Hindawi

Journal of Healthcare Engineering

Volume 2022, Article ID 2212020, 21 pages
https://doi.org/10.1155/2022/2212020

Research Article
An Ultrasonic-Based Sensor System for Elderly Fall
Monitoring in a Smart Room

Chokemongkol Nadee and Kosin Chamnongthai

Department of Electronic and Telecommunication Engineering, Faculty of Engineering, King Mongkut’s University of Technology,
Tonburi, Bangkok, Tailand

Correspondence should be addressed to Kosin Chamnongthai; [email protected]

Received 7 December 2021; Revised 20 April 2022; Accepted 11 May 2022; Published 7 November 2022

Academic Editor: Ikram Ud Din

Copyright © 2022 Chokemongkol Nadee and Kosin Chamnongthai. Tis is an open access article distributed under the Creative
Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the
original work is properly cited.

To reduce the risk of elderly people falling in a private room without relying on a closed-circuit television system that results in
serious privacy and trust concerns, a fall monitoring system that protects the privacy and does not monitor a person’s activities is
needed. An ultrasonic-based sensor system for elderly fall monitoring in a smart room is proposed in this study. An array of
ultrasonic sensors, whose ranges are designed to cover the room space, are initially installed on a wall of the room, and the sensors
are rotated to transmit and receive ultrasonic signals to measure the distances to a moving object while preventing ultrasonic
signal interference. Distance changes measured by ultrasonic sensors are used as time-independent patterns to recognize when an
elderly person falls. To evaluate the performance of the proposed system, a sensor system prototype using long short-term memory
was constructed, and experiments with 25 participants were performed. An accuracy of approximately 98% was achieved in this
experiment using the proposed method, which was a slight improvement over that of the conventional method.

1. Introduction preferred [4]. As elderly individuals are more susceptible to

fatal fall [5], they need immediate assistance when a fall
Te world population, which was approximately 7.7 billion occurs. Terefore, automatic fall detection and alarm sys-
people in mid-2019, is gradually increased [1]. Te number tems, which are used instead of caregivers to continuously
of elderly people will dramatically rise to approximately monitor elderly individuals in rooms, must be developed
11.1%–18.6% of the total population in the next 30 years [2] under the conditions of privacy and trust.
due to successful birth control policies and advanced Te authors of this study proposed the development of
medical care services. As a result, the ratio of the working an elderly fall monitoring system that can be installed in a
population to the elderly population is rapidly decreasing. private room. Te basic conditions of this system include
Tis will lead to a worldwide labor shortage for many in- trust, privacy, and safety. We attempted to utilize ultrasonic
dustries, including elderly care. At the beginning of the sensor arrays installed on the ceiling and walls of a room to
twenty-frst century, the European commissioner launched a detect human falls, and experiments were performed with a
project called the TeleCARE consortium [3] to design and static model [6–8]. A small room (30 × 30 cm2) with nine
develop a virtual community infrastructure platform for ultrasonic sensors was used to test the static model [6].
elderly support. Many other countries in other regions also Ultrasonic interference between neighboring sensors and
launched similar projects. sensor calibration were found to be problems. Tese
In the community infrastructure, the bedroom, living problems were considered, and the sensors were adjusted to
room, and restroom are the main rooms where elderly mitigate interference [7]. However, coverage gaps between
people spend most of their time, especially when they are at neighboring sensors in the array were found. Te overlap of
home, and it is assumed that during this time, privacy is neighboring sensors was proposed to address these gaps, and
2 Journal of Healthcare Engineering

experiments were performed using the model in a larger ultrasonic sound. Terefore, sound can be safely used to
room (60 × 60 cm2) [8]. It was confrmed that the methods measure distance using its refected wavelength [65]. As
proposed in this study could be used to recognize falls using ultrasonic sound is comparatively safe for the human body,
a static model. We also attempted to construct a prototype it is recommended in this study as a good technology to
system for human fall detection. Te duration of a fall is develop an automatic fall detection system. In fact, ultra-
approximately 0.7 s. Accordingly, the system must be im- sonic signals, which are mainly applied in point-to-point
proved to be able to complete all required processes during distance measurements by directional characters, currently
this short period of time. Terefore, we reconsidered the cannot be developed at a high resolution for object recog-
system with respect to the hardware and software, improved nition due to the constraints of its beam. Tis observation is
the hardware, installed an ultrasonic sensor on a sidewall, conversely a merit for this research problem, since, in
and developed an algorithm to sense distances to humans in general, elderly people prefer privacy in their living envi-
a point-by-point manner to detect falls using the state ronment. Te ultrasonic sensor becomes a solution to the
transition concept without identifying the specifc behaviors research problem.
of the person. In the research and development of the automatic fall
Tis study is organized as follows. Te related works and detection systems related to ultrasonic sensors, as shown in
smart room scenario, including the analysis of ultrasonic Table 2, Toshio Hori et al. [66] presented ultrasonic sensors,
array sensors, are presented in Sections 2 and 3, respectively. which utilized the speed of falling in a vertical direction, for
Our proposed design method is described in Section 3. elderly people and their caregivers in a nursing home. Tis
Experiments and results are reported in Section 4, and a method worked, especially in vertical falling applications,
discussion is presented in Section 5. Finally, the study is but might not be suitable for some complicated cases of
concluded in Section 6. diagonal falling, which is a limitation. Yirui Huang et al. [67]
suggested a method to detect falls and remote activity using
2. Related Works ultrasonic sensors. Tis work focused on a cost-efective and
intelligent hardware design for individual ultrasonic sensors.
Based on the research problem of elderly fall detection, many Tis method also utilized the speed and level of falling in
researchers have attempted to develop algorithms and each sensor so that it was not matched with complicated fall
systems to monitor elderly individuals and detect their falls. patterns. Chang and Shih [68] proposed human fall de-
As shown in Table 1, research works related to the devel- tection based on event pattern matching with ultrasonic
opment of fall detection systems are generally divided into array sensors. Tis method may be used to detect many
three approaches [37, 38]: vision-based approaches human fall patterns. However, the sensors were installed at
[14–20, 39–41], wearable device approaches [21–38], and the height of the human leg, which may not be robust
ambience sensor approaches [6–8, 39–41, 42–69]. Te vi- enough to detect all fall patterns. Ghosh et al. [69] proposed
sion-based approach is considered an excellent approach UltraSense, which is used to identify human activity using a
due to the benefcial 3D shape information it provides. heterogenous ultrasonic sensor grid for smart home envi-
Accurate results have also been achieved using the wearable ronments. Tis system identifed human activity well, which
device approach because of the dynamic personal position is not preferred from a privacy viewpoint. Moreover, the
information obtained. Due to the sensors and their locations, ultrasonic sensor grid was installed on the ceiling of the
sensed information using ambience sensor approaches is room. Since height information is used, the diference be-
mainly a point-to-point approach, which is a limitation. tween a fall and some normal activities in the vertical di-
However, the methods in this group were not originally rection may confuse the system.
utilized to clearly recognize an object shape. Terefore, these
methods are considered to better protect resident privacy. 3. Smart Room Scenario
Previous fall detection systems using ambience devices,
including infrared (IR) [39–41], IR and ultrasonic [42, 43], According to statistical data from the United Nations (UN)
radar [44–51], Wi-Fi [52–57], sound [58–62], and ultrasonic in 2019 [70], all societies in the world are in the midst of a
[49–55] systems, have been developed, and their advantages “longevity revolution,” and the number of elderly people
and disadvantages are discussed as follows. IR is mainly used worldwide has dramatically increased year by year. In ad-
to confrm the absence of objects [39–41, 42, 43], but it dition, children and relatives who may closely care for el-
cannot be used to measure distance, which makes 3D re- derly family members do not reside with them in the same
construction difcult. Moreover, IR has been proven to be household. Approximately half of adults aged 60 and older
harmful to the human eyes and body [39–41, 42, 43]. As it is live alone and approximately one-third live with a spouse
widely used, Wi-Fi is convenient. Radar is an excellent only (without children or relatives) in North America and
technique for scanning moving objects for a long time. Europe [71]. When an elderly person on their own has a vital
However, both Wi-Fi and radar operate at high frequencies, accident, such as a fall, it is difcult to obtain assistance as
which harm the human body and especially afect heart other people may not have noticed that the accident oc-
attack patients who use pacemakers [63, 64]. curred. Unless someone notices and helps the person in a
Other than the hearable frequency range, which causes timely manner, an unexpected death may occur.
noise harm, sound does not cause any serious harm to the To solve this problem, an automatic monitoring system
human body. Furthermore, the humans cannot hear an that protects user privacy and does not monitor activities
Journal of Healthcare Engineering 3

