to 216,000 approximately by 27%.
engganuradhasingh@[Link]).
Saeed Ur Rehman is with the College of Science and Engineering, The fear of falling itself leads to a restrictive way of living
Flinders University, Adelaide, SA 5042, Australia. for the old age people, e.g., restricted social life, reduced
Sira Yongchareon is with the Department of IT and Software daily activities, and depression [6]. The alarmingly increasing
Engineering, Auckland University of Technology, Auckland 1142,
New Zealand. rate of fall-related injuries necessitates the need for assistive
Digital Object Identifier 10.1109/JSEN.2020.2976554 care devices to detect and raise alarms in emergencies [7].

1558-1748 © 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See [Link] for more information.

Authorized licensed use limited to: Auckland University of Technology. Downloaded on October 10,2020 at 06:20:29 UTC from IEEE Xplore. Restrictions apply.
6890 IEEE SENSORS JOURNAL, VOL. 20, NO. 13, JULY 1, 2020

A robust fall detector can contribute in time to attend the fall. A detailed technical review on the status of present tech-
A person lying for hours after falls could result in hypothermia, nologies deployment can be informative for research and
dehydration and pressure sore, thus increasing the possibility useful in bridging the space between adaptation to the real
of admission in hospitals. After a fall, most of the older adults world and the laboratory environment. Unlike other review
cannot return to a standing position on their own [8]. The articles, our primary focus is on the comprehensive review of
response and rescue time of the caregiver affects the resultant present trends of technologies and offering detailed insight into
consequence of fall to a larger extent and critically determines each type of sensor technology. The segmentation approach is
the healthcare cost. This leads to an urgent need for fast fall taken for reviewing the fall detection system, and each stage of
detection and sending an alarm to the caregiver for immediate its architecture, including sensor, signal extraction, classifica-
health assistance that consequently reduces the healthcare cost. tion mechanism, experimentation dataset, and communication,
The robust fall detection system has considerable potential in is presented in detail. Our review article explains the basic
elderly care for their security and safety in homes. This has principle of sensor technology applied to the realization of the
been the motivation for researchers to develop reliable and fall detection system. The review highlights the implementa-
robust fall detection. The high sensitivity and accuracy in a tion challenges associated with each sensor technology. The
fall detection system are desirable, as high false fall detection power requirement, range of operation, cost, privacy aware-
may give an impression to caregivers that the device is useless ness, usability, and installation complexity are considered as
and may lead to even rejection. It may also result in the loss common evaluating parameters. The comparative analysis of
of confidence in the detectors, and the purpose of the fall sensor technologies is presented in Table XIII. Depending on
detection system may be diminished. the accuracy and sensitivity of the system, the review filters
Since the conception of the idea of a fall detection system, out the choice of extraction and classification algorithm. The
the methodologies have been evolved from non-automatic limited and restricted experimental environment raises the
to automatic fall detectors. The non-automatic fall detectors question of the adaptability of technologies in a real-world
require user intervention for triggering the alarm. It becomes scenario, which needs to be further explored. Better machine
useless in case the older adults lose consciousness after a fall learning with larger data sets or better detection algorithm
and are unable to trigger the alarm for help. Automatic fall may increase the accuracy, which shall increase the adapt-
detection technologies have addressed this issue to a more ability of fall detection systems. This review comprehensively
considerable extent. Furthermore, the adaptability of older documents the unaddressed issues open for future research.
people towards this is encouraging [9]. Automated monitoring Following are the main contributions of this review article
of fall reduces the time to attend the person, which in turn • The review has consolidated an insightful technologi-
minimize the damage caused by fall. cal analysis of the fall detection system. The working
Extensive technological research is carried out in the principle of each sensor and its development into a fall
automation of the fall detection system, but the adaptation detection system is presented.
in the real-world scenario is not teaming up with the pace. • The review aims to present the unaddressed open chal-
The reviews on fall detection, namely Noury et al. [10], lenges of the various technologies and provides reference
Perry et al. [11], Mubashir et al. [12], Birku and Survey [13] architecture in the field of automatic fall detection that
mainly provided a general outline of the fall detection status. would help in guiding future research.
However, the trends and challenges of specific technologies • This review presented a clear insight into the benefits and
deployment have left unaddressed. Madhubala et al. [14], limitations of different technologies. It provides a basic
Pannurat et al. [15], Bet et al. [16] provided insight on framework in deciding the fall detections system for a
wearable devices, but the rigorous approach to search method- specific scenario or location according to the prerequisites
ology and the inclusion of all types of technologies were not of the deployment.
considered. Cippitelli et al. [17] presented a detail discussion Section II presents the fall detection concept and sensor
on radar and depth cameras only. Atoyebi et al. [18] presented categorization. Section III details a comprehensive technical
a study on the use of information technology for fall detection review of studies available related to different sensor tech-
and fall prevention. Lapierre [19] have mostly focused on nologies for fall detection systems. Section IV presents the
the statistical representation and technology readiness level of discussion and future scope, and the review is concluded in
different technologies. Delahoz and Labrador [20] focusing section V.
on fall risk factors and general design issues of fall detection
systems. Chaccour et al. [21] discussed general overview of
II. FALL D ETECTION S YSTEM AND
different technologies and emphasized the need for common
S ENSOR C ATEGORIZATION
performance evaluation framework. Additionally, Vallabh and
Malekian [22] have discussed the various technologies and A. Fall Detection Architecture
presented their general limitations. Broadley et al. [23] pre- Uncontrolled dropping or toppling of the body to the ground
sented a specific review focusing on fall detection systems or lower level is classified as fall. This definition has been
applied in the real-world. They emphasized collaboration and the key trigger for most of the studies on the develop-
sharing of real fall data to have robust performance evaluation. ment of fall detection technology. Several medical studies
A summary of recent significant reviews or survey studies is have classified fall risk factors for older adults [24]–[27].
presented in Table I. Broadly all studies agreed on older age, bone fragility, chronic

Authorized licensed use limited to: Auckland University of Technology. Downloaded on October 10,2020 at 06:20:29 UTC from IEEE Xplore. Restrictions apply.
SINGH et al.: SENSOR TECHNOLOGIES FOR FALL DETECTION SYSTEMS: A REVIEW 6891

TABLE I
S UMMARY OF R ELEVANT R EVIEW ON FALL D ETECTION S YSTEMS

Authorized licensed use limited to: Auckland University of Technology. Downloaded on October 10,2020 at 06:20:29 UTC from IEEE Xplore. Restrictions apply.
6892 IEEE SENSORS JOURNAL, VOL. 20, NO. 13, JULY 1, 2020

TABLE I
(Continued.) S UMMARY OF R ELEVANT R EVIEW ON FALL D ETECTION S YSTEMS

Fig. 1. The architecture overview of the fall detection system.

diseases, effects of drugs or alcohol, poor vision, slipping focuses either detecting the critical phase or post-fall phase.
floors, poor lighting, and reduced body balance to be some However, the accuracy of fall detection is primarily determined
common risk factors [28]. Dizziness is one of the signifi- by the pre-fall phase. A daily activity such as suddenly sitting
cant causes of falls. Based on the initial or resting stage, on chairs, deliberate lying on the floor, etc. are also wrongly
fall can be categorized into four categories [29], fall from detected as fall.
standing or walking, fall from sitting, fall from sleeping, and A basic fall detection system performs four significant
fall from standing over height from floor, e.g., ladder, etc. functions: sensing or data acquisition, data processing, clas-
However, the entire fall duration or event is classified into sification, and communication. Fig. 1 gives the architectural
four phases [10], overview of a fall detection system. The first and foremost
• Pre-fall phase: normal daily activities are performed dur- step is data acquisition, i.e., the appropriate physical quantity
ing this phase, such as suddenly sitting, running, walking. is sensed using suitable sensors such as accelerometer, gyro-
• Critical phase: suddenly uncontrolled movement of the scope, pressure, acoustic, cameras, radar or other sensors.
body starts, and the body drops to the floor within a The second step is data processing that involves filtering,
second or two. amplification, background reduction, etc. The third and most
• Post-fall phase: inactivity period and body mostly keep crucial step is feature extraction and event classification. The
lying on the floor or ground. This duration should be features are extracted from the processed signal and based
minimized for reducing the post-fall damage on the features extracted, the classifier algorithm identifies
• Recover phase: body recovers to pre-fall phase on its own the event as fall or non-fall. Different microcontrollers and
or with the help of caregivers personal computers or laptops over a wireless network are
The rationality for the fall detection development system employed for computational tasks such as feature extraction
revolves around the detection of critical phase or post-fall and classification. The conventional wireless communication
phase. Almost all the technical solutions to monitor the fall protocols such as Bluetooth, Wi-Fi, IEE802.11, and ZigBee

Authorized licensed use limited to: Auckland University of Technology. Downloaded on October 10,2020 at 06:20:29 UTC from IEEE Xplore. Restrictions apply.
SINGH et al.: SENSOR TECHNOLOGIES FOR FALL DETECTION SYSTEMS: A REVIEW 6893

Fig. 2. Sensor categories.

are used to process the sensing data from multiple sensors sensors, and a processing unit. The values of yaw, pitch,
remotely. The last step of the system generates an alarm based and tilt angle orientation can be obtained from these sensors.
on the classifier outcome to alert the concerned person. The extracted values are sent to the processing unit for data
processing and event classification. The accelerometer and
gyroscope values are used to find the linear and angular
B. Sensor Categories acceleration as shown in equation (1) and (2) [31],
 
Fall detection systems can be broadly grouped into A = Lx 2 + Ly 2 + Lz 2 (1)
three types based on the location of device deployment:  
(a) Wearable, (b) Ambiance based, and (c) Hybrid sensing W = Gx 2 + Gy 2 + Gz 2 (2)
technologies. The sensor categorization followed in the
review is as charted in Fig. 2. The wearable systems mostly where, A is the linear acceleration, L x , L y , &L z are the
rely on the accelerometer, gyroscope, and magnetic sensors. acceleration along x, y & z-axis, respectively. Similarly,
The ambiance-based systems utilize acoustic sensors, floor W is angular acceleration, G x , G y ,&G z are the orientation
pressure sensors, near field imaging sensors, camera or image along x, y & z-axis, respectively. The magnetometer sensor
sensors, ultrasonic sensors, radar sensors, and PIR sensors, measures the magnetic field for all three physical axes (x, y, z)
while hybrid sensing utilizes a combination of wearable, due to the magnetic effect generated by Earth’s magnetic
ambiance-based sensors, and accelerometer sensors. The force [32]. The magnetic vectors of physical axes extracted
ambiance sensors suffer from blind spots, high installation through the magnetic sensor along with gravity information
cost, and area confinement. Also, image sensors have privacy from the accelerometer are utilized in evaluating the current
issues and are relatively expensive. In wearable devices, short orientation of the subject [33]. When the value of acceleration,
battery life is a primary concern [30]. Moreover, the entire orientation or a combination of both exceeds a threshold
objective is defeated if the subject forgets to wear the device. value, a fall is detected. The same concept is employed
The hybrid approach is characterized by combining different in wearable devices and smartphone systems. These devices
sensors with fusion methods at the data level, feature level, require that the person should always wear them and hence
and decision level. It significantly improves the performance puts the limitation on its usage. The next subsections discuss
of the fall detection system. the literature relevant to wearable sensors.
1) Wearable Sensor: Most of the ambiance-based fall
Appendix shows the review methodology adopted in this
detection systems are proposed for indoor environment and
review paper. Research published from 2008 to August 2019 is
suffer from installation complexity. To address the mon-
reviewed in this study. It is observed that the image-based
itoring requirement in the outdoor environment, a simple
systems and wearable devices are consistently researched each
accelerometer-based wearable fall detection system was pro-
year. These two technologies have high accuracy and are
posed [34]. The schematic architecture of wearable sensor sys-
reliable but come at the cost of privacy (for image sensors)
tems is shown in Fig. 3. Two accelerometer sensors, one at the
and portability (for wearable sensors). The research on other
chest, other at the thigh, were placed. Two-stage classifications
ambiance-based sensors is also gaining momentum, such as
were performed, first, if the accelerometer suddenly crossed a
radar and pressure sensor-based fall detection systems.
threshold, an interrupted signal for five seconds is generated,
and the system is paused. After five seconds, the system
III. S ENSOR T ECHNOLOGIES analyzed the body posture using the sensor’s orientation. If the
A. Wearable Fall Detection Systems x-axis of the sensor aligned with the thigh was zero, then
A wearable fall detection system consists of a set of it is inferred that the subject is standing. Similarly, if the
kinematic sensors like accelerometer, gyroscopes, magnetic x-axis of the sensor placed at the chest was high then lying is

Authorized licensed use limited to: Auckland University of Technology. Downloaded on October 10,2020 at 06:20:29 UTC from IEEE Xplore. Restrictions apply.
6894 IEEE SENSORS JOURNAL, VOL. 20, NO. 13, JULY 1, 2020

transmissions from the sensor device to FPGA based detection

hardware and high-power consumption were challenges
associated with design. The high-power consumption for large
data transmission and associated acquisition by processing
unit was addressed through compressed sensing (CS) [37],
orthogonal matching pursuits (OMP) algorithm was used for
reconstruction of the good quality signal. OMP is implemented
as a hardware core in addition to an earlier FPGA programmed
system [36]. The performance was evaluated only through
simulation and not assessed for real-life implementation.
To design a fall prevention system, the system must be
able to detect the pre-fall scenario instead of post-fall detec-
tion. To achieve this, a pre-fall detection system with three tri-
axial accelerometer sensors was designed [38]. The threshold
values of acceleration, tilt angle and orientation of the sensors
were set as per pre-fall conditions. It was found that the two
sensors were sufficient, and the third was providing redundant
Fig. 3. Working model of Wearable fall detection system. information. The tests were performed with limited data set in
the laboratory environment.
The analysis of the subject orientation, alone on the data
detected, and the system confirms the fall event. To conserve from a 3-axis gyroscope is affected due to the accumulation
battery power, the sensor’s output was sent to a PC through of white. Therefore, the magnetic field values of physical axes
the ZigBee network, and all the processing was performed from the magnetometer and gravity from the accelerometer
remotely. The device needs to be in the communicating area of is applied to correct the error in orientation value from the
the ZigBee network and posed a challenge for outdoor usage. gyroscope. This orientation and vertical acceleration derived
Highly complex human movement is difficult to monitor with from an accelerometer are then used to detect falls. Using
a single accelerometer. A new design was proposed with the same concept, a fall detection device employing a 3-axis
six accelerometers that were placed at the neck, waist, and accelerometer, 3-axis gyroscope, and 3-axis magnetometer was
wrists and thighs [31]. The accelerations amplitude value and designed [39], [40]. The Orientation filter [33], [39], along
the angle with reference to gravity determine the movement with the Madgwick filter [40], was used for signal processing.
posture of the body. Apart from the fall detection, the system The fall detection system suffers from a significant chal-
classifies standing, sitting, lying down and walking also. The lenge of real fall, having wide distinct characteristics as
proposed system used a self-learning approach, where the compared to simulated fall. The simulated falls are on the
classification rules, i.e., the threshold for accelerometer value cushion and have a huge difference in reflex actions and
and angle cluster for sensors were adaptively adjusted as per impacts. Furthermore, young adults usually perform simula-
the user’s motion behavior. The system offered high accuracy, tion contrary to real fall with older adults and have different
but due to many sensors, it was impractical to use in daily life. fall postures. Hence the accuracy and sensitivity in real-life
Different fall detection algorithms offered varied perfor- implementation degrade heavily. To address this issue, real
mance for a waist borne fall detection system. Twenty-one user data is collected using the crowd-sourcing methodol-
different algorithms were analyzed to access the best suitable ogy [41]. After a sufficient crowd-sourced central database is
among them [35]. The root sum of squares (RSS) of the generated, the classification algorithm is adaptively modified.
signal from the tri-axial accelerometer is defined as impact, The users were categorized and grouped as per age, height,
the integration of RSS as vertical velocity, and angle from weight, carrying position of sensors device. Whenever a false
the gravity as posture. These three features impact, velocity, positive/negative was observed, the classification algorithm
and posture were used to analyze the performance of different was modified for better accuracy. The system used a wearable
fall detection algorithm. It was found that a simple algorithm accelerometer sensor to detect falls based on the rapid change
with impact and posture produced reliable results. However, in acceleration of the user. The proposed system improved
the velocity, posture and impact algorithm achieved 100% accuracy.
sensitivity and lowest false positive rate (0.6 per day). The Fall prevention systems can considerably save the injury
tests were performed by young volunteers on the crash mat or impact of falls. A system i-prevent was proposed to uti-
to simulate real-life scenarios. Performance evaluation for a lize two wearable K-band frequencies modulated continuous
real-life scenario might change, specifically for older adults, wave (FMCW) radar sensors mounted on shoes [42]. The radar
where the movement does not occur suddenly. received signal followed a regular distance-to-go pattern for
To avoid the computational complexity of the fall detection the normal non-risk walk. Whenever some obstruction comes
algorithm, a simple field-programmable gate arrays (FPGA) in front of the shoe, suddenly, distance to go decreases, and
based hardware was designed [36]. The complex accelerom- radar mounted on the toe of shoe captures the signal and alerts
eter and angle threshold equation was simplified to addition the user. Similarly, when the person is about to miss the step-in
and multiplication equations. The efficiency of wireless stairs, there is a sudden increase observed in distance-to-go by