Table 1: Research works related to image-based and wearable fall detection systems.
Feature extraction
Algorithm name Sensor and equipment Classifcation algorithm Accuracy/error
technique
1. Image-based fall detection systems
Hernandez [11] 2D camera and OpenCV Rectangle enclosing Treshold-based Accuracy 85.37%
Acceleration and angular Accurately model an
Lin [12] GMM and MHI —
acceleration ellipse
MHI and ellipse Both ellipse approximation Accurately model an
Basavaraj [13] Accuracy 86.66%
approximation and MHI ellipse
ShanShan [14] Semi-contour distances Points on the vertical line SVM —
Sensitivity 77.98%,
Histogram of oriented
Chen [15] Depth camera SVM and ANN specifcity 87.58%,
gradient (HOG)
accuracy 82.84
Optical fow combining with Optical fow and residual
Cai [16] Softmax classifer Accuracy 92.6%
wide residual network network
Transfer learning, optical fow Displacement vector optical Fully connected neural Sensitivity 95%, specifcity
Marcos [17]
algorithm fow algorithm network 96%
Temporal motion feature,
Lu [18] 3D CNN was used SVM Accuracy 99.3%
3D CNN
Background subtraction Sensitivity 96% specifcity
Miguel [19] Low-cost device, Raspberry Pi KNN
Kalman flter 97%
Accuracy 99.2%,
Major and minor semi-axis of Multilayer perceptron
Lotf [20] Spatiotemporal sensitivity 99.5%,
ellipse ftting silhouette neural network
specifcity 97.3%
2. Wearable fall detection systems
BLE module and a
Freitas [21] Acceleration Fall signal to Web app —
microprocessor
Magnetic, angular rate, and Yaw, pitch, and roll, Accuracy 90.37%,
Pierleoni [22] Treshold and SVM
gravity (MARG) sensor Madgwick orientation flter sensitivity 80.74%
Using tri-dimensional Dynamic threshold Accuracy 97.40%,
Otanasap [23] Acceleration, ADL value
accelerometer model sensitivity 99.48%
Accuracy fall forward 95%,
Using tri-dimensional
Kurniawan [24] Yaw, pitch, and roll, alpha SVM accuracy fall backward
accelerometer
75%
Wearable device that Exponential smoothing
Chu [25] Acceleration —
combined BLE gray model (ESGM)
Shahiduzzaman [26] Smart helmet Biomedical sensing data SVM Accuracy 96.67%
Sensitivity 90% and
Nari [27] Accelerometer and gyroscope ACC and gyro Treshold-based
specifcity 86.7%
Fourfold cross-
Heart rate sensor and Cluster analysis-based user- Accuracy 97.51%,
Nho [28] validation, 13-
accelerometer adaptive fall detection sensitivity 99.17%
dimensional
Crowdsourcing-based Acceleration, inclination
Chen [29] Treshold-based Accuracy 97%
adaptive datasets angle
Tang [30] Radar sensors on shoe Distance-to-obstacle Treshold-based —
Extended nearest
Djelouat [31] Compressed sensing Acceleration Accuracy 91.73%
neighbor
Neural network classifcation 43 features, FFT, principal Multilevel fuzzy min- Sensitivity 97.29%,
Jahanjoo [32]
algorithm component analysis max neural network specifcity 98.7%
Magnetometer accelerometer Acceleration, Euler angle Accuracy 100% sensitivity
Mao [33] Treshold-based
gyroscope (orientation) 100%
Multiple power-saving
Ang [34] Acceleration Decision tree classifer Sensitivity 91%
algorithms
A neural network Linear and angular
Purushothaman [35] Neural network —
classifcation algorithm acceleration
Eight features from
Khojasteh [36] Treshold optimization SVM, RBS, and DT Accuracy 95.15%
acceleration
SVM-KNN with
Statistical features from Sensitivity 93%, specifcity
De Quadros [37] Madgwick’s decomposition Madgwick’s
acceleration 98%
decomposition
Accuracy 99.9%,
Two-segment feature Statistical 12 feature vectors Artifcial neural network
Saleh [38] sensitivity 99.1%,
extraction from acceleration and SVM
specifcity 99.9%
4 Journal of Healthcare Engineering

Table 2: Research works in ambience sensor-based systems.

Algorithm name Sensor and equipment Feature extraction technique Classifcation algorithm Accuracy/error
Ambience sensor-based fall detection systems
1.1 IR sensor
Floor pressure and infrared Average image pixel value
Tzeng [39] Image thresholding Accuracy 98.3%
image (mean)
K-nearest neighbor, GM- Sensitivity 98%,
Guan [40] Infrared signal-based Multi-sensor time
HMM, SVM specifcity 93%
Ogawa [41] IR array sensor Temperature distribution × 20 Machine learning Accuracy 97.75%
Multilayer perceptron
Asbjørn [42] IR array sensor 80 × 60 thermal array Accuracy 96.73%
model
Chen [43] Infrared arrays and ultrasonic 8 × 8-pixel thermal, RMS values SVM Accuracy 90.3%
1.2 Radar sensor
SVM and WT-based
Amin [44] Real fall data used MFCC, SWT —
algorithm
Liu [45] Doppler radar motion MFCC SVM and KNN —
Compensates the drawbacks of
Tomii [45] Multiple Doppler sensors SVM and KNN Accuracy 95.5%
mono-Doppler sensor
Pursuit decomposition, time-
Wu [47] Radar signal PCA, HMM, SVM —
frequency
Walking sequence
Wang [48] Pulse Doppler radar for passive Angle θ and ϕ afect —
selection/speed estimation
Gadde [49] Radar signal Time scale Wavelet transform —
Wu [50] Radar technology Extreme frequency magnitude Bayesian, SVM —
Su [51] Doppler radar MFCC features Wavelet transform —
1.3 Wi-Fi device
Sensitivity 92%,
Wang [52] Wi-Fi device Channel state information (CSI) SVM
specifcity 92%
Two-dimensional phase
Khan [53] Passive Wi-Fi sensing, Vi Wi Tremor classifcation Accuracy 98%
extraction system
Gu [54] Wi-Fi device Channel state information (CSI) Activity recognition Accuracy 94.58%
Wi-Fi, accelerometer, and foor
Ramezani [55] CSI, STD, MAD, IR, SRS SVM —
vibration
Cheng [56] Wi-Fi signals Channel state information (CSI) CNN, LSTM, GRU —
Hu [57] Wi-Fi Channel state information (CSI) SLN-DTW Accuracy 96%
1.4 Acoustic sensor
Pattern recognition, event
Zigel [58] Acoustic signal Event classifcation —
segmentation
Li [59] Circular array of 8 microphones MFCC Nearest neighbor AROC 0.98
Beamforming to increase signal Sensitivity 100%,
Li [60] MFCC Nearest neighbor
strength specifcity 97%
Li [61] 8-Microphone circular array iVAT clustering and GA-based Nearest neighbor —
Te spectrogram, MFCCs, LPC,
Chefena [62] Smartphone ANN Accuracy 98%
and MP
1.5 Ultrasonic sensor
Ultrasonic sensor network and Tracking a head of moving
Yoshio [66] Pattern trajectories —
foor mat sensor human
Huang [67] Ultrasonic sensor array, FPGA Distance, time duration Pattern matching —
Chang et al. [68] Arduino ultrasonic array Time energy SVM Accuracy 98%
Nadee et al. [6] Ultrasonic array: ceiling, sidewall Distance, time duration Treshold-based Accuracy 92%
Ultrasonic array: two
Nadee et al. [7] temperature sensor error Distance, time duration Treshold-based algorithm Accuracy 93%
correction
Nadee et al. [8] Ultrasonic array: octagonal array Distance, time duration Treshold-based algorithm Accuracy 94%
HC-SR04, LV-Max Sonar-EZ0
Ghosh et al. [69] Distance, time duration Decision tree Accuracy 90%
sensor
Ultrasonic array: MaxSonar
Tis work Distance, time duration LSTM Accuracy 98%
MB1010

should be installed in every house. In addition, the system recognize and record activities, is installed on the wall of a
must provide an immediate alert whenever a fall occurs. In room. Terefore, an ultrasonic node was selected as the
our proposed scenario, an automatic monitoring system, sensor in this study to continuously measure distances from
with sensors and devices that do not directly visually the ultrasonic nodes to the person in the room in a point-to-
Journal of Healthcare Engineering 5

Room

Human Body Module Control Point

Sensor Array

Figure 1: A human body and its control points sensed by a sensor array.

T1 T2 T3 T4 T5 T1 T2 T3 T4 T5

(a) (b)

T1 T2 T3 T4 T5 T1 T2 T3 T4 T5

(c) (d)

Figure 2: Control points during falls. (a) A forward fall. (b) A backward fall. (c) A sideway fall. (d) A fall from a chair.

point manner. In addition, external memory, which can the control point image. Elderly people may prefer that
store data permanently, is not utilized in this system to activities done during their private time are not monitored,
protect activity data. As shown in Figure 1, the fall moni- as an absolute condition to develop the fall monitoring
toring system consists of a processor, microcontroller, and system.
sensors that are installed in a room, which is called a smart Suppose that a video clip and an image of control points
room in this study. Te system may immediately alert of humans are not allowed to be used as input data in the fall
caregivers, relatives, and children via Wi-Fi, the cloud, and monitoring system. Instead, a point representing human
mobile phones whenever a fall occurs. body movement per frame, which contains the least amount
Based on the scenario of a smart room for elderly people, of data in a frame and from which, it is almost impossible to
privacy-based and distance-based human fall detection and determine activities, should be considered to enable fall
sensor blind zone are analyzed as follows. recognition. During a fall, a point that is closest to a sensor is
sensed, and this sensor may change based on the movement
of the falling person. Distances from those points closest to a
3.1. Privacy-Based Human Fall Detection. According to the sensor that is continuously collecting information during a
studies in [72], the human body comprises moving-inde- fall can be categorized into patterns based on the type of fall.
pendent parts or modules, and a point located on a module Figures 3(a)–3(d) show some examples of video frames
called a control point can represent the module position, as (upper row) and distances between sensors and a person
shown on the right side of Figure 1. If a control point (represented as a graph). Te distance changes during a fall
representing a module is detected and the distances from based on diferent fall types, e.g., forward, backward, and
the sensors to the control points are measured, the body of sideway falls and a fall from a chair, can be diferentiated as
a person is detected and monitored. If all control points of patterns in the graph. Tese diferent patterns can be used to
the person are continuously detected and the distances are classify a fall and a nonfall as well as recognize fall types. If a
measured all the time, then the person’s behavior is also classifer is trained with these distance change patterns in
monitored. Terefore, the patterns of control points when a advance, pattern matching between these trained patterns
fall occurs (e.g., a forward fall, a backward fall, a sideway and input data can always be used to classify the input as
fall, and a fall from a chair) can be observed and analyzed, either a fall or a nonfall.
as shown in Figure 2. Tese patterns can be trained in
advance and can possibly be used to recognize falls in
elderly fall monitoring systems. Although the control 3.2. Distance-Based Human Fall Detection. To apply the
points of a person do not show real pictures such as a video abovementioned concept to the human fall problem, since
clip captured by a closed-circuit television (CCTV), it may the time period for a human fall is as short as approximately
be possible to estimate the behaviors of the person even by 0.7 s [72], the whole room space should be sensed and
6 Journal of Healthcare Engineering

Tn Tn+1 Tn+2 Tn+3 Tn+4 Tn Tn+1 Tn+2 Tn+3 Tn+4

2,1
250
2,2
250
200

200
150

Distance (cm)
Distance (cm)

150
100

100
50

50
0
0 100 200 300 400 500 600 700 800 0 200 400 600 800 1000
Time (msec) Time (msec)

(a) (b)

Tn Tn+1 Tn+2 Tn+3 Tn+4 Tn Tn+1 Tn+2 Tn+3 Tn+4

2,1
250
2,2
250
200

200
Distance (cm)

150
Distance (cm)

150

100
100

50
50

0
0 100 200 300 400 500 600 700 0
0 100 200 300 400 500 600 700
Time (msec) Time (msec)

(c) (d)
Figure 3: Distance patterns during falls. (a) Pattern of a forward fall. (b) Pattern of a backward fall. (c) Pattern of a sideway fall. (d) Pattern of
a fall from a chair.