Authorized licensed use limited to: Auckland University of Technology. Downloaded on October 10,2020 at 06:20:29 UTC from IEEE Xplore. Restrictions apply.
SINGH et al.: SENSOR TECHNOLOGIES FOR FALL DETECTION SYSTEMS: A REVIEW 6895

the second sensor placed on the sole of the shoe. This sudden design, such as the mobile phones need to be put straight
change classifies the pre-fall scenario. This was the theoretical and in the front pocket of the shirt facing toward the user.
concept and lacked experimental or real-life validation. The application consumed sufficient power as the application
To achieve power optimization, the major challenge required to be running continuously.
observed was to reduce the input data set while simultaneously The limitation of fixing of mobile phone position in the
maintaining reliable accuracy for the single accelerometer- front pocket was addressed through the classification of the
based wearable fall detection system. Data acquisition through posture of using mobile phones [54]. The resultant changes in
compressed sensing along with K-nearest neighbor (KNN) and accelerometer values corresponding to fall from each posture
extended nearest neighbor (ENN) was used [43]. Similarly, were categorized, and different fall signatures were observed.
another approach has utilized the principal component For a fall, three sets of accelerometer values were tested
analysis (PCA) with the multilevel fuzzy min-max neural against a threshold to classify the event as a fall. The approach
network [44]. Though accuracy was maintained, however, offered simplicity, but the power requirement and false alarm
limited success was observed on the power optimization due to limitations of threshold-based algorithms were the main
issue. Another approach of using threshold optimization issues with the design.
and multiple power-saving algorithms on hardware and An E-fall ID system was designed to address the high
firmware were deployed [45]. A significant power saving was false alarm rate by adjusting threshold adaptively [55]. The
observed, but the false alarm rate become high after threshold application was interactive, where the user was asked to
optimization. A single 3-D accelerometer-based fall detection provide age, gender, height, and weight, and accordingly,
system was designed using a neural network classification the threshold observation and processing time window were
algorithm to address the power optimization [46]. The adjusted. When an event was classified as fall, a time delay for
approached increased the accuracy, but power efficiency 10 seconds was introduced before the alarm generation. During
reduced substantially. The position of the sensor on the chest this delay, the user was given the option of canceling the alarm
or waist may not be suitable for all older adults, thus limits in case the event was a daily activity and misclassified as a fall.
the placement of sensors on the wrist as a generalized option. The cancellation was used as data updating in the classification
The wrist-placed wearable devices suffer from low accu- algorithm, and false alarms were dynamically reduced.
racy due to motion artifacts and imbalanced training data The issues related to the threshold-based algorithm were
sets. To address the issue and maintain power optimization, addressed using the holdout method and support vector
classification of the event on the rule-based system was machine (SVM) [56]. Inbuilt kinematic sensors provided
proposed [47]. The publicly available data sets FARSEE- information on acceleration, angular velocity, and orientation.
ING [48] and UMAFall [49], along with synthetic minor- Fisher’s discriminant criterion along with the J3 criterion was
ity over-sampling technique (SMOTE) was used for training used for feature selection. The accuracy was improved up to
the classifier and balancing the data set for classes (fall or 95.03%. However, the approach puts a limitation on mobile
non-fall). The rule-based system approach provided reliable phone placement, such as the phone need to be worn on the
accuracy with good power optimization. However, the wrist- chest or placed in the pocket.
orientation affected accuracy. The limitation of threshold- The KNN classification algorithm was used to address the
based methods was addressed through the use of Madgwick’s issue of low accuracy of the threshold-based algorithm [59].
decomposition and machine learning methods for classifi- The signal features in time, (average, peak) time-statistical
cation algorithms. High accuracy was achieved through the (Skew, Kurt), and time-frequency domain (Wavelet energy,
approach, but the computational load & power requirements peaks) were extracted for accelerometer values. KNN used
increased substantially [50]. A two-segment feature extrac- these features for the classification of the event. To further
tion method that employs time characteristics of falls and reduce the false alarm rate, the user was given a time window
neural network-based classification algorithm was used to in which the false positive could be signaled. This false
reduce the computation load [51]. The low sampling frequency positive information was used to train the classifier adaptively.
reduced the power consumption also. The power optimization, An algorithm was proposed to adaptively reduce the sam-
24/7 wearing of the devices, and limitations of not being pling rate during the inactivity period of the accelerometer,
used in some circumstances such as restroom, bathrooms were which reduced the power consumption. In another study
significant challenges associated with the technology. Table II posture change, i.e., angle information along with acceleration
summarizes the literature review of wearable devices. was used to address low accuracy for threshold-based algo-
2) Smartphone Sensor: The essentiality of the mobile phone rithms [60]. The tilt angle changes more than a value (60◦)
in everyday life, for all ages of people, makes the pathway for was used to locate the time window of the signal for further
developing a standalone mobile application for fall detection. processing. In this time window, the tilt angle change pattern
The smartphone utilizes built-in sensors such as accelerometer or features were extracted, and weighted Hamming distance
or gyroscope [52]. The input from these sensors was taken, was used for pattern matching. Further, the accelerometer
and an application for fall detection was designed [53]. The values were observed, and if crossed a threshold, the event
accelerometer values are monitored, and if the value crosses was classified as fall. The approach was based on posture
a threshold and does not return to its usual value within a set change and hence, independent of weight, age, gender making
duration, the event is classified as fall. The prototype detected the setting of threshold easier and generalized. In addition to
fall satisfactorily, but there were certain limitations with the acceleration and orientation, angular velocity was also used

Authorized licensed use limited to: Auckland University of Technology. Downloaded on October 10,2020 at 06:20:29 UTC from IEEE Xplore. Restrictions apply.
6896 IEEE SENSORS JOURNAL, VOL. 20, NO. 13, JULY 1, 2020

TABLE II
W EARABLE FALL D ETECTION S YSTEMS

for better accuracy [61]. SVM was used as a classification The fine-tuning of the threshold for classification
algorithm for features extracted from the accelerometer and algorithms in smartphone-based fall detection systems
gyroscope signal. induces a tradeoff between false negative and positives.

Authorized licensed use limited to: Auckland University of Technology. Downloaded on October 10,2020 at 06:20:29 UTC from IEEE Xplore. Restrictions apply.
SINGH et al.: SENSOR TECHNOLOGIES FOR FALL DETECTION SYSTEMS: A REVIEW 6897

TABLE III
S MARTPHONE -B ASED FALL D ETECTION S YSTEMS

Standalone machine learning approaches require high 1) Image Sensor: In the image sensor-based system,
power consumption. A two-step approach based on a the subject is monitored in real-time through the camera, and
low computation load threshold-based algorithm followed by images are processed using various algorithms for posture
multiple–kernel learning support vector machine (MVL-SVM) identification. The choice of camera, its position, and the video
was used [62]. compression method are the key features affecting the perfor-
The smartphone-based design produced good results in mance of an image-based fall detection system. The image
terms of accuracy and simplicity. However, the requirement of sensors are subclassified into RGB/IP and depth camera. The
fixed orientation and position of mobile phone, unsuitability architecture of image-based systems is shown in Fig. 4.
in bathrooms, and frequent charging of the device are major a) RGB/IP camera: The popularization and low cost of web
challenges for the technology. Table III relates the merits and cameras have provided a way for designing a fall detection
limitations of the works associated with smartphone-based system through cameras. Though the use of silhouette in
systems. image processing has reduced the privacy concern to some
extent, yet the challenge to differentiate between a real fall
incident and a person lying deliberately is a major challenge
for camera-based fall detection systems. The critical time
B. Ambiance Fall Detection System difference between real fall and deliberate lying was observed
The ambiance fall detection system includes an image and used in the design to address the issue [63]. The human
sensor, acoustic sensor, pressure sensor, infrared sensor, radar body silhouette is adopted, and the bounding box height
sensor, or a combination of them. The sensor is installed to width ratio was used as a signal feature. As silhouette
in the area to be monitored. Each sensor uses a different removed most of the human body features, therefore upper
technological concept for the detection and implementation of limb activities affected the posture identification. The vertical
the fall system. When a person interacts with the environment projection histograms and statistical analysis of the silhouette
and comes in the proximity of the sensor zone, the sensor image were deployed to remove the upper limb activities’
detects the abnormal activity, and based on the sensed signal, effects. KNN was used for classifying the event.
categorizes the fall event, as shown in Fig. 1. The following The bounding box methodology of a vision-based system
subsections discuss different ambient sensors, its working requires the cameras to be placed sideways but suffered
principle, and related literature. from occlusion. Multiple camera-based approaches require

Authorized licensed use limited to: Auckland University of Technology. Downloaded on October 10,2020 at 06:20:29 UTC from IEEE Xplore. Restrictions apply.
6898 IEEE SENSORS JOURNAL, VOL. 20, NO. 13, JULY 1, 2020

Fig. 4. Schematic representation of the image-based fall detection system.

calibration, and most of the methods suffer from the inability the event using cyclic HMM [73]. The gait disorder detection
to differentiate between sitting down abruptly, deliberate lying accuracy was dependent on camera placement. The multiple
and real fall. To address the challenges a fall detection camera-based approaches require calibration, and the depth
system using one non-calibrated IP camera with a shape camera has narrow fields of view with limited depth. The
deformation analysis was designed [64], [65]. During the fall, feature extraction techniques have a limitation of dependency
the human shape changes drastically as compared to normal on the signal-to-noise ratio (SNR), high processing, etc. A fall
daily activities. For shape deformation analysis, first canny detection system using optical flow images as input to convo-
edge detector extracts the silhouette edge points. The shape lutional neural networks (CNN) was designed to address the
context matching [66] followed by fall Procrustes distance [67] above issue [74]. The optical flow algorithm [75] represents
provides the shape analysis. A peak in the fall Procrustes the patterns of the motion of objects by calculating the
distance curve is taken as a fall feature. One class Gaussian displacement vector fields between two consecutive images.
mixture model (GMM) trained with normal activities was The CNN works as an automatic feature extractor was fed
used for classification of the event. To reduce the false alarm, with optical flow images as input and classified the event.
post-fall lack of significant movement was added as a feature The use of optical flow images made the system independent
for threshold. The design was robust and produced accurate of the environmental condition. The design required a heavy
results, but the computational load was high as compared to computational load and posed a significant challenge. The
other vision-based methods. conventional two-dimensional convolutional neural network
The image processing issues such as shadowing, high com- requires a large dataset to train a classifier and is influ-
pression, and occlusion were dealt through the design of a enced by the image quality. To address the issue, a three-
fall detection system based on head tracking. The 3-D head dimensional CNN was employed, which can work with smaller
image was projected as a 2D ellipsoid, and the particle filter datasets [76]. The approach increased the computational cost
technique was used for tracking the ellipsoid [68]. The high as a major challenge for the system design.
velocity of the head and the distance from the ground is lower The summarized description of the merits and limitations of
than a threshold triggered the fall alarm. The privacy invasion the works related to RGB/IP image-based devices is presented
was the main challenge observed during adaptability check in Table IV.
with older adults. The neural network classification algorithm b) Depth Camera: The depth camera provides spatial
was used on simple silhouette extraction and bounding box information where a mono-camera vision system lacks this
ratio approach for a fall detection system [69]. The system feature. The use of multiple cameras may resolve the issue but
was simple and sufficiently robust but suffered from biased requires calibration to relate the camera coordinate system to
training data and occlusion. Furthermore, neural networks a practically useful coordinate system. It was observed that the
are computationally expensive; therefore, employing it in a fall images have a specific spatial-temporal signature. Using
standalone ambient system would be expensive. the observation, a fall detection system based on a spatial-
The fall detection system based on motion history imaging temporal signature using a convolutional neural network was
(MHI), which extracts the motion form video sequence, suffers designed [77]. The use of spatial-temporal features provided
from limitations such as motion self-occlusion, motion over- high accuracy than the silhouette-based approaches.
writing, and dependency on the temporal duration value. These The spatial information is useful in fall detection systems,
issues were addressed through a timed motion history image and the use of a depth camera also addressed the issue of
(tMHI) [70]. The use of tMHI was extended [71] through the high sensitivity to illumination variation and privacy protection
extraction of 7hu’s feature [72] for gait disorder and classified to some extent as only skeleton information of the image is

Authorized licensed use limited to: Auckland University of Technology. Downloaded on October 10,2020 at 06:20:29 UTC from IEEE Xplore. Restrictions apply.
SINGH et al.: SENSOR TECHNOLOGIES FOR FALL DETECTION SYSTEMS: A REVIEW 6899

TABLE IV
RGB/IP I MAGE -B ASED FALL D ETECTION S YSTEMS

processed [78]. The head joint tracking approach by SVM Thermal sensors were used for addressing the privacy issue,
and mean value matching of the image was used to classify especially in toilet rooms. The thermal image was divided into
the event. The accuracy of the RGB-based system was further subareas, and the temperature intensity variations near the floor
improved using a depth camera with five fall detection fea- area were termed as fall events [83]. The subjects to follow the
tures [79], [80]. The five features were aspect ratio, change regular pattern of using the room as trained to the classification
in aspect ratio fall angle, the velocity of the center of mass, algorithm. The design was limited to a maximum temperature
and the speed of the head. These features are set as input of 31 ◦ C.
to SVM for classification of the event. The designs suffered A time-of-flight (TOF) 3-D camera, along with a self-
from occlusion and undetected event when a person falling calibration algorithm, is used in [84], where the 3D camera
on furniture classify the event. The designs suffered from utilizes the phase-shift of the infrared modulated signal in
occlusion placed higher than the floor plane. getting the depth information of the subject. The phase shift is
To capture human fall from the bed, depth camera observed during the round-trip from the device to the subject
along with large margin nearest neighbor classification was and back. The self-calibration automatically detects the floor
used [81]. The depth camera along with FFT encryption for surface in the image. The issue of false alarm due to the
transmitting and processing the images was used to address the deliberate lying on the floor was a major challenge to the
low accuracy under low light conditions and privacy issues of design. A simple RGB-D camera minimizes the processing
RGB camera, respectively [82]. complexity and high cost. The system used a hidden Markov

Authorized licensed use limited to: Auckland University of Technology. Downloaded on October 10,2020 at 06:20:29 UTC from IEEE Xplore. Restrictions apply.
6900 IEEE SENSORS JOURNAL, VOL. 20, NO. 13, JULY 1, 2020

model (HMM) for designing a fall detection system [88]. non-fall based on the delay reaching the sensors. Using Mel-
Three features; the center of mass, vertical speed, and standard frequency cepstral coefficients (MFCC) along with KNN was
deviation from the moving points of the 2D image were deployed to classify the event as fall or non-fall. The system
extracted using component labeling and Kalman filter. These produced a substantially high false alarm rate of 5 alarms
features were used by HMM for the classification of events. per hour. Sound source height was used as a threshold to
The system was simple and inexpensive. However, it suffered reduce the false alarm rate [93]. For the signals arriving
from occlusion and biased due to the training data set of from above 2 feet height, the signal was labeled “non-fall,”
healthy young people. Post-fall detection does not avoid and no further processing was performed. The sound height
the injuries and first aid requirements. The pre-impact fall information improved accuracy from 30 false alarms per hour
detection integrated with an on-demand fall protection system to 5 false alarms per hour. The prototype design of the acoustic
(inflatable nip protector or wearable airbag) may address the sensor for fall detection was demonstrated in a controlled
issue [89]. The essential requirement of a pre-impact fall environment. The real-life implementation tests, reduction of a
detection system is a longer lead time for response. If the false alarm, training the classifier for a different type of people
imbalance initiation is detected, the longer lead time may on different fall surfaces, performance evaluation with varying
be achieved. A pre-impact fall detection system based on sources of noise (e.g. TV, radio) are the main challenges
a zero-moment point (ZMP) criterion and dynamic support toward its feasibility.
area (DSA) was proposed. ZMP is defined as a point on The realistic fall sound was difficult to simulate even by the
the ground where the net moment of inertial and gravity stunt actors as the instinct of humans is to avoid painful fall.
forces has no values along the axis on the ground. The area Fall dummies (Rescue Randy) cannot generate realistic full
between the projections of both feet was defined as DSA. sound because of the difference in skin, bone density. This
From the images, a skeletal model of the human was obtained, was the biggest challenge observed for acoustic-based design.
and the ankle, knee, other joints coordinates provided center Two possible solutions; the fuzzy rule where no training data
of mass. Through the coordinates and extraction algorithm is needed, and other classifiers trained with non-fall data were
relation between ZMP and DSA for balance, the condition explored [93]. Motion detection was used in addition to sound,
was obtained. Through the relation imbalance state or pre- and movement observed within a set time would classify the
impact fall condition and by velocity and height of the center event as non-fall, reducing the false alarm rate. The fuzzy
of mass, fall event was classified. The setup was laboratory rule was developed based on the observation that the spectral
limited, and real-life implementation and high computational energy of the fall signal lies in the low-frequency range (below
load were challenges associated with technology. 300 Hz). The idea of sound height reduced the false alarm
The occlusion and issue of falling in the direction of the rate but increased the computation load. Moreover, the height
camera were dealt with the use of the depth camera. A fall estimation accuracy is affected by the acoustic properties of the
detection algorithm was developed utilizing the major axis environment and the conical beam shape. Eight microphones
orientation of subjects [90]. The mean position of different circular array was used to address the sound source localiza-
points in the skeleton of the body as taken to calculate the tion of acoustic sensors [94]. The spectrum response power-
major axis. The spine distance from the estimated floor was phase transform (SRP-PHAT) [95], [96] is used for sound
derived. A threshold-based algorithm analyzed the orientation localization along with earlier test setup and classification
of axis and distance of spline as parallel to the floor to algorithm [97].
decide an event as a fall. The design shows good accuracy; The height estimation accuracy of the sound source was
however, tracking algorithm analyzed the orientation of axis affected profoundly by environmental noise such as TV, radio,
and distance of spline as parallel to the floor to decide an phones. The beamforming (BF) technique to improve the
event as a fall. The design shows good accuracy; however, classification performance was used to mitigate the challenge
tracking error was introduced whenever the person leaves of background noise and strong interferences [98]. The steered
the frame of the time window. The depth image sensors response power SRP-PHAT algorithm improved the localiza-
address the occlusion and privacy issues to an extent. Yet tion accuracy followed by simple delay and sum, and the
the limited trust over data leakage on social platforms and beamforming (BF) technique was applied to the earlier used
related companies, the adaptation of camera-based devices is experimental setup [94] to have a better-quality signal for
still an open challenge, especially in certain areas such as detection of an event. Though the interference due to known
bedrooms/bathrooms, etc. Table V summarizes the merits and sources was significantly reduced with the BF technique, yet
limitations of the works related to depth image-based devices. the real-world interference from unknown sources was not
2) Acoustic Sensor: Acoustic sensors are used to differen- addressed. The approach failed when the fall comes from the
tiate different sound classes, e.g., baby noise, breaking glass, close proximity of interference. The BF approach also did not
door sound, human sound. These sound classes were extended address the issue of occlusion of fall.
to identify fall and non-fall activities. Non-negative matrix factorization tool for blind source
A vertically arranged linear array of three acoustic sensors separation was used to address the challenges of the BF
(microphone), (named as FALL DEtector, FADE) was used to approach [99]. The phase information of multichannel was also
capture sound energy [91]. The energy of the acoustic signal utilized in making the system more robust against background
was tracked in a time window of one-second duration, and noise and interferences. The design was made simpler by using
test against a threshold. The signals were identified as fall or 4 Microsoft Kinect microphone sensors in random order. This

Authorized licensed use limited to: Auckland University of Technology. Downloaded on October 10,2020 at 06:20:29 UTC from IEEE Xplore. Restrictions apply.
SINGH et al.: SENSOR TECHNOLOGIES FOR FALL DETECTION SYSTEMS: A REVIEW 6901

TABLE V
D EPTH I MAGE -B ASED FALL D ETECTION S YSTEMS

avoided the installation complexity of circular array. In the interference. Two microphones along with interferences as TV
most challengeable acoustic scenario of −5 dB to -10 dB SNR, sets were used, and performance was found better than earlier
the accuracy of the system improved. The design, however, systems [93], [99]. The sound source localization challenge
suffered from low accuracy when the fall comes from the same and noise interference of environment sound were addressed
direction of interference (TV, radio), or the fall occurs outside through a floor acoustic sensor. The sensor has a membrane
the field of view. and resonance chamber generating an electrical voltage when
The difficulty in getting realistic fall data set for training the membrane vibrates. A high SNR signal was obtained by
was addressed by using a one-class classifier [100]. The passing the signal from a high pass filter as the maximum
background noise and interference were removed through energy of fall signals is concentrated below 1.5 kHz
the two-stage source separation technique. MFCC features This improved the detection accuracy of the acoustic fall
from non-fall sounds were used for training the OCSVM detection system. The computational complexity of the MFCC
(one-class SVM) classifier, which classifies the foot sound feature extraction process was reduced through represent-
from abnormal (fall) sounds to classify the event as fall. In the ing the acoustic signal through acoustic local ternary pat-
first stage of source separation, the reverberation in signal terns [101], and SVM classification with the Hidden Marker
due to sound reaching the microphones through multiple model (HMM) was used for fall event detection [102].
paths was de-reverberated. The de-reverberation was achieved The efficiency of an acoustic-based fall detection system
through a spectral subtraction method. The source separation depended on the accuracy of sound source localization. The
technique was used in the second stage to remove noise and environmental noise effect and the issue of fall coming close or

Authorized licensed use limited to: Auckland University of Technology. Downloaded on October 10,2020 at 06:20:29 UTC from IEEE Xplore. Restrictions apply.
6902 IEEE SENSORS JOURNAL, VOL. 20, NO. 13, JULY 1, 2020

TABLE VI
ACOUSTIC S ENSOR -B ASED FALL D ETECTION S YSTEMS

Fig. 5. Architectural model of a floor pressure sensor-based fall detection systems.

from the same direction of interference remained unaddressed. transmitted through the floor. Also, the vibration signal pro-
The realistic fall sound simulation was the main challenge duced by human fall differs significantly from the fall of
associated with technology. The technology was limited with other objects. The piezo element is one of the choices among
a hypothesis of only one person in the room, and even the issue pressure sensors. It works on the principle that when a force is
of large pet presence remained unaddressed. The summarized applied, it deforms and a resultant change in its atomic struc-
description of the merits and limitations of the works related ture causes the transfer of charge. A voltage signal is induced
to acoustic sensor-based systems is presented in Table VI. between the two poles of the piezoelectric sensor proportional
3) Floor Pressure Sensor: The impact of the human fall to the applied pressure. An overview of the floor-pressure
on the ground generates measurable vibrations that are sensor is shown in Fig. 5. A piezoelectric transducer was

Authorized licensed use limited to: Auckland University of Technology. Downloaded on October 10,2020 at 06:20:29 UTC from IEEE Xplore. Restrictions apply.
SINGH et al.: SENSOR TECHNOLOGIES FOR FALL DETECTION SYSTEMS: A REVIEW 6903

used to generate an electric signal corresponding to floor sensor-based design was used to address this challenge, where
vibration [103]. The vibration pattern (frequency, amplitude, the differential pressure detector chambers were outside the
duration, succession) matches with the human fall pattern, carpet [107]. More than one person walking simultaneously
the design detects a fall. Anthropomorphic dummies were used over the sensor-generated a false alarm and posed a challenge
for simulating falls, and due to similar mass distribution as of for the design. Furthermore, such carpets cannot be utilized in
humans, the dummies provided an excellent representation of certain areas such as the stair, restrooms.
human fall. The different vibration amplitudes were observed A low cost unobtrusive and waterproof fall detection system
for mezzanine and concrete slab floors. This required the based on a pressure-sensitive fiber sensor was designed [108].
setting of threshold differently and produced a challenge When pressure is exerted on the fiber, the received voltage
making the design dependent on the type of floor. The sensor is changed. The grid of such fibers was deployed on the
placement no closer than a minimum distance to floor posed floor. Each grid represented a pixel of the complete floor
another challenge as then it may trigger false alarms. The image. The back-projection algorithm was used to construct
design was tested under laboratory set up, and no real field the floor pressure image by distributing and summing the
trials were conducted. The low impact or soft falls were not voltage received by the fiber to pixels uniformly. The threshold
detected by design. The limitation of the lack of explicit differ- applied to the number and value of non-zero pixels classified
ence between falls and similar activities like jumping, sudden the posture or fall events. The scalability, reduced installation
lying triggered alarms was addressed through the use of seis- complexity, water resistance, and independency on type of
mic sensor network [104]. Footstep categorization, vibration floor were some essential prerequisites for designing a robust
signatures both in time and frequency domain were analyzed floor pressure-based fall detection system. Piezoelectric bands
and used as a feature for the classification algorithm. Though of 60 cm width and variable-length as per installation were
the approach increased specificity, the real-life trials and instal- used to address the scalability issue [109]. The bands, the thin
lation complexity was the major challenge of system design. piezoelectric film was placed under the floor covering. When
The issue of no real field trials was little addressed through the pressure was exerted, each band emitted an electric signal,
deploying a piezoelectric grid-based floor pressure monitoring and the signal from all the bands was summed up. The Fourier
system in a lab “The future care lab” [105]. This deployment transform and temporal information of signal were used as
provided realistic data for training the classifier. The living lab a feature, and the random forest algorithm was chosen for
approach was interactive with the user, and iterative training classification based on its ease of implementation. Though
data sets were obtained. The normal activity signals were the laboratory-based data analyze provided good accuracy,
continuously recorded, and the dataset was updated. Wherever yet the real-life implementation and the biased data set for
the fall happened, the event was classified, and the user classification were the critical challenges of design.
notified false alarms through an interactive wall attached to the The scalability of a smart carpet or mat was achieved
floor. This process made the training interactive and improved through stitching the secondary or tertiary mat with the pri-
accuracy. mary mat. This stitching was done by piezo-resistive fabrics
The gap in the real-time experiments, deployment perfor- but posed a challenge as a slight error results in considerable
mance and assess the adaptability to real-life was addressed degradation in the accuracy. An easily scalable mat was
using a high sensitivity micro-electromechanical accelerome- designed using “velostat” sandwiched between two grids of
ter (MEMS) based floor sensor [106]. The performance was aluminum tapes placed orthogonal to each other [110]. The
evaluated in a real-life deployment for a total of 507 days grid points made by aluminum tapes above and below of
in 15 types of floors in living flats. Though no real fall velostat were conductive only when the pressure was exerted
occurred, and the sensitivity of the system was evaluated on over velostat, and the resultant resistance drop had occurred.
a false positive basis. The variations in floor types produced When a person falls or moved over the mat, the floor pressure
different fall signatures for the same person. The accuracy image is constructed, and the classify on the pattern matching
of floor pressure-based technology depends on the type of principles. To address the issue of large pets, a face recognition
floor, and this key challenge was observed in the long-running algorithm was also used through deploying a web camera.
experiment. The common classification algorithm for all types Nevertheless, the introduction of the camera made the system
of floors produced more inadequate results. The classification intrusive and complex.
threshold is optimized according to the type of floor to Though the design addressed the issue of scalability yet
decrease dependency on the type of floors. The approach the common challenge of floor-based technology such as
resulted in better performance; however, it would require undetected low impact or soft fall, large pets triggering a
manually tuning for different types of floor. The main issue of false alarm, real-life implementation, and particular case while
floor type dependency on the accuracy remained unresolved. designing smart mats and accuracy dependency on the type
Another approach used to tackle the floor dependence chal- of floor remained unaddressed. Table VII summarizes the
lenge is to embed a pressure sensor inside the carpet, called floor-based fall detection systems.
smart carpets. The adaptability of the floor pressure based fall 4) PIR Sensor: The human body emits the Infra-red (IR)
detection system realized as smart carpets are largely affected radiation. The crystalline material of the passive infrared (PIR)
by the maintainability of the sensor under the floor. The force sensor generates a surface charge when exposed to heat in the
resistive or capacitive sensors are implanted inside and impose different IR patterns and track human motion. The form of
difficulty in maintenance. Differential piezo-resistive pressure IR. Therefore, PIR sensors are deployed to detect working

Authorized licensed use limited to: Auckland University of Technology. Downloaded on October 10,2020 at 06:20:29 UTC from IEEE Xplore. Restrictions apply.
6904 IEEE SENSORS JOURNAL, VOL. 20, NO. 13, JULY 1, 2020

TABLE VII
F LOOR P RESSURE S ENSOR -B ASED FALL D ETECTION S YSTEMS

but with a delay as compared to the upper sensors. A set of

observations were recorded, and with this training data set,
HMM was used for classification of the event. The laboratory
set up presented the proof-of-concept and could be useful for
tight spaces such as a bathroom. The trade-off between the
detection accuracy and range and dependence of the placement
of the sensor surfaced as the challenges to the design of the
real-life implementation.
The PIR sensors produce low dimensional temporal data
stream without sensing the redundant background, and conse-
quently, the high dimensional processing is avoided. However,
this low dimensional sensor output is deprived of spatial
information, and therefore, low accuracy of the design is
observed. To address the issue and increase the spatial infor-
mation, reference structure tomography was used to design
the spatially aware fall detection system [112], [113]. The
Fig. 6. Working principle of PIR based fall detection system. reference structure tomography segmented the object space
into a lot of sampling cells by using a Fresnel lens and opaque
mask, covering PIR sensors in a defined pattern. Different
principle of PIR systems is shown in Fig. 6. The requirement human motions such as sitting, standing, walking, and fall
of non-invasive, unobtrusive, and passive fall detection system produced different output signatures for sampling cells. These
was addressed by utilizing PIR sensors [111]. The fall has a patterns were classified with the two-layer HMM classifier.
unique pattern where the human motion shifts gradually from The system had advantages of low- dimensional processing,
a height towards the ground in the vertical axis, as shown in immunity to illumination condition, and privacy awareness.
Fig. 6. This pattern was explored by deploying PIR sensors The design was accurate as compared to the first PIR based
vertically on the wall. The motion shift during fall activated design by Sixsmith and Johnson [114], where the detection
the PIR sensors due to variation in the signal strength near was based on vertical velocity information and could not
the ground. The lowest sensors observed maximum strength discriminate the real fall from similar daily activities.