monitored so that the fall detection system can be pro- orderly manner in a matrix form, as shown in Figure 5. In
cessed and an alarm immediately activated during this this case, coverage ranges almost cover the sensing wall, but
period. In the case that a distance sensor is installed on a blind spaces or gaps exist between neighboring cells. To fll in
sidewall to measure the shortest distance from a sensor to a the blind spaces, ultrasonic nodes should be shifted in an
human at a point (s) in the room space, as shown in orderly manner along a straight line and simultaneously the
Figure 4, the change in measured distances during the fall given distance between consecutive ultrasonic nodes should
duration (Tfl ) is mathematically expressed as the following be maintained to prevent interference, as shown in
fall pattern (F). Figure 6(a). Additionally, the sensors can also be shifted in
Tfl zigzag scanning lines to maintain balance in the horizontal
F�􏽚 f st 􏼁 dt. (1) and vertical directions. As an example, an ultrasonic node or
t�0
more than one ultrasonic node can be scanned along a zigzag
To cover the whole room space with ultrasonic signals line, as illustrated by the black and red dashed arrows in
without interference, multiple sensors are installed in an Figure 6(b).
Journal of Healthcare Engineering 7

ultrasonic rays at a distance smaller than the width of the

ft(s) human as the design condition. Tis guarantees the detec-
tion of human falls even in a blind zone. Tus, the range (d)
S ft+1(s) between consecutive ultrasonic nodes can be obtained.
Sensor S
B × 2cos(90 − θ/2)
d� . (5)
cos(θ/2)
In addition, the power transmitted by a string wave (E)
Figure 4: Pattern of distance changes. can be determined as follows [73, 74]:
1
E � 􏼐μω2 A2 ]􏼑 � 2μπ2 f2 A2 ], (6)
M Sensing wall 2
Blind space
Sensing space where μ, ω, A, and v represent the mass per unit length of the
string, angular frequency of the wave, wave amplitude, and
N wave propagation velocity, respectively.
sensors

I 4. Proposed Ultrasonic-Based Human Fall

Monitoring System
Sensor side wall
Based on the abovementioned concept, a fall detection
Figure 5: Distance sensing matrix.
design and implementation method using ultrasonic sensors
for the monitoring system is explained in this section. It is
assumed that the system must be not only nonintrusive,
Terefore, a human fall pattern can be expressed by
noninvasive, and device-free but also protect the user’s
distance changes during a fall:
privacy. In the following, the system design and imple-
Tfl N−1 M−1 mentation are divided into hardware and software parts.
F�􏽚 􏽚 􏽚 f smn 􏼁 dmdndt, (2)
t�0 n�0 m�0

where m and n are the ultrasonic node positions in the 4.1. Hardware Design. Te hardware system mainly consists
horizontal and vertical directions, respectively, and Tfl is the of two units, a sensor array and a signal processing unit. Te
human fall duration. sensor array installed on a wall must be designed to cover the
Tis fall pattern can then be converted into a digital form whole room with the smallest blind zone. In addition, the
to realize the results using a digital computer. signal processing unit must provide enough ports for re-
Tfl ceiving signals from all sensors and must be designed to have
⎨ N−1 M−1
⎧ ⎬
⎫ enough ability to process those signals. Te design of the
F � ⎩ 􏽘 􏽘 f smn 􏼁⎭ . (3)
n�0 m�0
sensor array and signal processing unit is explained as
t�0
follows.

3.3. Geometric Concept for Sensor Array Installation. 4.1.1. Sensor Array. Suppose the scale of a room in which the
From the basic condition of the ultrasonic signal application, ultrasonic sensor array for the human fall monitoring system
which must be generated without interference, the coverage is installed is M × N × l. Ultrasonic sensor nodes should be
range (k) on the sensing wall, as shown in Figure 7, can be geometrically installed on a sensor wall (M × N) in the room
obtained: under the condition of noninterference. Te range between
θ consecutive ultrasonic nodes in the horizontal (φN ) and
k � 2l cos , (4) vertical (φM ) directions can be simply determined.
2
θ θ
where l and θ are the room width and sensor transmission φN ≥ 2l tan , φM ≥ 2l tan . (7)
2 2
angle, respectively.
Te coverage range can be used to determine the number If B represents the blind zone for human sensing, which
of ultrasonic generating nodes in the frst static step (t1). has to be determined in advance, the range between con-
Since it is assumed that nodes were shifted to cover all blind secutive sensors in the array (d) can be obtained based on
zones on the wall as previously described, the distance (5).
between a pair of ultrasonic sensor nodes will be calculated In the frst step, the number of sensors in the horizontal
based on the condition of an unavoidable blind zone (B), as (zN ) and vertical (zM ) directions for sensing the distance in
shown in Figure 8. a frame, which must be limited due to the interference of
To apply this concept in fall detection applications, users ultrasonic signals from diferent sensors in a frame, can be
should consider setting blind zones between consecutive simply determined.
8 Journal of Healthcare Engineering

M M

Ultrasonic node

N N
Coverage

(a) (b)
Figure 6: Scanning lines: (a) straight lines and (b) zigzag lines.

θ
l tan shifting mainly depends on the average human fall duration
2
(Tfl ). If a sensing process takes a duration of time (Tu ), a
Sensing side wall shifting range (d), which is regarded as the dense sensor
k node range, can be calculated.
Tfl
d � k/􏼠 􏼡. (9)
Tu

Te total number of dense node sensors (Si ) is therefore

determined.
θ
θ θ θ
M×N
B 2
θ θ Si � . (10)
d2
t1 t2 t3 t1 t2 t3
Sensor side wall
d
sensor
4.1.2. Signal Processing Unit. Te signal processing unit
consists of a power supply unit, an ultrasonic sensor array,
Figure 7: Coverage and shifting ranges. a microprocessor, memory, and a classifer, as shown in
Figure 8. Te ultrasonic sensor array, which is supplied
power by the power supply unit, always senses a moving
Power
supply object in the room. Analog signals representing the dis-
tance from ultrasonic sensors to a moving object are
transmitted to the microprocessor. In the microprocessor,
Ultrasonic the analog signals are translated into distance data and
Sensor Microprocessor Classifier Output
Array stored in terms of a matrix in the memory. Te matrix of
distance data is fnally fed to the classifer as a feature for fall
classifcation.
Memory

In the power supply unit, the power needed to drive an

ultrasonic sensor can be calculated based on equation (6). In
Figure 8: Signal processing unit.
equation (6), μ, f, A, and ] can be determined as the mass
per unit length of the string [75], the frequency of the ul-
trasonic signal (defned above the sound frequency or set to
20 kHz), the amplitude of the ultrasonic signal, and the
N
zN ≥ , ultrasonic velocity [76], respectively.
2l tan θ/2 Since this study proposes the utilization of the change
(8)
M patterns of distance from the ultrasonic sensor nodes to the
zM ≥ . closest point of a person who has fallen, the ultrasonic
2l tan θ/2
signals should be transmitted to refect the falling person in
Based on our proposed concept to shift the active sensors as many round trips as possible. Te number of scanning
in the dense ultrasonic sensor array instead of dynamically frames (τ) for ultrasonic signal transmission in one second
scanning the ultrasonic sensors, the time duration for can be simply estimated:
Journal of Healthcare Engineering 9

BEGIN
Input: m: number of nodes in row, n: number of nodes in column,
disA [row][col]: distance from node A to object
disB [row][col]: distance from node B to object, X: input data for classifier
REPEAT
PROCEDURE Zigzag scanning ⊳Zigzag scan process
for row � 1 to m do ⊳Conversion of 2D to 1D matrix
for col � 1 to n do
CALCULATE: X[2{(row − 1) · n} + col] � disA [row][col]
CALCULATE: X[2{(row − 1) · n} + (col + 1)] � disB [row][col]
End for
End for
COMPUTE: Classifcation based on input data of measured distances
UNTIL all signal samples
END

ALGORITHM 1: Comprehensive system of the proposed fall detection and classifcation.