Authorized licensed use limited to: Auckland University of Technology. Downloaded on October 10,2020 at 06:20:29 UTC from IEEE Xplore. Restrictions apply.
SINGH et al.: SENSOR TECHNOLOGIES FOR FALL DETECTION SYSTEMS: A REVIEW 6905

TABLE VIII
PIR S ENSOR -B ASED FALL D ETECTION S YSTEMS

The PIR sensors are direction sensitive, and the side view The sensing mechanism works on the sample cell segmen-
sensor deployment [113] could only detect fall efficiently tation of object space using RST.
if occurring perpendicular to the Field of view (FOV) of The sample cell segmentation is carried out using
the sensor node. This limitation was also addressed through multi-mask over PIR sensors. The multi-masking enhanced the
sampling cell segmentation of object space and measuring the spatial information and coverage of the side view of vertically
thermal variations at the boundaries of sampling cells. The installed PIR sensors. The design achieved higher sensitivity,
ceiling-mounted sensor node avoided the occlusion issue and but the noisy output of the PIR sensor, the requirement of
made the design, view independently. The sensing cone of high pre-processing, design limitation to single person fall
sensor node was of 3m radius, and to achieve ubiquitous moni- and performance evaluation in real-world deployment were
toring scalability was required, then the fusion of sensor node’s the major challenges that were not addressed. Table VIII
outputs as needed. The performance evaluation of the system summarizes PIR-based fall detection systems.
after the fusion of sensors outputs was not addressed. Due to 5) Near Field Imaging Sensor: Sensing the human body
the Fresnel lens geometry, PIR sensors have a 3-D pattern of impedance by the electric field has found limited application in
wedge-shaped detection zones with embedded dead zones. The the biomedical or electronic interface systems. A method was
sensors are not sensitive to the line of sight movement along demonstrated for human tracking with electric field sensing,
the axis of the sensor’s field of view. Moreover, the detection which is invisible to the user, scalable, and low cost [117].
sensitivity also depends on the distance of the subject from A thick film sensor matrix made of the metal element of
the sensor. 300 × 300 mm was used. The primary metal element is
The human movement detection through the PIR sensor fed with a low frequency, low voltage signal, and the rest
depends on the recognition of subtle temperature differences of others (secondary) were grounded. The electrical coupling
between the human body and the environment. The PIR sensor between these primary and secondary elements varied with
output is low dimensional, noisy, and if not taken care of, a human presence on the top of the sensor. This electrical
would produce erroneous results. To address the issue, a fall coupling variance resulted in the current flow between the
detection system using median, Gaussian, and wavelet filters sensor element and the film matrix was scanned sequentially to
for data preprocessing and neural network for classification track the human. The proposed system is validated for walk
were designed [115]. The coverage of the sensor depends and fall. The performance of the system in the real world
on deployment and in turn, affects the efficiency of a PIR and actual usability remained a challenge to the designed
based fall detection system. The coverage of the sensor area system. The location of the person was approximated through
was efficiently increased by using reference structure tomog- the coordinate of activated sensors weighted by the signal
raphy (RST) along with stereo-view based infrared radiation strength. This estimation is sufficiently accurate when the
changes (IRC) sensing mechanism [116]. Each person behaves person is standing because the centroid of observation is
as a distributed infrared source, and the geometrical shape of between the legs. However, when the person is walking,
a person determines the distribution. Therefore, the activity the observation of centroid becomes noisy as only one foot is
pattern can be inferred out from the induced IRC pattern. contributing.

Authorized licensed use limited to: Auckland University of Technology. Downloaded on October 10,2020 at 06:20:29 UTC from IEEE Xplore. Restrictions apply.
6906 IEEE SENSORS JOURNAL, VOL. 20, NO. 13, JULY 1, 2020

TABLE IX
N EAR F IELD I MAGING S ENSOR -B ASED FALL D ETECTION S YSTEMS

An accurate reference positioning system was designed to 6) Radar Sensor: The radar transmits electromagnetic waves
improve accuracy and observation [118]. An experimental set at certain frequencies, which is reflected from the moving
up was made for reference position based on triangulation object, a frequency shift is observed in the reflected wave. This
methodology. The Kalman filter was used for estimating the frequency shift depends on the velocity of the body. Different
position of the subject. The test was performed for two people body movement generates distinct Doppler signatures for a
walking randomly for three minutes. Though the introduction person, and hence fall characteristics are obtained through
of the reference position system improved the accuracy, sup- these signatures. The Doppler signatures or the Doppler fre-
pressed the noise, yet some challenges remain unaddressed, quency shift for a body moving with velocity v are observed
such as tracking at high speed. as in equation 3 [121],
A fall detection system was developed and deployed in
2 fc v
Helsinki, Finland [119]. The design also addressed the issue fd = cos θ (3)
of battery charging for wearable devices, privacy concern with c
camera-based devices, insensitivity of low impact fall, and where c is the propagation speed of light, fc is the carrier
furniture shading of floor pressure-based system. The signal frequency of the radar, and θ is the angle between the velocity
features were taken as the number of observations or cells, vector of the target and the line-of-sight of the radar. When
the longest dimension of cell coordinates, and the sum of the the angle approaches 90◦ , the Doppler signature is strongly
magnitude of cell voltages. To reduce the noise effect, the pose degraded as it depends on the cosine of the aspect angle, hence
of the person was estimated using Bayesian filtering, and state the classification performance is severely affected. The multi-
evolution was modeled as two-state Markov Chain. Sitting, ple radars can address the mentioned issue. The architecture
falling on chair or bed, sliding against a wall were detected overview of radar-based fall detection systems is shown in
with poor performance as near field imaging (NFI) has a range Fig. 7. The primary challenge of any fall detection system is to
of 5 cm above floor. minimize false fall alarms and detecting all the true falls. In the
The interference due to conducting an object or electrical radar sensor, the placement of the sensor critically affects the
appliance near the floor affects the performance. Water seepage performance of fall detection systems. The floor-based radar
issue in the bathroom under the floor majorly degrades the sensors are not sensitive to the falls lateral to its central axis.
performance and calibration of reference positioning remained However, the floor sensor is useful in analyzing the fall risk by
the critical challenges for the technology. To address the monitoring the gait motion of the subject. While the sensors
mentioned challenges of NFI technology, a system was devel- installed on the ceiling are less sensitive to the falling direction
oped based on capacitive proximity sensing and threshold as there will always be the downward component of the radar
algorithm [120]. A matrix of thin planar electrodes was placed signal. To address the issue, two range control radar (RCR)
under the floor, and capacitive change due to a human on sensors were placed orthogonal to each other on the floor.
the electrode was measured to compute the human position. The configuration helped in reducing the false alarm [122].
A stable voltage of 1.8 V was observed without human By placing the second sensor on the ceiling, the perfor-
contact on the sensor as the capacitance is negligible. When mance can be improved further. MFCC, along with SVM and
human occupies the sensor, a 5 V is observed. A threshold- KNN was used for classification of the fall event. Doppler
based algorithm was used to classify the occupancy. Though radar operates efficiently in line-of-sight conditions, and due
the NFI floor sensor is completely untraceable, immune to to spatial variance, their performance degrades in multipath
furniture shading, and needs low computing power. It has and non-line-of-sight (NLOS) [123]. Particularly, the radio
the advantage of privacy awareness and better adaptability, wave easily gets reflected by obstacles in high-frequency
but the challenges of NFI technology such as unclear fall, bands such as 24 GHz. The temporal-spatial characteristic
electrical or conducting devices interferences, limitation due of the signals using an array sensor was utilized to enhance
to deployment environment, real-life experimental verification the performance. The subspace method was used to make the
remained unaddressed. Table IX describes the merits and design spatially invariant [124]. The approach reduces the data
limitations of the works related to near field imaging systems. correlation matrix into orthogonal signal and noise subspaces

Authorized licensed use limited to: Auckland University of Technology. Downloaded on October 10,2020 at 06:20:29 UTC from IEEE Xplore. Restrictions apply.
SINGH et al.: SENSOR TECHNOLOGIES FOR FALL DETECTION SYSTEMS: A REVIEW 6907

SVM is the most common algorithm used for classification

of the fall event. The size of the training data sets controls
the requirement of a number of support vectors, and it grows
linearly, which makes it inappropriate for processing large
data sets [129]. Sparse Bayesian learning methods based on
the relevance vector machine (RVM) was used that improved
the performance as compared to the SVM utilizing fewer
relevance vectors. Since the radar detects the radial velocity
component of the moving object, therefore, sensor position
and fall direction affects the performance of the system.
A real-living environment poses many challenges for fall
detection like antenna configuration, building material, fur-
niture shading, pets, and multi-person presence. To evaluate
the performance of a radar-based fall detection system with
reference to the radar positioning in a real-life environment,
two radars were deployed in six apartments [130].
One radar sensor was placed on the floor for optimal
gait detection for fall risk and others on the ceiling for fall
detection. The slotted foil was wrapped around the floor
RCR to exclude the potential sources of false alarm such as
Fig. 7. Working principle of radar sensor-based fall detection System.
moving or vibrating objects like fans and the electric field
formed by high voltage equipment and motors. The slotted
via the eigenvalue decomposition. A 24 GHz continuous- foil was not used for ceiling radar to allow for a wider FOV.
wave Doppler radar and an array of sensors, 800 MHz band Feature alignment along with MFCC coefficients was used for
dipole antenna is used. The correlation between eigenvectors, classification. The performance variation in false alarm due to
eigenvalues, and RSS is taken as a signal feature from an array the elderly using walkers, pets jumping off beds and couch,
of sensors. MFCC is extracted from the radar signal, and then posed a challenge for the design.
SVM is used for classification for both Doppler radar and an As the angle between the object-motion trajectory and the
array of sensors. radial path between the radar and the object approaches 900 ,
For scalability and better adaptability, the radar-based the performance of the system decreases rapidly. This direction
devices must satisfy the requirement of low processing and the dependence and occlusion challenge of radar-based systems
spectrum mask. To achieve this, the signal processing is carried were addressed through multi-sensor deployment along with
out with a wirelessly connected base station that reduced the fusion methods at the data level, feature level, and decision
complex onboard processing, size, cost, and power consump- level [131]. The multi-sensor approach eliminates the blind
tion of the system [125]. The hybrid waveform designed by spot to some extent as the target remains visible to at least
alternating a single 5.8 GHz signal with a stepped frequency one radar. Information lost due to variation in target distance
continuous wave (SFCW) was used as a transmit signal. and aspect angle is recovered with data-level fusion. In data-
An incremental step of 25 MHz in the 6-7 GHz band was level fusion, data acquired from each sensor is averaged and
taken to design the SFCW waveform. The single tone and the fused using discrete wavelet transform (DWT) [132], [133].
SFCW wave determined the speed and absolute distance of the Single sensor-based systems have low accuracy for hard to
target, respectively. This hybrid waveform approach effectively detect fall or occlusion issues. To increase accuracy, two radars
determined all the gait characteristics. Also, the noise or (GE RCR 50), one mounted at the ceiling, and others above the
interference corrupted data can affect the reliability of Doppler ground were used [134]. The Doppler radar is highly propor-
signatures and consequently misclassify the fall. tional to the velocity component on the direction of the human
Since during the fall event, a high-frequency signal for movement. Since fall is a downward motion, ceiling-mounted
short duration and then low-frequency signal for a longer radar can better align the radar-human direction with velocity
duration is generated, therefore, the time-frequency charac- than radar mounted at the torso level. Therefore, ceiling radar
teristics were explored to improve the performance [126]. produced better performance than the radar placed at the
Leveraging the time-frequency characteristics of fall signals, torso level and reduces the false alarms to half. However,
the simple and efficient approach of discrete stationary wavelet the concatenation of signal features of both the radars resulted
transform (SWT) was used as a feature extraction technique. in better efficiency of the overall system.
The time-frequency representation was also utilized through In order to gain the minimum preprocessing requirement
STFT and SWT for feature extraction along with SVM for and high accuracy, multilevel feature learning from radar time-
classification of the event with high accuracy [127]. As an frequency representations was used. A radar time series was
extension to the concept, the hybrid approach of compressed derived from the radar backscattered matrix, and a capsule
sensing and multi-window analysis based on Slepian or Her- network for automating feature learning was used [135]–[137].
mite functions was used to improve the performance even with The capsule network provided a better hierarchical image
the incomplete or noise corrupted data [128]. representation and effectively dealt with small and imbalance

Authorized licensed use limited to: Auckland University of Technology. Downloaded on October 10,2020 at 06:20:29 UTC from IEEE Xplore. Restrictions apply.
6908 IEEE SENSORS JOURNAL, VOL. 20, NO. 13, JULY 1, 2020

TABLE X
R ADAR S ENSOR -B ASED FALL D ETECTION S YSTEMS

training dataset. The radar technology carries great potential signal after a certain time interval. The measured time (t) com-
to be the competitive fall detection technique. To name a few putes the distance of the object. The distance d is calculated
of its attractive features are privacy preservation, especially as in (4) and (5):
in restrooms, safety, non-obstructive, not affected by lighting cair ∗ t
conditions. Recently advance signal processing techniques d = (4)
2 
at micro, and macro movement also makes it useful for
cair = 20.0457 ∗ (273.15 + t) (5)
measuring vital health parameter. It also can suppress noise
due to the scattering of electromagnetic waves from the room’s where cair is the velocity of the wave. Generally, the ultrasonic
furniture, walls, ceiling, and floor. They are small, low cost, sensors use the approximate frequency of 40 kHz with an
low power, and can be easily installed on a wall or ceiling eight-pulse signal waveform. The maximum feasible distance
depending on the logistic need and the required signal strength. of the sensor is 254 inches. The architecture of Ultrasonic
Table X summarized the merits and limitations of the works systems is shown in Fig. 8. The concept of distance mea-
related to radar-based systems. surement through an ultrasonic sensor was used to develop
7) Ultrasonic Sensor: The Ultrasonic wave sensor works on a fall detection system [138]. An array of ultrasonic sensors
the principle of time-of-flight. The sensor radiates a pulse was placed at the ceiling, and latitude information of the plane
signal towards the object and then receives back the reflected facing the sensor was obtained using the time of flight method.

Authorized licensed use limited to: Auckland University of Technology. Downloaded on October 10,2020 at 06:20:29 UTC from IEEE Xplore. Restrictions apply.
SINGH et al.: SENSOR TECHNOLOGIES FOR FALL DETECTION SYSTEMS: A REVIEW 6909

was classified based on the number of activated detectors;

the area occupied by the time series of that signal strength.
The system suffered from large pets, environmental effects on
ultrasonic waves, and the real-life performance evaluation.
Further reducing the installation complexity, two different
types of ultrasonic sensors HC-SR04 and LV-max Sonar EZ0
were used for fall detection [143]. Four HC-SR04 sensors at
the corner and one LV max Sonar EZ0 were used to cover the
complete area. Utilizing the same principle [139]–[141],
the distance of the subject from the ground was calculated,
and the decision tree algorithm was used for the classification
of fall events. This design reduced the installation complexity.
However, the environmental effect on ultrasonic waves, false
alarm due to large pets, fall near the walls, single person-based
designs, and real-life performance evaluation were some sig-
Fig. 8. The system architecture of the Ultrasonic sensor-based fall nificant challenges that require further investigation. Table XI
detection system.
summarizes the ultrasonic sensor fall detection systems.

The latitude of return sound beam, when compared with the

default scenario (i.e., empty room), determines the posture. For C. Hybrid Sensor
the default scenario pattern, the system was calibrated as per The multi-sensor fusion can potentially improve the relia-
user, the environment, and room size. The rapid transition from bility and specificity of the fall detection system. Due to com-
standing or seating latitude to ground latitude was classified plexity in fall kinematics and diversity in fall characteristics,
as fall. The system achieved workable accuracy after six the desired sensitivity and specificity is hard to achieve with
months of trial setup in an assisted living care room. High a single sensor system. The challenges and issues in imple-
calibration requirements and customization of system, person menting a multi-sensor fall detection system are discussed in
falling near the walker, high false alarm rates were the main detail by Koshmak et al. [144]. A summary of hybrid sensors
challenges observed with the design. The method detects the is given in Table XII. The accuracy of the PIR sensor-based
fall by a decrement in the height of the subject but was not system was increased using a multi-sensor approach, where
successful in classifying the gesture and position of the fall. the data from the floor vibration sensor was fused to classify
To address the issue, an array of 16 Ultrasonic sensors was the event [145]. Two PIR sensors placed vertically at a height
placed on the ceiling and wall [139]. The distance information detected the human motion. First, the floor vibration signature
achieved by sensors was placed in an array, and a sudden for fall was observed, and if sequential activation of the lower
decrease in the distance from the ground was classified as a fall PIR sensor and inactivation of the upper PIR sensor was
event. The 16 Ultrasonic sensor arrays though improved the also observed then the event was classified as fall. Moreover,
accuracy and classified the gesture and position of the fall yet a fusion of floor vibration sensors introduced the challenges of
suffered from the issue of installation complexity. Moreover, low impact fall, undetected large pets triggering a false alarm.
the sensors did not cover the complete area, and a 7% rejection The author in [146] comes out with a unique system based
error was observed. The challenge was addressed through the on the fusion of floor vibration and acoustic sensing and uses
optimization of array design and high rejection error [140]. pattern recognition algorithm to discriminate against the fall
An octagonal sensor array was designed and deployed on the event. A microphone and an accelerometer (vibration sensor)
roof and wall of a room. This arrangement provided an overlap are attached to the floor may be in one corner, which is
area of 20% between the sensors and covered the complete sensitive to the entire room. When a significant event appears
room area. The roof and wall sensors provided the posture in the vibration signal, both the vibration and sound signals are
information through measuring the distance between subject analyzed to trigger the classification algorithm. The system can
and ground. Through the octagonal design addressed the detect the fall with high accuracy for distance up to 5 m. The
coverage issue, yet the installation complexity, environmental system is designed to be adaptive and can be easily calibrated
dependent accuracy of the system, fall near the wall, low to any floor type. However, it may not be sensitive to low
accuracy of the design were some major challenges associated impact human falls in some cases. Also, the challenge of
with the ultrasonic technology-based system. The installation large pets triggering false alarms of floor vibration techniques
complexity and impracticability of complete coverage of the remained unresolved.
wall by an array of the ultrasonic sensor was addressed through To improve the detection for low impact falls, i.e., sensi-
an alternative design [142]. Arduino based ultrasonic array tivity, Huan-Wen et al. [147] employed the infrared camera
sensors were used and fixed on the wall above 15 cm from the along with the floor pressure to eliminate any blind spot. The
ground. The time interval for observation was set at 500 ms, linear variable differential transformer (LVDT) is employed for
and signal strength was collected in the cycle based on the time floor pressure sensing. The output from the pressure sensor is
sequence. Time history signal features were extracted, and analyzed and is compared with the predefined threshold limit.
SVM was used for posture classification. The falling situation In case of a fall event, the signal is sent to start the camera

Authorized licensed use limited to: Auckland University of Technology. Downloaded on October 10,2020 at 06:20:29 UTC from IEEE Xplore. Restrictions apply.
6910 IEEE SENSORS JOURNAL, VOL. 20, NO. 13, JULY 1, 2020

TABLE XI
U LTRASONIC S ENSOR -B ASED FALL D ETECTION S YSTEMS

tracking. The feature extracted from the infra-red image is To increase the accuracy and assess the fall impact,
processed by the threshold-based classification algorithm to the fusion of two sensors, accelerometer, and cardio-
identify the fall. The design improved the sensitivity, but the tachometer, were used [153]. Fall detection based on
high sensitivity also detected a drop of objects and other accelerometer value, trunk angle and heart rate through cardio
low impact events, making a trade-off between sensitivity and tachometer defined the fall and its impact. The orientation
accuracy. of the subject plays a critical role in accurate fall detection
The fusion of a barometric altimeter with an accelerometer but difficult to derive from the single 3-axis accelerometer.
sensor was able to achieve higher accuracy for low impact In another study, the fusion of the wearable and ambient
falls [148]. The drift-free vertical velocity estimation along sensors is employed to characterize the human motion [154].
with height and posture information, i.e., the gait analysis was The data from the accelerometer is combined with the optical
performed through a threshold-based algorithm prior-to-fall. flow from the surrounding cameras to provide local and global
Further, the sophisticated classification algorithm increases motion detection. Key postures are learned using HMM, and
accuracy but priced with high processing complexity and posture classification are done based on the relative likelihood
large power consumption. The electronic compass with an of recorded postures.
accelerometer is used to discriminate fall events from ADL Since the detection accuracy of PIR sensors is dependent on
at the initial stages of the four-state machine algorithm [149]. the distance of the person and sensitive to along the axis of the
This reduces the complexity, and then if the state machine sensor’s field of view only. To address these issues, microwave
processes the data at the third stage to be of fall, SVM is and PIR-based dual technology sensor (DTS) was used [155].
applied at the fourth stage for the final classification of the The PIR sensor detects the human and the microwave sensor
event. detects the movement of that human for tracking. An array
The accuracy of only accelerometer-based wearable devices of such sensors can identify multiple people moving in the
is low when the classification is based on thresholding. residence & can detect multiple falls also based on the sensor
To address the issue, the wearable device along with the activation pattern. The earlier algorithms had the limitation
smartphone was used [150]. The system first detected the that if after the fall, the person is conscious and moving,
acceleration values of wearable devices as higher than a the sensor shall be active and not triggering the alarm because
threshold to trigger the posture analysis process through a of the absence of the inactivity period. This was addressed
smartphone. When posture was confirmed as of lying, then using two DTS placed vertically, one DTS facing upwards,
the event was classified as fall. Additionally, based on the monitoring the upper half of the room and the other one
lying duration, the level of severity of falls was identified. observing the lower half of the room. Thus, the small move-
Ozcan et al. [151] introduced a new concept of the wearable ment of a fallen person shall not activate the upper DTS, and
camera, thus not restricting it to a limited area. Further, the fall shall be detected. The two DTS arrangement employed
the fusion of the accelerometer and camera data improves the with the heuristic decision tree classification model was able
detection accuracy as compared to the accelerometer-only or to track multiple people fall with high accuracy. The design
camera-only systems [152]. The classifiers of both the sensors had a limitation of false alarm when someone has left the
need to be uncorrelated, and the fusion algorithm based on the building, and the sensor was not able to detect this scenario.
sum rule of two normalized classifiers is applied to eliminate The large size and difficulty in the practical deployment of
false positives. the previously proposed acoustic FADE system [98] along with

Authorized licensed use limited to: Auckland University of Technology. Downloaded on October 10,2020 at 06:20:29 UTC from IEEE Xplore. Restrictions apply.
SINGH et al.: SENSOR TECHNOLOGIES FOR FALL DETECTION SYSTEMS: A REVIEW 6911

TABLE XII
H YBRID S ENSOR FALL D ETECTION S YSTEMS

Authorized licensed use limited to: Auckland University of Technology. Downloaded on October 10,2020 at 06:20:29 UTC from IEEE Xplore. Restrictions apply.
6912 IEEE SENSORS JOURNAL, VOL. 20, NO. 13, JULY 1, 2020

the detection accuracy of sound source localization and direc-

tion of arrival degrade in the multi-interference environment.
The authors extended the study and used a Microsoft Kinect
device, which consists of a 4-element linear microphone array
and a depth sensor [99]. The superior source localization
is achieved by employing minimum variance distortion-less
response (MVDR) adaptive BF (ABF), which resulted in
higher selectivity and sound quality from the direction of
interest. Doukas et al. [156] utilize the fusion of audio, video
and motion data to implement the patient status awareness
system. Video and audio information is captured by installing
cameras and microphone arrays at the monitoring site. The
motion data and patient-generated sound are collected through
special wearable sensor nodes with networking capability. The Fig. 9. Relationship in accuracy and experimental dataset.
separate feature extraction and classification techniques are
applied to the data captured from each sensor. Many sensors The dataset used for training and testing of classifier
increased the overall complexity of the system. played a significant role in deciding the performance of the
designed system. The evaluation of performance parameter
varies widely based on the number of participants, the number
of fall events to ADL activity in an experimental dataset, and
IV. D ISCUSSION
the duration of the experiment. The accuracy for a group of
A. General Evaluation of Sensor Technologies experimental datasets can be broadly clubbed into two clusters
The fall detection approaches can be subdivided into three of 96%-100% and 89%-96%, respectively. As the experimental
parts: (i) type of sensor technology, (ii) signal feature extrac- dataset increases, a general trend of accuracy degradation is
tion technique and processing to mitigate the noise effects, observed as shown in Fig. 9. For instance, in [91], the accuracy
(iii) the classification algorithm employed for extracted fea- of 100% is achieved with a very small data set of 23 tests only.
tures to classify the event. Most of the research is based on While the accuracy decreases to 98% in [99] with 240 tests for
the nine types of sensors that were discussed in this article. the same sensor type. As the tests increases to 1870 in [135],
The vision-based sensors were mostly researched, followed the accuracy decreases to 91%. The trend is visible in both
by the accelerometer and gyroscope-based system. The signal clusters. It broadly indicates that most of the systems which
processing and extraction technique varied from simple volt- reported very high accuracy were tested with limited datasets,
age extraction to MFCC and digital signal processing-based and the actual robustness of the system can only be inferred out
techniques. The most widely used signal feature extraction when validated with the larger dataset. The data set varied from
technique is MFCC. The signal processing techniques mostly the publicly available fall data sets to practically performed in
govern the complexity and power requirement of the designed a laboratory environment. Almost all the studies performed
system. Simple voltage extraction of the signal has the lowest the simulation of falls and activities of non-fall or daily living
processing complexity in conjunction with a threshold-based in a controlled environment. Most of the dataset has a larger
classification algorithm. However, these designs suffer from occurrence of falls to justify the working of their proposed
low accuracy. Support vector machine, K-nearest neighbor, and systems. This performance may not be available when the
neural networks are the commonly used high-end classification systems are deployed in a real-life environment, where the fall
algorithm; however, the threshold-based algorithm is mostly event happens rarely. The experimental dataset is widely varied
used and provides a reasonable tradeoff between sensitivity among all the studies and is limited. For better evaluation of
and accuracy. the system, the balanced dataset for the classifier algorithm
There is a wide variation in the context of technology assess- may be produced to mitigate the trade-off between sensitivity
ment and the type of participants involved in experiments. and accuracy.
The fall detection system should have high sensitivity and The systems are not tested in a real-life environment, and
robustness. For different sensor technologies, the tradeoff is the real-life deployment and performance evaluation remained
compared taking the area under the ROC curve, sensitivity a pivotal issue to be addressed. The comparative analysis of
accuracy, and specificity as performance parameters. The sensor technologies is shown in Table XIII. It provides a basic
results for performance parameters are mostly in the same framework in deciding the technology for a specific scenario or
range for each technology. Sensitivity varied from (80.7% location according to the prerequisites for real-life deployment.
to 100%) and accuracy reached from 80% to 100% for all For example, the implementation of fall detection systems in
the technologies. The performance parameters are mostly bathrooms has a critical prerequisite of privacy awareness. The
technology independent and mainly governed by signal feature real-time performance of the system must be assessed through
extraction technique, type of deployment of sensors, and real-life deployment, and specific adaptation rate by older
classification algorithms. The signal feature extraction tech- adults should be analyzed. The information on fall gesture and
nique mainly controls processing complexity and the power position of the fall to the caretaker can help in deciding about
requirements of the system. the type of care required. This shall reduce the post-fall impact

Authorized licensed use limited to: Auckland University of Technology. Downloaded on October 10,2020 at 06:20:29 UTC from IEEE Xplore. Restrictions apply.
SINGH et al.: SENSOR TECHNOLOGIES FOR FALL DETECTION SYSTEMS: A REVIEW 6913

TABLE XIII
C OMPARATIVE A NALYSIS OF S ENSOR T ECHNOLOGIES

and injury repercussions. Research on fall or fall detection is and number of sensors. However, as the number of sensors
mainly focused on technical outcomes. In the majority of the increases, the cost and installation complexity also increase
studies, the performance parameters considered are sensitivity, substantially. The tradeoff between the accuracy and placement
specificity, and accuracy, while minimal consideration is given of sensors is another area to be researched. Specifically,
to health parameters. The hospital admittance rate, post-fall the occlusion, intrusion issues, and cost-quality tradeoff for
care time, post-fall hospitalization rate, the caregiver response image sensors poses a significant challenge for the wide usage
has not been touched upon in the studies. The cost of deploy- of image sensing for fall detection systems. The miniatur-
ment and the designed system is also not explicitly described, ization of radar sensing technology and its increased ability
and fewer studies provide insight into the cost. The studies to classify micro and macro movement makes it suitable for
have covered to provide the range of detection, but to bar a the measurement of vital health parameters of humans. The
few, the majority have not addressed the power requirement utilization of radar for monitoring vital health parameters
and the ease of deployment of the fall detection systems. along with fall detection for designing a robust fall detection
system as well as a wellness preservation system could lead
to a new area of the research.
B. Future Scope and Open Issues The false alarms due to the inability of sensors in dif-
Different sensor technologies offer a wide variation in ferentiating between the fall of a human and a large pet is
performance outcomes. Each sensing technology wearable, another major challenge for acoustic, floor pressure and ultra-
ambiance or hybrid has its strengths and weakness, as summa- sonic sensor-based fall detection systems. The signal features
rized in Table XIII. The studies leave space for improvements generated by the pets and humans differ in context to vibration
in performance parameters through better designs or fusion or Doppler effects. The classification algorithms based on
with different technology. The wearable sensors are battery processing the features from different sensors reduces the false
operated. The power starving wireless communication and sig- fall event detection. Hence the fusion of other environmental
nal processing reduce the operating duration of these devices. sensors is proposed as mitigation for this issue. However, this
The device needs to be in the coverage area of the communica- increases the installation complexity, and significant process-
tion network. The power optimization of wearable devices by ing is required at classification and feature extraction stages.
exploring novel classification and signal processing techniques Therefore, the false alarm triggering due to large pets remains
with minimum processing complexity is the biggest challenge an open issue for these sensing technologies.
for these devices. Though the evolution of smartphones has The limited and biased data set and evaluation of the
addressed the power optimization issue a little, yet the tradeoff system in a real-world environment are common issues in
between the placement of a smartphone at a specific location every technology. The performance and robustness of design
and accuracy of detection is an open issue to be addressed. depend on the availability of the testing data set. Most of the
The accuracy of ambiance-based fall detection systems, experiments are performed with young adults, whereas the fall
which are primarily based on image sensors, radar, ultrasonic, pattern of the elderly might be different. Therefore, data set
acoustic, and floor pressure is governed by the placement from elderly subjects would help to validate the algorithms and