1 matrix in steps 7–12, and it is fed to a classifer for fall

τ� , (11) classifcation in Step 13. Te details of the rotation of active
(2l/v) + P
ultrasonic sensors and fall classifcation are explained in
where v, P, and l are the ultrasonic signal velocity, fre- Sections 4.2.1 and 4.2.2, respectively.
quency, processing time, and room depth, respectively.
In human fall detection and classifcation, the number of
scanning frames is one of the crucial keys to fundamentally 4.2.1. Rotation of Active Ultrasonic Sensors. To implement a
guaranteeing quality. Te number of scanning frames is software unit of the fall monitoring system, multiple ul-
practically an initial condition to design the signal pro- trasonic nodes in an array are expected to simultaneously
cessing unit. Although the more scanning frames there are, sense distances from the sensors to objects. However, in-
the more robust the system is, and users may select an terferences among ultrasonic signals may occur and lead to
appropriate set of devices that is normally limited by pro- errors in the case where coverage areas overlap. To prevent
cessing time. For example, the processing time of a frame is ultrasonic signal interference, coverage areas must not
limited on an approach of real time. Tus, the number of overlap. Te possible number of ultrasonic nodes for the
scanning frames should be calculated in real time during the simultaneous distance measurement can be determined by
period of a human falling, which is approximately 0.7 s. the minimum distance between working ultrasonic nodes, as
Suppose users fx the number of scanning frames, the shown in equation (7). To efciently scan ultrasonic nodes in
maximum processing time allowed in each frame is what an array by maintaining a minimum distance, zigzag
users may need next to select electronic devices in the scanning is recommended to balance the horizontal and
hardware design step. Tus, users can expect to be able to vertical directions; notably, zigzag scanning is demonstrated
select devices by specifcations based on the allowed pro- for a pair of ultrasonic nodes in Figure 8. Suppose that a
cessing time. Te processing time per frame (P) of devices couple of scanning lines on ultrasonic nodes in an array
allowed in the system design can be calculated. simultaneously start from points A and B; they synchro-
nously move along a zigzag scanning line to the next nodes,
Nt 2l which are labeled in the same colors. If the coordinates of
P� − , (12)
τ v ultrasonic sensor nodes are represented by row (R), column
where v, τ, and l are the ultrasonic signal velocity, fre- (C), and current node counting (i), patterns of nodes moving
quency, number of scanning frames, and room depth, in a couple of zigzag scanning lines (A and B) can be logically
respectively. illustrated by rows (RA, RB) and columns (CA, CB) of A and
B.

4.2. Software Design. Te software system for retrieving ul- Group 1. Shift from the frst node (represented by a scan-
trasonic signals representing the distance to a moving object ning line A in Figure 9)
can be designed and created, as shown in Algorithm 1. First, Initially, the scanning node will be shifted to the adjacent
the initialization and declaration of variables are registered for node in the edge row as initial couple nodes. Tis logic can be
values measured by the sensors and shown in steps 2–4. Ten, simply expressed as follows.
an infnite loop (steps 5–14) is run to read a distance value on
Pattern a:
an ultrasonic sensor, store it in a matrix, shift the active sensor
to another neighboring sensor according to the zigzag di- IF I � 0 THEN RA(i+1) � RA(i), CA(i+1) � CA(i) + 1,//
rection in Step 6 for all sensor nodes, and then classify the fall. A : Shift straight right. RB(i+1) � RB(i) + 1, CB(i+1) �
Te 2D matrix of distances in Step 6 is converted into a 1D CB(i) – 1.//B : Shift diagonally left down
10 Journal of Healthcare Engineering

[0,0] [0,1] [0,2] [0,3] (R(i+1), C(i+1)) will be shifted down on the 0th column. Tis
a
c logic can be expressed as follows.
A
Pattern d:
b f IF {CA(i) � 0} and {|RA(i) – RA(i–1)| � odd} and {|CA(i)
b
[1,0] [1,1] [1,2] [1,3] – CA(i–1)| � even}
THEN RA(i+1) � RA(i+1), CA(i+1) � CA(i),//Shift
down
e RB(i+1) � RB(i) – 1, CB(i+1) � CB(i) + 1.//Shift
d e f diagonally
[2,0] [2,1] [2,2] [2,3]
B Shift in diagonal right up direction (d on scanning line A
in Figure 9)
When the previous sensing node (R(i–1), C(i–1)) located
a d out of the 0th column is shifted down to the current sensing
c
node (R(i), C(i)) located on the 0th column, the next sensing
[3,0] [3,1] [3,2] [3,3] node (R(i+1), C(i+1)) will be shifted up in the diagonal
b b direction. Tis logic can be expressed as follows.
Pattern e:
Figure 9: Zigzag scanning patterns for a pair of ultrasonic nodes.
IF {CA(i) � 0} and {|RA(i) – RA(i–1)| � odd} and {|CA(i)
Group 2. Cases where the current sensing node is located in – CA(i–1)| � even}
the 0th row (RA(i) � 0); Subgroup 2.1 Shift in the diagonal THEN RA(i+1) � RA(i) – 1, CA(i+1) � CA(i) + 1,//Shift
down direction (b on scanning line A in Figure 9) diagonally
When the previous sensing node (RA(i–1), CA(i–1)) RB(i+1) � RB(i) – 1, CB(i+1) � CB(i) + 1.//Shift diago-
located in the 0th row is shifted right to the current sensing nally right
node (RA(i), CA(i)) located in the 0th column, the next
sensing node (R(i+1), C(i+1)) will be shifted down in the
diagonal direction. Tis logic can be expressed as follows. Group 4. Cases where the current sensing node is not lo-
cated on the edge (CA(i)) ≠ 0 and (RA(i)) ≠ 0; Subgroup 4.1
Pattern b: Shift in the diagonal right direction (e on scanning line A in
IF {RA(i) � 0} and {|RA(i)– RA(i–1)| � even} and {| Figure 9)
CA(i)– CA(i–1)| � odd} When the previous sensing node (RA(i–1)), CA(i–1)) is
THEN RA(i+1) � RA(i) + 1, CA(i+1) � CA(i) – 1,//Shift shifted diagonal right up to the current sensing node (RA(i),
diagonally left down CA(i)) located out of the 0th column and out of the 0th row,
RB(i+1) � RB(i), CB(i+1) � CB(i) + 1.//Shift right the next sensing node (RA(i+1), CA(i+1)) will be shifted up in
the diagonal direction. Te logic can be expressed as follows.
Subgroup 2.2. Shift right on the 0th row (f on scanning
line A in Figure 9) Pattern f:
When the previous sensing node (RA(i–1), CA(i–1)) IF {CA(i) ≠ 0} and {RA(i) ≠ 0} and {|RA(i)
located outside of the 0th row approaches the current – RA(i–1)| � odd}and{|CA(i) – CA(i–1)| � odd}
sensing node (RA(i), CA(i)), the next sensing node (RA(i+1), THEN RA(i+1) � RA(i) – 1, CA(i+1) � CA(i) + 1,//
CA(i+1)) will be shifted right on the 0th row. Tis logic can Shift diagonally up. RB(i+1) � RB(i) + 1, CB(i+1) �
be expressed as follows. CB(i).//Shift down
Pattern c:
IF {RA(i) � 0} and {|RA(i)– RA(i–1)| � even} and {| Subgroup 4.2. Shift in the diagonal left direction
CA(i)– CA(i–1)| � odd} When the previous sensing node (RA(i–1), CA(i–1)) is
THEN RA(i+1) � RA(i), CA(i+1) � CA(i) + 1,//Shift shifted diagonal right up to the current sensing node (RA(i),
right CA(i)) located out of the 0th column and out of the 0th row,
RB(i + 1) � RB(i) + 1, CB(i+1) � CB(i) – 1.//Shift diag- the next sensing node (RA(i+1), CA(i+1)) will be shifted
onally left down down in the diagonal direction. Tis logic can be expressed
as follows.
Pattern g:
Group 3. Cases where the current sensing node is located in
the 0th column (CA(i) � 0); Subgroup 3.1 Shift down on the IF {CA(i) ≠ 0} and {RA(i) ≠ 0} and {|RA(i) – RA(i–1)| �
0th column (c on scanning line A in Figure 9) odd} and {|CA(i) – CA(i–1)| � odd}
When the previous sensing node (R(i–1), C(i–1)) located THEN RA(i+1) � RA(i) + 1, CA(i+1) � CA(i) – 1,//Shift
out of the 0th column approaches the current sensing node diagonally left down. RB(i+1) � RB(i) – 1, CB(i+1) �
(R(i), C(i)) located on the 0th column, the next sensing node CB(i) + 1.//Shift diagonally right up
Journal of Healthcare Engineering 11