Authorized licensed use limited to: Auckland University of Technology. Downloaded on October 10,2020 at 06:20:29 UTC from IEEE Xplore. Restrictions apply.
6914 IEEE SENSORS JOURNAL, VOL. 20, NO. 13, JULY 1, 2020

sensing technologies. The hardware non-idealities, parameter and cost impact of fall-related injuries, the research question
drifts for long term operation also makes the generation of is framed and described in detail in section 1.
datasets non-uniform, and hence, a biased testing dataset is
available for the evaluation of design robustness. Different B. Search Criteria
scenarios of noise, daily activities, and fall events are needed To conform to the second stage of the framework, i.e., to
for generating a well-defined testing data set. identify the relevant studies and to achieve the objective of
the review to be as comprehensive as possible, a strategy
V. C ONCLUSION for identifying the primary studies through searching via
Falls are a major concern for maintaining a better healthy different electronic databases is adopted. This includes Google
life for older adults. Broadly based on sensor technologies, Scholar, Scopus, Web of Science, Medscape, Ovid, Medline,
the fall detection system is characterized into acoustic, floor IEEE Xplore, Bibliographies within references. Textbooks of
pressure, image, PIR, near field imaging, radar, ultrasonic, the related field were also searched. Peer-reviewed journals,
smartphone-based, and wearable sensors-based systems. The studies of health care institutions, independent research of
quality research article is reviewed based on our inclusion and reputed institutions were selected to get the articles for this
exclusion criteria. We have critically reviewed the existing sen- review. The search strategy objective used the following terms:
sor technologies in terms of feature extraction, classification, fall, fall detector, fall detection, fall detection technology,
performance, and experimental dataset. We have summarized Biosensor, elderly care, smart homes, wearable sensors, fall
the strengths and weaknesses of each technology, which would detection (acoustic, floor pressure, near field imaging, camera,
help researchers to select a sensor technology for different video, ultrasonic, radar, PIR, smartphone, wearable), old age
applications. Future research direction includes evaluation of health care, assisted living and fear of falling.
the testbed with elderly subjects, consideration to deployment
and installation complexity, and its adaptability in a real-life C. Inclusion and Exclusion Criteria
environment.
After getting the relevant literature, the inclusion and exclu-
sion criteria are established and applied to the studies as per
A PPENDIX the third stage of the framework.
R EVIEW M ETHODOLOGY • Research published during the interval 2008 to
The framework suggested by Daudt et al. [157] is adopted August 2019.
for reviewing fall detection systems. He suggested that the • The research focused on technology adaptation to address
review is meant for mapping the fields of study in an open human fall.
range of available literature. A review details and disseminates • Studies focused on technology, either environmental or
the findings for research scholars or system designers in the wearable.
area of interest. Another aspect touched upon is to analyze • Studies focused on the detailed insight of technological
the existing literature comprehensively and filter out the aspects.
research gaps. Daudt et al. incorporated suggestions from • Detectors useful for indoor or outdoor.
Levac et al. [158] and modified the five-stage framework • Conference proceedings review and review papers pub-
described by Arksey and O’Malley [159] for reviews. The lished in peer-reviewed journals.
modification resulted in enhancing the review for evaluating 1) Inclusion Criteria: Studies published before 2008 were
the quality of studies or research itself, and modification excluded as they may not be relevant to present technological
of the application of the inclusion /exclusion criteria. This advancements, or they may already be covered through recent
review incorporates the framework proposed by Daudt et al. references.
for methodology and material selection. The process is • English is the most commonly used language in interna-
iterative, and six stages are emphasized for conducting a tional conferences, journals, and covers almost all works
comprehensive and extensive review through the framework. related to all types of technologies. Hence to avoid
The framework focused on the primary sources, key concepts, resource and time, research work in other languages is
and the gaps in the research area as a standalone subject. The excluded.
fall detection technology fits suitably in the above-mentioned • Theoretical papers and technology not ready for imple-
adopted framework. mentation are excluded.
• Research work not available in full detail or papers.
A. Identify the Research Problem • Studies not having enough detail for chart preparation.

The framework defined the first stage as identifying the 2) Exclusion Criteria: After getting relevant articles, dupli-
research question. The motivation of the review is to study the cates are removed through online citation management soft-
technical evaluation of the sensor technology for the fall detec- ware. Inclusion/Exclusion criteria are applied. The selection
tion system and their associated implementation challenges. algorithm is illustrated in Fig. 10.
The challenges and unaddressed issues of different sensor tech-
nologies used for fall detection systems are comprehensively D. Charting Data
reviewed and presented for future research and applications. As per the fourth stage of the framework, a chart was
Through an exhaustive review of fall, its characteristics, health, prepared to extract data from the final screened relevant

Authorized licensed use limited to: Auckland University of Technology. Downloaded on October 10,2020 at 06:20:29 UTC from IEEE Xplore. Restrictions apply.
SINGH et al.: SENSOR TECHNOLOGIES FOR FALL DETECTION SYSTEMS: A REVIEW 6915

TABLE XIV
C ATEGORIZATION OF R EVIEW P ROCESS I NTO D IFFERENT S ENSORS

• Key results and measurements of the studies.

• Implementation challenges, solutions, and open issues for
further research.

E. Search Results
Summarizing and reporting the results is the fifth stage of
the framework adopted for the review and is demonstrated
through the search results. The process resulted in a total
of 634 relevant articles, and after screening and applying inclu-
sion/exclusion criteria, 103 articles are selected. The literature
on the topic shows a massive interest of the researcher towards
the problem. Exploring technology has excellent potential to
help older adults to lead a better lifestyle. Chart data approach
resulted in categorizing the articles according to the types of
Fig. 10. Selection algorithm for the review of studies. technology used for fall detection. A summary of the search
results is presented in Table XIV.
studies. The data obtained were broadly categorized into four
parts:
• General characteristics such as the journal information, F. Research Gap
the title of the article, year of publication, journal publi- Daudt et al. added the sixth stage as research gap identifi-
cation ranking, number of citations. cation and reporting this information to the policymakers and
• Technical characteristics such as sensor technology, type research community. The future scope and open issue section
of technology (ambiance, wearable, floor-based), trans- under discussion of this review addresses the sixth stage of
mission methods, machine learning process. the framework.

Authorized licensed use limited to: Auckland University of Technology. Downloaded on October 10,2020 at 06:20:29 UTC from IEEE Xplore. Restrictions apply.
6916 IEEE SENSORS JOURNAL, VOL. 20, NO. 13, JULY 1, 2020

R EFERENCES [26] I. Gringauz et al., “Risk of falling among hospitalized patients with
[1] World Population Prospects: The 2015 Revision, Key Findings and high modified morse scores could be further stratified,” BMC Health
Advance Tables, document ESA/P/WP 241, United Nations, Depart- Services Res., vol. 17, no. 1, p. 721, Dec. 2017.
ment of Economic and Social Affairs, Population Division, 2015. [27] C. Boelens, E. E. G. Hekman, and G. J. Verkerke, “Risk factors for falls
[2] F.A.D., HQSC, New Zealand, (2018). Falls in People Aged 50 and of older citizens,” Technol. Health Care, vol. 21, no. 5, pp. 521–533,
Over. [Online]. Available: [Link] Oct. 2013.
health-quality-evaluation/projects/atlas-of-healthcare-variation/falls/. [28] S. Abbate, M. Avvenuti, P. Corsini, J. Light, and A. Vecchio, “Moni-
[3] WHO Library Cataloguing-in-Publication Data. (2007). WHO Global toring of human movements for fall detection and activities recognition
Report on Falls Prevention in Older Age. [Online]. Available: in elderly care using wireless sensor network: A survey,” in Wireless
[Link] Sensor Networks: Application-Centric Design, G. V. Merret and Y. K.
[4] W. He, D. Goodkind, and P. Kowal. (2016). An Aging World. US Cen- Tan, Eds. Rijeka, Croatia: InTech, 2010, pp. 147–166.
sus Bureau. [Online]. Available: [Link] [29] X. Yu, “Approaches and principles of fall detection for elderly and
Census/library/publications/2016/demo/[Link] patient,” in Proc. HEALTHCOM, Jul. 2008, pp. 42–47.
[5] T. Blakely et al., “Updated New Zealand health system cost estimates [30] C.-C. Yang and Y.-L. Hsu, “A review of accelerometry-based wearable
from health events by sex, age and proximity to death: Further motion detectors for physical activity monitoring,” Sensors, vol. 10,
improvements in the age of ‘big data,”’ NZ Med. J., vol. 128, no. 1422, no. 8, pp. 7772–7788, 2010.
pp. 13–23, 2015. [31] C.-F. Lai, Y.-M. Huang, J. H. Park, and H.-C. Chao, “Adaptive body
[6] A. C. Scheffer, M. J. Schuurmans, N. van Dijk, T. van der Hooft, and posture analysis for elderly-falling detection with multisensors,” IEEE
S. E. de Rooij, “Fear of falling: Measurement strategy, prevalence, risk Intell. Syst., vol. 25, no. 2, pp. 20–30, Mar. 2010.
factors and consequences among older persons,” Age Ageing, vol. 37, [32] J. Lenz and A. Edelstein, “Magnetic sensors and their applications,”
no. 1, pp. 19–24, Jan. 2008. IEEE Sensors J., vol. 6, no. 3, pp. 631–649, Jun. 2006.
[7] M. E. Tinetti, “Predictors and prognosis of inability to get up after falls [33] Y. S. Suh, “Inertial and magnetic sensor data compression considering
among elderly persons,” JAMA, J. Amer. Med. Assoc., vol. 269, no. 1, the estimation error,” Sensors, vol. 9, no. 8, pp. 5919–5932, 2009.
pp. 65–70, Jan. 1993. [34] T. Le and R. Pan, “Accelerometer-based sensor network for fall detec-
[8] N. B. Alexander, J. Ulbrich, A. Raheja, and D. Channer, “Rising from tion,” in Proc. Biomed. Circuits Syst. Conf., Nov. 2009, pp. 265–268.
the floor in older adults,” J. Amer. Geriatrics Soc., vol. 45, no. 5, [35] A. K. Bourke et al., “Evaluation of waist-mounted tri-axial accelerom-
pp. 564–569, May 1997. eter based fall-detection algorithms during scripted and continuous
[9] J. van Hoof, H. S. M. Kort, P. G. S. Rutten, and M. S. H. Duijnstee, unscripted activities,” J. Biomech., vol. 43, no. 15, pp. 3051–3057,
“Ageing-in-place with the use of ambient intelligence technology: Nov. 2010.
Perspectives of older users,” Int. J. Med. Informat., vol. 80, no. 5, [36] M. Neggazi, A. Amira, and L. Hamami, “A wireless reconfigurable
pp. 310–331, May 2011. system for falls detection,” in Proc. 11th Int. Conf. Inf. Sci., Signal
[10] N. Noury, P. Rumeau, A. K. Bourke, G. ÓLaighin, and J. E. Lundy,
Process. Appl. (ISSPA), Jul. 2012, pp. 77–82.
“A proposal for the classification and evaluation of fall detectors,”
[37] H. Rabah, A. Amira, and A. Ahmad, “Design and implementaiton
IRBM, vol. 29, no. 6, pp. 340–349, Dec. 2008.
[11] J. T. Perry, S. Kellog, S. M. Vaidya, J.-H. Youn, H. Ali, and H. Sharif, of a fall detection system using compressive sensing and shimmer
“Survey and evaluation of real-time fall detection approaches,” in Proc. technology,” in Proc. ICM, Dec. 2012, pp. 1–4.
6th Int. Symp. High Capacity Opt. Netw. Enabling Technol. (HONET), [38] A. Lopez-Yunez et al., “A novel approach for high speed wireless pre-
Dec. 2009, pp. 158–164. fall detection multisensory system,” in Proc. MWSCAS, Aug. 2014,
[12] M. Mubashir, L. Shao, and L. Seed, “A survey on fall detection: pp. 857–859.
Principles and approaches,” Neurocomputing, vol. 100, pp. 144–152, [39] J. A. Tropp and A. C. Gilbert, “Signal recovery from random mea-
Jan. 2013. surements via orthogonal matching pursuit,” IEEE Trans. Inf. Theory,
[13] Y. Birku and A. Agrawal, “Survey on fall detection systems,” Int. vol. 53, no. 12, pp. 4655–4666, Dec. 2007.
J. Pure Appl. Math., vol. 118, no. 18, pp. 2537–2543, 2018. [40] S. O. H. Madgwick, A. J. L. Harrison, and R. Vaidyanathan, “Esti-
[14] J. S. Madhubala, A. Sree, and B. J. A. Rani, “A survey on technical mation of IMU and MARG orientation using a gradient descent
approaches in fall detection system,” Nat. J. Physiol., Pharmacy algorithm,” in Proc. IEEE Int. Conf. Rehabil. Robot., Jun./Jul. 2011,
Pharmacol., vol. 5, no. 4, pp. 275–279, 2015. pp. 1–7.
[15] N. Pannurat, S. Thiemjarus, and E. Nantajeewarawat, “Automatic fall [41] H.-M. Chen, C.-K. Chen, H.-I. Wu, and R.-S. Tsay, “An accurate
monitoring: A review,” Sensors, vol. 14, no. 7, pp. 12900–12936, 2014. crowdsourcing-based adaptive fall detection approach using smart
[16] P. Bet, P. C. Castro, and M. A. Ponti, “Fall detection and fall risk devices,” in Proc. IEEE Int. Conf. Healthcare Inform. (ICHI),
assessment in older person using wearable sensors: A systematic Oct. 2016, pp. 456–460.
review,” Int. J. Med. Informat., vol. 130, Oct. 2019, Art. no. 103946. [42] Y. Tang, Z. Peng, L. Ran, and C. Li, “IPrevent: A novel wearable radio
[17] E. Cippitelli, F. Fioranelli, E. Gambi, and S. Spinsante, “Radar and frequency range detector for fall prevention,” in Proc. RFIT, Aug. 2016,
RGB-depth sensors for fall detection: A review,” IEEE Sensors J., pp. 1–3.
vol. 17, no. 12, pp. 3585–3604, Jun. 2017. [43] H. Djelouat, H. Baali, A. Amira, and F. Bensaali, “CS-based fall
[18] O. A. Atoyebi, A. Stewart, and J. Sampson, “Use of information detection for connected health applications,” in Proc. ICABME, 2017,
technology for falls detection and prevention in the elderly,” Ageing pp. 1–4.
Int., vol. 40, no. 3, pp. 277–299, Sep. 2015. [44] A. Jahanjoo, M. N. Tahan, and M. J. Rashti, “Accurate fall detection
[19] N. Lapierre, N. Neubauer, A. Miguel-Cruz, A. R. Rincon, L. Liu, and using 3-axis accelerometer sensor and MLF algorithm,” in Proc. 3rd
J. Rousseau, “The state of knowledge on technologies and their use for Int. Conf. Pattern Recognit. Image Anal. (IPRIA), Apr. 2017, pp. 90–95.
fall detection: A scoping review,” Int. J. Med. Inform., 2018,” vol. 111, [45] C. Wang, W. Lu, S. J. Redmond, M. C. Stevens, S. R. Lord, and N. H.
pp. 58–71, Mar. 2018. Lovell, “A low-power fall detector balancing sensitivity and false alarm
[20] Y. Delahoz and M. Labrador, “Survey on fall detection and fall
rate,” IEEE J. Biomed. Health Informat., vol. 22, no. 6, pp. 1929–1937,
prevention using wearable and external sensors,” Sensors, vol. 14,
Nov. 2018.
no. 10, pp. 19806–19842, 2014.
[21] K. Chaccour, R. Darazi, A. H. El Hassani, and E. Andres, “From [46] A. Purushothaman, K. V. Vineetha, and D. G. Kurup, “Fall detection
fall detection to fall prevention: A generic classification of fall-related system using artificial neural network,” in Proc. ICICCT, Apr. 2018,
systems,” IEEE Sensors J., vol. 17, no. 3, pp. 812–822, Feb. 2017. pp. 1146–1149.
[22] P. Vallabh and R. Malekian, “Fall detection monitoring systems: [47] S. Khojasteh, J. Villar, C. Chira, V. González, and E. de la Cal,
A comprehensive review,” J. Ambient Intell. Humanized Comput., “Improving fall detection using an on-wrist wearable accelerometer,”
vol. 9, no. 6, pp. 1809–1833, 2017. Sensors, vol. 18, no. 5, p. 1350, 2018.
[23] R. Broadley, J. Klenk, S. Thies, L. Kenney, and M. Granat, “Methods [48] J. Klenk et al., “The FARSEEING real-world fall repository: A large-
for the real-world evaluation of fall detection technology: A scoping scale collaborative database to collect and share sensor signals from
review,” Sensors, vol. 18, no. 7, p. 2060, 2018. real-world falls,” Eur. Rev. Aging Phys. Activity, vol. 13, no. 1, p. 8,
[24] World Health Organization, Geneva, Switzerland. (2004). WHO Global Dec. 2016.
Report on Falls Prevention in Older Age. [Online]. Available: [49] E. Casilari, J. A. Santoyo-Ramón, and J. M. Cano-García, “Analysis
[Link] of a smartphone-based architecture with multiple mobility sensors for
[25] J. S. Cassells and R. L. Berg, The Second Fifty Years: Promoting Health fall detection,” PLoS ONE, vol. 11, no. 12, 2016, Art. no. e0168069.
and Preventing Disability. Washington, DC, USA: National Academy [Online]. Available: [Link]
Press, 1992. 1371/[Link].0168069

Authorized licensed use limited to: Auckland University of Technology. Downloaded on October 10,2020 at 06:20:29 UTC from IEEE Xplore. Restrictions apply.
SINGH et al.: SENSOR TECHNOLOGIES FOR FALL DETECTION SYSTEMS: A REVIEW 6917

[50] T. de Quadros, A. E. Lazzaretti, and F. K. Schneider, “A move- [75] B. Andres, F. A. Hamprecht, and C. S. Garbe, “Selection of local opti-
ment decomposition and machine learning-based fall detection system cal flow models by means of residual analysis,” in Pattern Recognition,
using wrist wearable device,” IEEE Sensors J., vol. 18, no. 12, Berlin, Germany: Springer, 2007.
pp. 5082–5089, Jun. 2018. [76] N. Lu, Y. Wu, L. Feng, and J. Song, “Deep learning for fall detection:
[51] M. Saleh and R. L. B. Jeannes, “Elderly fall detection using wearable Three-dimensional CNN combined with LSTM on video kinematic
sensors: A low cost highly accurate algorithm,” IEEE Sensors J., data,” IEEE J. Biomed. Health Informat., vol. 23, no. 1, pp. 314–323,
vol. 19, no. 8, pp. 3156–3164, Apr. 2019. Jan. 2019.
[52] Z. Ma, Y. Qiao, B. Lee, and E. Fallon, “Experimental evaluation [77] H. Alkittawi, “A deep-learning-based fall-detection system to
of mobile phone sensors,” in Proc. ISSC. Letterkenny, Ireland: LYIT support aging-in-place,” Dept. Comput. Sci., Univ. Jordan,
Letterkenny, Jun. 2013, p. 49. Amman, Jordan, Tech. Rep., 2013, p. 128. [Online]. Available:
[53] M. Kaenampornpan, T. Anuchad, and P. Supaluck, “Fall detection [Link]
prototype for thai elderly in mobile computing era,” in Proc. ECTI, [78] Z.-P. Bian, J. Hou, L.-P. Chau, and N. Magnenat-Thalmann, “Fall
2011, pp. 446–449. detection based on body part tracking using a depth camera,” IEEE
[54] Q. V. Vo, G. Lee, and D. Choi, “Fall detection based on movement J. Biomed. Health Informat., vol. 19, no. 2, pp. 430–439, Mar. 2015.
and smart phone technology,” in Proc. IEEE RIVF Int. Conf. Comput. [79] L. Yang, Y. Ren, and W. Zhang, “3D depth image analysis for indoor
Commun. Technol., Res., Innov., Vis. Future, Feb./Mar. 2012, pp. 1–4. fall detection of elderly people,” Digit. Commun. Netw., vol. 2, no. 1,
[55] Y. Cao, Y. Yang, and W. Liu, “E-FallD: A fall detection system using pp. 24–34, 2016.
[80] G. Debard, M. Mertens, T. Goedemé, T. Tuytelaars, and B. Vanrumste,
Android-based smartphone,” in Proc. 9th Int. Conf. Fuzzy Syst. Knowl.
“Three ways to improve the performance of real-life camera-based fall
Discovery, May 2012, pp. 1509–1513.
detection systems,” J. Sensors, vol. 2017, pp. 1–15, Oct. 2017.
[56] Y. He and Y. Li, “Physical activity recognition utilizing the built-in
[81] F. Zhao, Z. Cao, Y. Xiao, J. Mao, and J. Yuan, “Real-time detection of
kinematic sensors of a smartphone,” Int. J. Distrib. Sensor Netw., vol. 9, fall from bed using a single depth camera,” IEEE Trans. Autom. Sci.
no. 4, 2013, Art. no. 481580. Eng., vol. 16, no. 3, pp. 1018–1032, Jul. 2019.
[57] P. Pierleoni, A. Belli, L. Palma, M. Pellegrini, L. Pernini, and S. Valenti, [82] X. Kong, Z. Meng, L. Meng, and H. Tomiyama, “A privacy protected
“A high reliability wearable device for elderly fall detection,” IEEE fall detection IoT system for elderly persons using depth camera,” in
Sensors J., vol. 15, no. 8, pp. 4544–4553, Aug. 2015. Proc. Int. Conf. Adv. Mech. Syst. (ICAMechS), Aug. 2018, p 31-35.
[58] A. Mao, X. Ma, Y. He, and J. Luo, “Highly portable, sensor-based [83] S. Kido, T. Miyasaka, T. Tanaka, T. Shimizu, and T. Saga, “Fall
system for human fall monitoring,” Sensors, vol. 17, no. 9, p. 2096, detection in toilet rooms using thermal imaging sensors,” in Proc.
2017. IEEE/SICE Int. Symp. Syst. Integr. (SII), Nov. 2009, pp. 83–88.
[59] P. Tsinganos and A. Skodras, “A smartphone-based fall detection [84] G. Diraco, A. Leone, and P. Siciliano, “An active vision system
system for the elderly,” in Proc. ISPA, 2017, pp. 53–58. for fall detection and posture recognition in elderly healthcare,” in
[60] W.-X. Weng and S.-C. Lo, “Fall detection based on tilt angle and Proc. Design, Automat. Test Eur. Conf. Exhib. (DATE), Mar. 2010,
acceleration variations,” in Proc. IEEE Trustcom/BigDataSE/ISPA, pp. 1536–1541.
Aug. 2016, pp. 1712–1717. [85] G. Baldewijns, G. Debard, G. Mertes, T. Croonenborghs, and
[61] A. Hakim, M. S. Huq, S. Shanta, and B. S. K. K. Ibrahim, “Smartphone B. Vanrumste, “Improving the accuracy of existing camera based fall
based data mining for fall detection: Analysis and design,” Procedia detection algorithms through late fusion,” in Proc. 39th Annu. Int. Conf.
Comput. Sci., vol. 105, pp. 46–51, Jan. 2017. IEEE Eng. Med. Biol. Soc. (EMBC), Jul. 2017, pp. 2667–2671.
[62] A. Shahzad and K. Kim, “FallDroid: An automated smart-phone-based [86] K. de Miguel, A. Brunete, M. Hernando, and E. Gambao, “Home
fall detection system using multiple kernel learning,” IEEE Trans. Ind. camera-based fall detection system for the elderly,” Sensors, vol. 17,
Informat., vol. 15, no. 1, pp. 35–44, Jan. 2019. no. 12, p. 2864, 2017.
[63] C.-L. Liu, C.-H. Lee, and P.-M. Lin, “A fall detection system using [87] A. Lotfi, S. Albawendi, H. Powell, K. Appiah, and C. Langensiepen,
k-nearest neighbor classifier,” Expert Syst. Appl., vol. 37, no. 10, “Supporting independent living for older adults; employing a visual
pp. 7174–7181, Oct. 2010. based fall detection through analysing the motion and shape of the
[64] C. Rougier, J. Meunier, A. St-Arnaud, and J. Rousseau, “Robust video human body,” IEEE Access, vol. 6, pp. 70272–70282, 2018.
surveillance for fall detection based on human shape deformation,” [88] A. Dubois and F. Charpillet, “Automatic fall detection system with
IEEE Trans. Circuits Syst. Video Technol., vol. 21, no. 5, pp. 611–622, a RGB-D camera using a hidden Markov model,” in Inclusive Soci-
May 2011. ety: Health and Wellbeing in the Community, and Care at Home.
[65] C. Rougier, A. St-Arnaud, J. Rousseau, and J. Meunier, “Video sur- ICOST (Lecture Notes in Computer Science), vol. 7910, J. Biswas, H.
veillance for fall detection,” in Proc. Video Surveill., 2011, pp. 46–51. Kobayashi, L. Wong, B. Abdulrazak, and M. Mokhtari, Eds. Berlin,
[66] J. Canny, “A computational approach to edge detection,” IEEE Germany: Springer, 2013.
Trans. Pattern Anal. Mach. Intell., vol. PAMI-8, no. 6, pp. 679–698, [89] M. Li, G. Xu, B. He, X. Ma, and J. Xie, “Pre-impact fall detection
Nov. 1986. based on a modified zero moment point criterion using data from Kinect
[67] M. Rudemo, “Statistical shape analysis,” in Statistics in Medicine, sensors,” IEEE Sensors J., vol. 18, no. 13, pp. 5522–5531, Jul. 2018.
vol. 19, no. 19, I. L. Dryden and K. V. Mardia, Eds. Chichester, U.K.: [90] R. Planinc and M. Kampel, “Introducing the use of depth data for fall
Wiley, 1998, p. 17, 347, and 2716–2717. detection,” Pers. Ubiquitous Comput., vol. 17, no. 6, pp. 1063–1072,
[68] D. F. DeMenthon and L. S. Davis, “Model-based object pose in 25 Aug. 2013.
[91] M. Popescu, Y. Li, M. Skubic, and M. Rantz, “An acoustic fall detector
lines of code,” Int. J. Comput. Vis., vol. 15, nos. 1–2, pp. 123–141,
system that uses sound height information to reduce the false alarm
Jun. 1995.
rate,” in Proc. 30th Annu. Int. Conf. IEEE Eng. Med. Biol. Soc.,
[69] L. Alhimale, H. Zedan, and A. Al-Bayatti, “The implementation of
Aug. 2008, pp. 4628–4631.
an intelligent and video-based fall detection system using a neural [92] J. S. Madhubala and A. Umamakeswari, “A vision based fall detection
network,” Appl. Soft Comput., vol. 18, pp. 59–69, May 2014. system for elderly people,” Indian J. Sci. Technol., vol. 8, no. S9,
[70] S. Albawendi, K. Appiah, H. Powell, and A. Lotfi, “Video based p. 172, 2015.
fall detection with enhanced motion history images,” in Proc. 11th [93] M. Popescu and A. Mahnot, “Acoustic fall detection using one-class
PErvasive Technol. Rel. Assistive Environ. Conf., New York, NY, USA, classifiers,” in Proc. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc.,
2016, pp. 529–536. Sep. 2009, pp. 3505–3508.
[71] H. Le Uyen Thuc, P. Van Tuan, and J.-N. Hwang, “An effective video- [94] Y. Li, Z. Zeng, M. Popescu, and K. C. Ho, “Acoustic fall detection
based model for fall monitoring of the elderly,” in Proc. Int. Conf. Syst. using a circular microphone array,” in Proc. Annu. Int. Conf. IEEE
Sci. Eng. (ICSSE), Jul. 2017, pp. 48–52. Eng. Med. Biol., Aug. 2010, pp. 2242–2245.
[72] H. Foroughi, A. Rezvanian, and A. Paziraee, “Robust fall detection [95] B. Mungamuru and P. Aarabi, “Enhanced sound localization,” IEEE
using human shape and multi-class support vector machine,” in Proc. Trans. Syst., Man, Cybern. B, Cybern., vol. 34, no. 3, pp. 1526–1540,
6th Indian Conf. Comput. Vis., Graph. Image Process., Washington, Jun. 2004.
DC, USA, Dec. 2008, pp. 413–420. [96] H. Do and H. F. Silverman, “A fast microphone array SRP-PHAT
[73] H. X. Hai and H. L. U. Thuc, “Cyclic HMM-based method for source location implementation using coarse-to-fine region contraction
pathological gait recognition from side view gait video,” Int. J. Adv. (CFRC),” in Proc. IEEE Workshop Appl. Signal Process. Audio Acoust.,
Res. Comput. Eng. Technol., vol. 4, no. 5, pp. 2171–2176. 2015. Oct. 2007, pp. 295–298.
[74] A. Núñez-Marcos, G. Azkune, and I. Arganda-Carreras, “Vision-based [97] E. Principi, D. Droghini, S. Squartini, P. Olivetti, and F. Piazza,
fall detection with convolutional neural networks,” Wireless Commun. “Acoustic cues from the floor: A new approach for fall classification,”
Mobile Comput., vol. 2017, pp. 1–16, Dec. 2017. Expert Syst. Appl., vol. 60, pp. 51–61, Oct. 2016.

Authorized licensed use limited to: Auckland University of Technology. Downloaded on October 10,2020 at 06:20:29 UTC from IEEE Xplore. Restrictions apply.
6918 IEEE SENSORS JOURNAL, VOL. 20, NO. 13, JULY 1, 2020