Output class (12 classes)

Output
Layer

Hfc1
Hidden nodes (2/3 Hfc2
of input size per Hfc3
FC Layer Sofmax
series

Fully Connected (FC) Hfcn

and Sofmax Layer

Hidden
Layer H21 H31 Hn1
H01 H11
H22 H32 Hn2
H02 H12
Hidden nodes (2/3 H03 H13 H23 H33 Hn3
of input size per
series
H0n H2n H3n Hnn
H1n
LSTM
Layer LSTM 1 LSTM 2 LSTM 3 LSTM n

X11 X21 X31 Xn1

X12 X22 X32 Xn2
X13 X23 X33 Xn3 Input size per series
(Closest distance to
moving object)
Input
Layer X1n X24 X3n Xnn

1 2 3 M

Sequence Length (Number of frames)

Figure 10: Long short-term memory for classifcation.

Tese logics cover the possible movement of a couple of appropriately for the networks. In the input layer, the batch
current sensing nodes (RA(i), CA(i)) and (RB(i), CB(i)) to the size, input size per series, input feature, and learning rate are
next sensing nodes (RA(i+1), CA(i+1)) and (RB(i+1), trained with a number of 2n within the capability of the
CB(i + 1)) in a couple of zigzag scanning lines. As an ex- graphics processing unit (GPU) memory, total number of
ample, an algorithm for shifting a couple of sensing nodes to input data of all features, input data dimension, and ap-
the next sensing nodes in a couple of zigzag scanning lines is propriate rate for gradient descent that considers an ap-
described in Algorithm 2. propriate time, respectively, without overshooting [78].

5. Experiments and Results

4.2.2. Fall Classifcation. Due to the time limitation and time
invariance of a human fall, long short-term memory To evaluate the performance of the proposed method, a
(LSTM), which is regarded as an excellent classifer for time room with ultrasonic sensors installed was constructed, and
invariance, is utilized in this study. Distance information a data processing system was implemented based on the
from sensors to a human is assumed to be input data for the experimental specifcations, as shown in Table 4. For the
classifer. As shown in Figure 10, the node number in the experiments, a representative group of participants were
input layer is determined based on the number of sensors selected based on sex and age. Tese participants were
that sense a moving object during a human fall, which is trained to walk, sit, and fall in the room before experiments
assumed to take approximately 0.7 s, and the output nodes were performed. Te LSTM classifer was set up based on the
for the fall detection system should be set as many to one, specifcations shown in Tables 1 and 2. Photographs of an
with a number of preference choices, such as a backward fall, empty room, a room with a participant, and our design
forward fall, or walking. Te hidden layer of the LSTM is interface with a memory card are shown in Figures 11(a)–
used to set the input size per series to 2/3 of its original value 11(c), respectively. Te experiments were performed by 25
[77]. participants using some behavior criteria, including falling,
Pretests must be performed on some samples using walking, and sitting in the constructed room, as shown in
possible parameters in the training state to determine the Table 5. Te experimental results obtained in distance data of
experimental parameter settings. As shown in Table 3, the one and two points per frame reveal the human fall rec-
activation function (e.g., softmax, ReLU, sigmoid, and tanh) ognition rate based on the number of training and testing
should be pretested on some samples in advance and selected samples, as shown in Tables 6 and 7, respectively. Te results
12 Journal of Healthcare Engineering

(i) BEGIN
DATA: i: counting variable
disA , disB : dis tan ce from A an d B
i=0
READ: distances (disA [i], disB [i]) from A and B
SHIFT A straight right
SHIFT B diagonally left down
CALCULATE: i � i + 1
READ: DISTANCE distances (disA [i], disB [i]) from A and B
for i � 2 to(N × M/2) − 1 do
Switch Pattern do
case b
SHIFT A diagonally left down
SHIFT B straight right
case c
SHIFT A straight right
SHIFT B diagonally left down
case d
SHIFT A straight down
SHIFT B diagonally right up
case e
SHIFT A diagonally right up
SHIFT B diagonally right up
case f
SHIFT A diagonally right up
SHIFT B straight down
case g
SHIFT A diagonally left down
SHIFT b diagonally right up
READ: distancedisA [i], disB [i]
END for
END

ALGORITHM 2: Zigzag scanning.

Table 4: Experimental specifcations.

Table 3: Classifcation specifcation.
Devices/software/participants Specifcations
Layer Parameter Value
Computer system Aspire VX15
Batch size 12 CPU : Intel Core i7-7700HQ
Input size per series 16 GPU : NVIDIA GeForce GTX
Input
Input feature 1 dimension 1050
Learning rate 0.001 Memory size: 12 GB DDR4
LSTM hidden layer Hard disk drive: SSD 512 GB
Hidden layer Hidden node 100 Basic programming MATLAB_R2019b
Activation Participants 25 people
LSTM model Age and height ranges of 21–30 years; male: 8; female: 2;
Output layer participants 158–175 cm
Output class 12
31–40 years; male: 4; female: 0;
165–178 cm
are divided into the following groups: 20–40 years of age and 41–50 years; male: 3; female: 1;
158–175 cm
41 years of age and older. Training and testing data from the
51–60 years; male: 2; female: 0;
experiment were processed with ratios of 90 : 10, 80 : 20, 70 : 158–175 cm
30, 60 : 40, and 50 : 50 to access the accuracy of the training >60 years; male: 3; female: 2;
and testing criteria. An accuracy of approximately 99.14% 158–175 cm
was achieved using the ratio of 90 : 10 after the training and Experiment 1 (walk) Participants: 25 people: 8 groups
testing experiments. Compared with conventional methods, Experiment 2 (fall) Participants: 25 people: 4 groups
the proposed method exhibited an improvement in accuracy Room size Size: 200 × 200 cm2
by approximately 1.14%, as shown in Table 8. Examples of Sensors 16 nodes
distances captured by two node sensors in the cases of a Distance between sensors 50 cm
forward fall, backward fall, fall from a chair, and walking Blind spot distance 15.2 cm
Journal of Healthcare Engineering 13

Ultrasonic sensors

200 cm Control
signal
pins

Signal

200 cm
reflection
wall

Analog
A0-A15

(a) (b) (c)

Figure 11: Photographs of the experiments.