[98] Y. Li, K. C. Ho, and M. Popescu, “A microphone array system for [119] H. Rimminen, J. Lindstrom, M. Linnavuo, and R. Sepponen, “Detection
automatic fall detection,” IEEE Trans. Biomed. Eng., vol. 59, no. 5, of falls among the elderly by a floor sensor using the electric near field,”
pp. 1291–1301, May 2012. IEEE Trans. Inf. Technol. Biomed., vol. 14, no. 6, pp. 1475–1476,
[99] Y. Li, K. C. Ho, and M. Popescu, “Efficient source separation algo- Nov. 2010.
rithms for acoustic fall detection using a Microsoft Kinect,” IEEE [120] A. Arshad, S. Khan, A. H. M. Z. Alam, A. F. Ismail, and R. Tasnim,
Trans. Biomed. Eng., vol. 61, no. 3, pp. 745–755, Mar. 2014. “Capacitive proximity floor sensing system for elderly tracking and fall
[100] M. S. Khan, M. Yu, P. Feng, L. Wang, and J. Chambers, “An unsuper- detection,” in Proc. IEEE 4th Int. Conf. Smart Instrum., Meas. Appl.
vised acoustic fall detection system using source separation for sound (ICSIMA), Nov. 2017, pp. 1–5.
interference suppression,” Signal Process., vol. 110, pp. 199–210, [121] F. Fioranelli, M. Ritchie, and H. Griffiths, “Aspect angle depen-
May 2015. dence and multistatic data fusion for micro-Doppler classification of
[101] A. Irtaza, S. M. Adnan, S. Aziz, A. Javed, M. O. Ullah, and armed/unarmed personnel,” IET Radar, Sonar Navigat., vol. 9, no. 9,
M. T. Mahmood, “A framework for fall detection of elderly people pp. 1231–1239, 2015.
by analyzing environmental sounds through acoustic local ternary pat- [122] L. Liang, M. Popescu, M. Skubic, M. Rantz, T. Yardibi, and P. Cuddihy,
terns,” in Proc. IEEE Int. Conf. Syst., Man, Cybern. (SMC), Oct. 2017, “Automatic fall detection based on Doppler radar motion signature,” in
pp. 1558–1563. Proc. 5th Int. ICST Conf. Pervas. Comput. Technol. Healthcare, 2011,
[102] X. Tan and B. Triggs, “Enhanced local texture feature sets for face pp. 222–225.
recognition under difficult lighting conditions,” IEEE Trans. Image [123] J. Hong, S. Tomii, and T. Ohtsuki, “Cooperative fall detection using
Process., vol. 19, no. 6, pp. 1635–1650, Jun. 2010. Doppler radar and array sensor,” in Proc. IEEE 24th Annu. Int.
[103] M. Alwan et al., “A smart and passive floor-vibration based fall detector Symp. Pers., Indoor, Mobile Radio Commun. (PIMRC), Sep. 2013,
for elderly,” in Proc. 2nd Int. Conf. Inf. Commun. Technol., Apr. 2006, pp. 3492–3496.
pp. 1003–1007. [124] H. Krim and M. Viberg, “Two decades of array signal processing
[104] J. Clemente, F. Li, M. Valero, and W. Song, “Smart seismic sensing research: The parametric approach,” IEEE Signal Process. Mag.,
for indoor fall detection, location, and notification,” IEEE J. Biomed. vol. 13, no. 4, pp. 67–94, Jul. 1996.
Health Informat., vol. 24, no. 2, pp. 524–532, Feb. 2020. [125] M. Mercuri et al., “Analysis of an indoor biomedical radar-based
[105] L. Klack, C. Möllering, M. Ziefle, and T. Schmitz-Rode, “Future care system for health monitoring,” IEEE Trans. Microw. Theory Techn.,
floor: A sensitive floor for movement monitoring and fall detection in vol. 61, no. 5, pp. 2061–2068, May 2013.
home environments,” in Wireless Mobile Communication and Health- [126] B. Y. Su, K. C. Ho, M. J. Rantz, and M. Skubic, “Doppler radar fall
care (Lecture Notes of the Institute for Computer Sciences, Social activity detection using the wavelet transform,” IEEE Trans. Biomed.
Informatics and Telecommunications Engineering), vol. 55, J. C. Lin Eng., vol. 62, no. 3, pp. 865–875, Mar. 2015.
and K. S. Nikita, Eds. Berlin, Germany: Springer, 2011. [127] M. G. Amin, Y. D. Zhang, F. Ahmad, and K. C. D. Ho, “Radar
[106] F. Werner, J. Diermaier, P. Panek, and S. Schmid, “Fall detection signal processing for elderly fall detection: The future for in-home
with distributed floor-mounted accelerometers: An overview of the monitoring,” IEEE Signal Process. Mag., vol. 33, no. 2, pp. 71–80,
development and evaluation of a fall detection system within the Mar. 2016.
project eHome,” in Proc. 5th Int. ICST Conf. Pervas. Comput. Technol. [128] B. Jokanovic, M. G. Amin, Y. D. Zhang, and F. Ahmad, “Multi-
Healthcare, 2011, pp. 354–361. window time–frequency signature reconstruction from undersampled
[107] K. Chaccour, R. Darazi, A. H. El Hassans, and E. Andres, “Smart continuous-wave radar measurements for fall detection,” IET Radar,
carpet using differential piezoresistive pressure sensors for elderly fall Sonar Navigat., vol. 9, no. 2, pp. 173–183. 2015.
detection,” in Proc. IEEE 11th Int. Conf. Wireless Mobile Comput., [129] Q. Wu, W. Tao, Y. D. Zhang, and M. G. Amin, “Radar-based fall detec-
Netw. Commun. (WiMob), Oct. 2015, pp. 225–229. tion based on Doppler time–frequency signatures for assisted living,”
[108] G. Feng, J. Mai, Z. Ban, X. Guo, and G. Wang, “Floor pressure imaging IET Radar, Sonar Navigat., vol. 9, no. 2, pp. 164–172, Feb. 2015.
for fall detection with fiber-optic sensors,” IEEE Pervas. Comput., [130] L. Liu, M. Popescu, M. Skubic, M. Rantz, and P. Cuddihy, “An
vol. 15, no. 2, pp. 40–47, Apr. 2016. automatic in-home fall detection system using Doppler radar signa-
[109] L. Minvielle, M. Atiq, R. Serra, M. Mougeot, and N. Vayatis, “Fall tures,” J. Ambient Intell. Smart Environ., vol. 8, no. 4, pp. 453–466,
detection using smart floor sensor and supervised learning,” in Proc. Jul. 2016.
39th Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. (EMBC), Jul. 2017, [131] B. Erol, M. G. Amin, and B. Boashash, “Range-Doppler radar sensor
pp. 3445–3448. fusion for fall detection,” in Proc. IEEE Radar Conf. (RadarConf),
[110] B.-C. Yeo, V. Kumar, W.-S. Lim, J. Emerson Ra, and K.-B. Koh, May 2017, pp. 819–824.
“Development of electronic floor mat for fall detection and elderly [132] R. C. Gonzalez, R. E. Woods, and S. L. Eddins, Digital Image
care,” Asian J. Sci. Res., vol. 11, no. 3, pp. 344–356, Jun. 2018. Processing, vol. 976, 3rd ed. London, U.K.: Pearson, 2009.
[111] M. Popescu, B. Hotrabhavananda, M. Moore, and M. Skubic, [133] S. Aviyente and W. J. Williams, “Minimum entropy time-frequency
“VAMPIR—An automatic fall detection system using a vertical PIR distributions,” IEEE Signal Process. Lett., vol. 12, no. 1, pp. 37–40,
sensor array,” in Proc. 6th Int. Conf. Pervas. Comput. Technol. Health- Jan. 2005.
care, 2012, pp. 163–166. [134] B. Y. Su, K. C. Ho, M. Rantz, and M. Skubic, “Radar placement for
[112] D. J. Brady, N. P. Pitsianis, and X. Sun, “Reference structure tomog- fall detection: Signature and performance,” J. Ambient Intell. Smart
raphy,” J. Opt. Soc. Amer. A, Opt. Image Sci., vol. 21, no. 7, Environ., vol. 10, no. 1, pp. 21–34, Jan. 2018.
pp. 1140–1147, Jul. 2004. [135] H. Sadreazami, M. Bolic, and S. Rajan, “CapsFall: Fall detection
[113] X. Luo, T. Liu, J. Liu, X. Guo, and G. Wang, “Design and implemen- using ultra-wideband radar and capsule network,” IEEE Access, vol. 7,
tation of a distributed fall detection system based on wireless sensor pp. 55336–55343, 2019.
networks,” EURASIP J. Wireless Commun. Netw., vol. 2012, no. 1, [136] H. Sadreazami, M. Bolic, and S. Rajan, “Fall detection using standoff
pp. 1–13, Dec. 2012. radar-based sensing and deep convolutional neural network,” IEEE
[114] A. Sixsmith and N. Johnson, “A smart sensor to detect the falls Trans. Circuits Syst. II, Exp. Briefs, vol. 67, no. 1, pp. 197–201,
of the elderly,” IEEE Pervas. Comput., vol. 3, no. 2, pp. 42–47, Jan. 2020.
Apr. 2004. [137] H. Sadreazami, M. Bolic, and S. Rajan, “Residual network-based
[115] X. Fan, H. Zhang, C. Leung, and Z. Shen, “Robust unobtrusive supervised learning of remotely sensed fall incidents using ultra-
fall detection using infrared array sensors,” in Proc. IEEE Int. Conf. wideband radar,” in Proc. IEEE Int. Symp. Circuits Syst. (ISCAS),
Multisensor Fusion Integr. Intell. Syst. (MFI), Nov. 2017, pp. 194–199. May 2019, pp. 1–4.
[116] Q. Guan, C. Li, X. Guo, and B. Shen, “Infrared signal based elderly [138] Y. Huang and K. Newman, “Improve quality of care with remote
fall detection for in-home monitoring,” in Proc. 9th Int. Conf. Intell. activity and fall detection using ultrasonic sensors,” in Proc. Annu.
Human-Machine Syst. Cybern. (IHMSC), Aug. 2017, 373-376. Int. Conf. IEEE Eng. Med. Biol. Soc., Aug. 2012, pp. 5854–5857.
[117] R. Henry, L. Matti, and S. Raimo, “Human tracking using near [139] C. Nadee and K. Chamnongthai, “Multi sensor system for automatic
field imaging,” in Proc. 2nd Int. Conf. Pervasive Comput. Technol. fall detection,” in Proc. Asia–Pacific Signal Inf. Process. Assoc. Annu.
Healthcare, Jan./Feb. 2008, pp. 148–151. Summit Conf. (APSIPA), Dec. 2015, pp. 930–933.
[118] H. Rimminen, J. Lindström, and R. Sepponen, “Positioning accu- [140] C. Nadee and K. Chamnongthai, “Octagonal formation of ultra-
racy and multi-target separation with a human tracking system using sonic array sensors for fall detection,” in Proc. Asia–Pacific Sig-
near field imaging,” Int. J. Smart Sens. Intell. Syst., vol. 2, no. 1, nal Inf. Process. Assoc. Annu. Summit Conf. (APSIPA), Dec. 2016,
pp. 156–175, 2009. pp. 1–4.

Authorized licensed use limited to: Auckland University of Technology. Downloaded on October 10,2020 at 06:20:29 UTC from IEEE Xplore. Restrictions apply.
SINGH et al.: SENSOR TECHNOLOGIES FOR FALL DETECTION SYSTEMS: A REVIEW 6919

[141] C. Nadee and K. Chamnongthai, “Ultrasonic array sensors for mon- Saeed Ur Rehman (Senior Member, IEEE)
itoring of human fall detection,” in Proc. 12th Int. Conf. Electr. received the M.E. and Ph.D. degrees in electrical
Eng./Electron., Comput., Telecommun. Inf. Technol. (ECTI-CON), and electronic engineering from the University
Jun. 2015, pp. 1–4. of Auckland, New Zealand, in 2009 and 2015,
[142] Y.-T. Chang and T. K. Shih, “Human fall detection based on event respectively. He is a Senior Lecturer with the
pattern matching with ultrasonic array sensors,” in Proc. 10th Int. Conf. College of Science and Engineering, Flinders
Ubi-Media Comput. Workshops (Ubi-Media), Aug. 2017, pp. 1–4. University, Adelaide, Australia. Prior to joining
[143] A. Ghosh et al., “On automatizing recognition of multiple human Flinders University, he has worked as a Lecturer
activities using ultrasonic sensor grid,” in Proc. 9th Int. Conf. Commun. at the Auckland University of Technology from
Syst. Netw. (COMSNETS), Jan. 2017, pp. 488–491. 2017 to 2019 and a Senior Lecturer/Academic
[144] G. Koshmak, A. Loutfi, and M. Linden, “Challenges and issues in Leader at the Unitec Institute of Technology from
multisensor fusion approach for fall detection,” J. Sensors, vol. 2016, 2014 to 2017, Auckland, New Zealand. His doctoral research was
pp. 1–12, Dec. 2016. focused on physical layer security of wireless networks and ME on the
[145] A. Yazar, F. Erden, and A. E. Cetin, “Multi-sensor ambient assisted MAC of WSN. He has published articles in high-quality journals and
living system for fall detection,” in Proc. IEEE Int. Conf. Acoust., international conferences with an H-index of ten (google scholar). His
Speech, Signal Process. (ICASSP), May 2014, pp. 1–3. current research interests include physical layer security (in wireless
[146] D. Litvak, I. Gannot, and Y. Zigel, “Detection of falls at home using devices and embedded systems), privacy-aware embedded systems in
floor vibrations and sound,” in Proc. IEEE 25th Conv. Electr. Electron. health care, and visible light communication.
Eng. Isr., Mar. 2008, pp. 514–518.
[147] H.-W. Tzeng, M.-Y. Chen, and J.-Y. Chen, “Design of fall detection
system with floor pressure and infrared image,” in Proc. Int. Conf. Syst.
Sci. Eng., Jul. 2010, pp. 131–135.
[148] A. M. Sabatini, G. Ligorio, A. Mannini, V. Genovese, and L. Pinna,
“Prior-to-and post-impact fall detection using inertial and barometric
altimeter measurements,” IEEE Trans. Neural Syst. Rehabil. Eng.,
vol. 24, no. 7, pp. 774–783, Jul. 2016.
[149] L.-J. Kau and C.-S. Chen, “A smart phone-based pocket fall accident
detection, positioning, and rescue system,” IEEE J. Biomed. Health
Informat., vol. 19, no. 1, pp. 44–56, Jan. 2015.
[150] G. Fortino and R. Gravina, “Fall-MobileGuard: A smart real-time fall Sira Yongchareon (Senior Member, IEEE)
detection system,” in Proc. 10th EAI Int. Conf. Body Area Netw., 2015, received the Ph.D. degree in IT from the
pp. 44–50. Swinburne University of Technology, Australia,
[151] K. Ozcan, S. Velipasalar, and P. K. Varshney, “Autonomous fall detec- in 2012. He is a Senior Lecturer with the IT and
tion with wearable cameras by using relative entropy distance mea- Software Engineering Department, Auckland
sure,” IEEE Trans. Human-Mach. Syst., vol. 47, no. 2017, pp. 31–39. University of Technology, New Zealand. He has
Nov. 2017. over 40 research publications in referred journals
[152] K. Ozcan and S. Velipasalar, “Wearable camera-and accelerometer- and reputable international conferences, includ-
based fall detection on portable devices,” IEEE Embedded Syst. Lett., ing Information Systems (Elsevier), WWWJ,
vol. 8, no. 1, pp. 6–9, Mar. 2016. ACM TMIS, Computers in Industry, Computing,
[153] J. Wang, Z. Zhang, B. Li, S. Lee, and R. S. Sherratt, “An enhanced fall
CCPE, Sensors, IEEE TrustCom, IEEE SCC,
detection system for elderly person monitoring using consumer home BPM, CoopIS, and WISE. His research interests include the Internet
networks,” IEEE Trans. Consum. Electron., vol. 60, no. 1, pp. 23–29, of Things, ubiquitous computing, ambient intelligence, business process
Feb. 2014. analytics, and service-oriented computing. He is an ACM member.
[154] D. McIlwraith, J. Pansiot, and G.-Z. Yang, “Wearable and ambient
sensor fusion for the characterisation of human motion,” in Proc.
IEEE/RSJ Int. Conf. Intell. Robots Syst., Oct. 2010, pp. 5505–5510.
[155] A. Ariani, S. J. Redmond, D. Chang, and N. H. Lovell, “Simulated
unobtrusive falls detection with multiple persons,” IEEE Trans. Biomed.
Eng., vol. 59, no. 11, pp. 3185–3196, Nov. 2012.
[156] C. Doukas, L. Athanasiou, K. Fakos, and I. Maglogiannis, “Advanced
sound and distress speech expression classification for human status
awareness in assistive environments,” J. Inf. Technol. Healthcare, vol. 7,
no. 2, pp. 111–117, 2009. Peter Han Joo Chong (Member, IEEE) received
[157] H. M. L. Daudt C. van Mossel, and S. J. Scott, “Enhancing the scop- the Ph.D. degree from the University of
ing study methodology: A large, inter-professional team’s experience British Columbia, Canada, in 2000. He was
with Arksey and O’Malley’s framework,” BMC Med. Res. Methodol., an Associate Professor (tenured) with the
vol. 13, no. 1, p. 48, Mar. 2013. School of Electrical and Electronic Engineering,
[158] D. Levac, H. Colquhoun, and K. K. O’Brien, “Scoping studies: Nanyang Technological University, Singapore.
Advancing the methodology,” Implement. Sci., vol. 5, no. 1, p. 69, From 2011 to 2013, he was an Assistant Head
Dec. 2010. of the Division of Communication Engineering.
[159] H. Arksey and L. O’Malley, “Scoping studies: Towards a methodolog- From 2013 to 2016, he was the Director of
ical framework,” Int. J. Social Res. Methodol., vol. 8, no. 1, pp. 19–32, the Centre for Infocomm Technology (Infinitus).
Feb. 2005. From February 2001 to May 2002, he was a
Research Engineer with the Nokia Research Center, Finland. From
July 2000 and January 2001, he worked at the Advanced Networks
Division, Agilent Technologies Canada Inc., Canada. He is currently a
Full Professor, the Head of the Department of Electrical and Electronic
Anuradha Singh received the [Link]. degree from the Engineering, and the Associate Head of School (Research) with the
Zakir Hussain College of Engineering and Technology, A.M.U, School of Engineering, Computer and Mathematical Sciences, Auckland
in 2003, and the [Link]. degree from Sharda University, Greater Noida, University of Technology, New Zealand. He is also an Adjunct Professor
in 2017. She is currently pursuing the Ph.D. degree with the Department with the Department of Information Engineering, Chinese University
of Electrical and Electronic Engineering, Auckland University of of Hong Kong, Hong Kong. His research interests are in the areas
Technology. She has worked at Syscom Technologies and Honeywell of mobile communications systems, including MANETs/VANETs, V2X,
for more than three years. Her research interests include embedded and the Internet of Things/Vehicles and artificial intelligence for wireless
systems for health care, analog electronics, and VLSI systems. networks.

Authorized licensed use limited to: Auckland University of Technology. Downloaded on October 10,2020 at 06:20:29 UTC from IEEE Xplore. Restrictions apply.
View publication stats