Table 5: Samples for training and testing.

No. Posture Category No. of samples
1 Forward fall 1 100
2 Backward fall 2 100
3 Left and right sideway falls 3 100
4 Fall from a chair 4 100
5 Walk to chair 5 25
6 Walk to sensor 6 25
7 Walk away from sensor 7 25
8 Walk in a circle 8 25
9 Walk left to right 9 25
10 Walk right to left a 25
11 Walk diagonal left to right b 25
12 Walk diagonal right to left c 25
Total of samples 600

Table 6: Experimental results based on one node.

1 node
TP FP FN
Train/test Recognition (%) Error (%) SD Age Age Age
Total Total Total
20–40 >40 20–40 >40 20–40 >40
90/10 95.74 4.26 0.52 23 34 57 1 2 3 0 0 0
80/20 89.58 10.42 0.56 43 64 107 4 6 10 1 2 3
70/30 86.43 13.57 0.63 62 94 156 7 11 18 3 4 7
60/40 79.72 20.28 0.60 77 114 191 14 24 38 5 6 11
50/50 70.33 29.67 0.65 84 127 211 24 38 62 12 15 27

Table 7: Experimental results based on two nodes.

2 nodes
TP FP FN
Train/test Recognition (%) Error (%) SD Age Age Age
Total Total Total
20–40 >40 20–40 >40 20–40 >40
90/10 98.15 1.85 0.50 24 35 59 0 1 1 0 0 0
80/20 96.47 3.53 0.53 46 70 116 2 2 4 0 0 0
70/30 93.43 6.57 0.57 67 101 168 5 7 12 0 0 0
60/40 90.58 9.42 0.57 88 130 218 7 13 20 1 1 2
50/50 86.56 13.44 0.63 104 156 260 14 21 35 2 3 5
14 Journal of Healthcare Engineering

Table 8: Performance comparison with conventional methods.

Method Recognition ratio (%) Precision (%) Recall (%) SD
Huang et al. 2012 [67] 92 — — —
Chang et al. 2017 [68] 98 — — —
Ghosh et al. 2019 [69] 90 74 100 —
Proposed method 98.15 98.30 100 0.500

Stand Fall
Fall 0,0
0,0 250 0,1
250 0,1
0,2
0,2 Top Top
0,3 0,3
Layer 200 1,0 Layer
200 1,0 Sensors
Sensors 1,1
1,1
1,2 1,2
150 1,3 150 1,3

Distance
Distance

2,0 2,0
2,1 2,1
100 100 2,2
2,2 Bottom Bottom
2,3 Layer 2,3 Layer
50 3,0 Sensors 3,0 Sensors
3,1 50 3,1
3,2 3,2
0 3,3 3,3
0
0 100 200 300 400 500 600 700 800 0 200 400 600 800 1000
Time (ms) Time (ms)

(a) (b)
Fall
0,0 Sit on Chair Fall 0,0
0,1 250 0,1
250 0,2 Top 0,2 Top
0,3 Layer 0,3 Layer
1,0 Sensors 200 1,0 Sensors
200 1,1 1,1
1,2 1,2
1,3 150 1,3
Distance
Distance

150 2,0 2,0

2,1 2,1
2,2 100 2,2
100 2,3 Bottom Bottom
2,3
3,0 Layer Layer
3,0
3,1 Sensors 50 3,1 Sensors
50
3,2 3,2
3,3 3,3
0 0
0 200 400 600 800 1000 0 100 200 300 400 500 600 700
Time (ms) Time (ms)

(c) (d)

0,0
0,1
250 0,2 Top
0,3 Layer
1,0 Sensors
200 1,1
1,2
150 1,3
Distance

2,0
2,1
100 2,2
2,3 Bottom
50 3,0 Layer
3,1 Sensors
3,2
0 3,3

0 1000 2000 3000 4000

Time (ms)

(e)
Figure 12: Distances in fall cases: (a) forward fall, (b) backward fall, (c) sideway fall, (d) fall from a chair, and (e) walking.

during 1,000 ms are shown in Figure 12. Tese graphs show on the opposite sidewall. All measured distances during
patterns of distance changes for each case. While ultrasonic 1,000 ms are shown in the graphs. In these measured distance
signal transmitters are located on a wall scan in the zigzag data, diferences among a forward fall, backward fall, sideway
direction to transmit ultrasonic signals by two nodes each time, fall, fall from a chair, and walking were observed. Tese data
distances are always measured by 16 sensors (0,0–3,3) located were input into the LSTM for training and classifcation.
Journal of Healthcare Engineering 15

Table 9: Error analysis based on one node.

Ratio of training/testing 90:10 80:20 70:30 60:40 50:50

Type Correct Error Correct Error Correct Error Correct Error Correct Error
Forward fall 10 19 1 24 6 29 11 35 15
Backward fall 10 18 2 25 5 33 7 38 12
Left and right sideway fall 10 17 3 26 4 35 5 33 17
Fall from a chair 8 2 17 3 24 6 32 8 34 16
Walk left to right 3 5 7 1 7 3 9 4
Walk right to left 2 4 1 7 7 3 9 3
Walk in a circle 3 4 1 7 1 8 2 10 3
Walk diagonal left to right 2 5 7 1 8 2 8 4
Walk diagonal right to left 3 3 2 7 1 8 2 7 6
Walk to sensor 2 5 7 7 3 9 3
Walk away from sensor 3 5 8 9 1 10 3
Walk to chair 2 5 7 8 2 9 3
Sum 58 2 107 13 156 25 191 49 211 89

Table 10: Error analysis based on two nodes.

Ratio of training/testing 90:10 80:20 70:30 60:40 50:50

Type Correct Error Correct Error Correct Error Correct Error Correct Error
Forward fall 10 20 29 1 38 2 47 3
Backward fall 10 20 28 2 37 3 46 4
Left and right sideway fall 10 19 1 29 1 38 2 46 4
Fall from a chair 9 1 18 2 28 2 37 3 45 5
Walk left to right 3 5 7 1 8 2 10 3
Walk right to left 2 5 6 1 8 2 9 3
Walk in a circle 3 5 8 9 1 11 2
Walk diagonal left to right 2 5 6 1 8 2 9 3
Walk diagonal right to left 3 5 7 1 9 1 10 3
Walk to sensor 2 5 7 10 10 2
Walk away from sensor 3 5 7 1 8 2 9 4
Walk to chair 2 4 1 6 1 8 2 8 4
Sum 59 1 116 4 168 12 218 22 260 40

In the experiments, the measured distances in con- adequate. Users must consider the coverage range of the
tinuing frames that showed behaviors, such as falling, ultrasonic signals with respect to the room scale as one of the
walking, and sitting done by the 25 participants (as shown in design conditions. Te results shown in Tables 5 and 6
Table 4), were used to train and test with various ratios. confrm the efective range of sensor coverage and were
Errors occurred based on the use of one node or two nodes, used to evaluate the performance of the proposed method.
as shown in Table 9, respectively (Table 10). Tese errors are Acceptable accuracy was achieved. Te accuracy increased
analyzed in the discussion. according to the increasing ratio of training samples, and an
approximately 98% accuracy was achieved using the 90 : 10
6. Discussion training and testing ratio. Although elderly individuals
(individuals over 40 years of age in this experiment) com-
To build a smart room that can be used to detect the falls of prehensively caused both positive and negative faults, they
elderly people without an intrusion of privacy, this study contributed a large number of true positives to the results
proposes installing an array of ultrasonic sensors on a wall, compared with the other group. Terefore, the proposed
activating the sensors to sense distance information, and method was considered to be applicable to elderly indi-
classifying the falls of elderly people based on distance viduals. Te proposed method was compared with con-
change patterns. Te performance of the proposed method ventional methods, and it was obvious that higher precision
was evaluated, and the accuracy was more than 90% in the and recall were obtained using the proposed method, as
cases of training more than 50% of the 2 node-based sample shown in Table 8. Tese high accuracy results were ana-
data, as shown in Table 7. Based on the results of the 1 node- lytically caused by the distance change patterns of forward,
based data shown in Table 6, the accuracy using this method backward, and sideway falls, falls from a chair, and especially
was worse than that of the 2 node-based data because sensing walking, as shown by the examples in Figure 12. As observed
based on one node was insufcient to cover the whole room. change patterns among consecutive frames in the time domain,
If the room was much smaller so that the range of the ul- classifers for video were confrmed to be appropriate tools for
trasonic signal covered the room, one node would be this kind of fall detection and classifcation problem.
16 Journal of Healthcare Engineering

Table 11: Error analysis based on two nodes.

Train/test FP FN
Correct Beside to fall Sum Correct — —
1 90/10 Error Side fall left and right Error — —
1 1
Correct Beside to fall Sum Correct — —
Error Side fall left and right Error — —
2 2
Correct Side fall left and right
2 80/20 Error Beside to fall
1 1
Correct Walk to sit
Error Walk to sensor
1 1
Correct Side fall Correct — —
Sum
Error Forward fall Blackward fall Error — —
1 1 2
Correct Blackward fall
Error Forward fall Side fall left and right
1 1 2
Correct Forward fall
Error Backward fall
1 1
Correct Side fall left and right
Error Beside to fall
1 1
Correct Walk left to right
Error Walk right to left
1 1
3 70/30
Correct Walk right to left
Error Walk left to right
1 1
Correct Walk diagonal left to right
Error Walk diagonal right to left
1 1
Correct Walk diagonal right to left
Error Walk diagonal left to right
1 1
Correct Walkout sensor
Error Walk to sensor
1 1
Correct Walk to sit
Error Walk diagonal left to right
1 1
Correct Side fall left and right Sum Correct Walk to sit Sum
Error Forward fall Blackward fall Error Beside to fall
1 1 2 2 2
Correct Blackward fall
Error Forward fall Side fall left and right
2 1 3
Correct Forward fall
Error Backward fall Side fall left and right
1 1 2
4 60/40
Correct Beside to fall
Error Forward fall Side fall left and right
1 2 3
Correct Walk left to right
Error Walk right to left
2 2
Correct Walk right to left
Error Walk left to right
2 2
Journal of Healthcare Engineering 17

Table 11: Continued.

Train/test FP FN
Correct Walk diagonal left to right
Error Walk diagonal right to left
1 1
Correct Walk diagonal right to left
Error Walk diagonal left to right
1 1
Correct Walkout sensor
Error Walk to sensor
2 2
Correct Walk to sit
Error Walk diagonal left to right
2 2
Correct Side fall left and right Sum Correct Walk to sit Sum
Error Forward fall Beside to fall Error Beside to fall
2 2 4 3 3
Correct Blackward fall Correct Beside to fall
Error Forward fall Side fall left and right Error Walk to sit
2 2 4 2 2
Correct Forward fall
Error Backward fall Side fall left and right
1 1 2
Correct Beside to fall
Error Forward fall Side fall left and right
1 2 3
Correct Walk left to right
Error Walk right to left Walk out sensor
2 1 3
Correct Walk right to left
Error Walk left to right
3 3
5 50/50
Correct Walk circle
Error Forward fall Backward fall
1 1 2
Correct Walk diagonal left to right
Error Walk diagonal right to left
3 3
Correct Walk diagonal right to left
Error Walk diagonal left to right
3 3
Correct Walkout sensor
Error Walk to sensor
2 2
Correct Walk to sensor
Error Walkout sensor Walk right to left
3 1 4
Correct Walk to sit
Error Walk diagonal left to right
2 2

Analytically, errors decreased when the number of training number of positive errors in many error patterns, such as
samples was increased compared with testing samples, and misclassifying sitting as falling and misclassifying the type of
accuracy was considered reliable in the 90 :10 training and fall. Tese errors may cause a caregiver to be alerted to
testing ratio, as shown in Tables 9 and 10. Terefore, the provide help to an elderly person who has fallen. Although
proposed method was proven to be efective for fall these errors were considered a waste of time and energy for
classifcation. the caregivers, they were counted as positive errors and were
In addition, the fault-positive (FP) and fault-negative considered a safety measure. However, the FN column in
(FN) errors shown in the middle column of Tables 6 and 7 Table 11 indicates some cases of falling from a chair that
were analyzed, and causes of these errors were found, as were misclassifed as sitting on a chair. After analyzing the
shown in Table 11. Te FP column in Table 10 indicates the photographs and signals of this case, as shown by the
18 Journal of Healthcare Engineering

(a)

(b)
Figure 13: Photographs representing the system misclassifcation of falling as sitting: (a) sitting on a chair and (b) falling from a chair.

250 250

200 200

150
Distance (cm)

Distance (cm)

150

100
100

50
50
0
0
0 200 400 600 800 1000 -200 0 200 400 600 800 1000 1200 1400 1600
Time (msec) Time (msec)
2,2 2,2
(a)

250 250

200 200
Distance (cm)
Distance (cm)

150 150

100 100

50 50

0 0
-200 0 200 400 600 800 1000 1200 1400 1600 0 200 400 600 800 1000
Time (msec) Time (msec)
2,2 3,1
(b)
Figure 14: Similar distance change patterns of sitting and falling.

examples in Figure 13, respectively, and Figure 14, the video limitation of the proposed method. Additional features, such
shots and signal patterns that represented falling and sitting as distance changes measured from the roof, should be
looked similar and were hard to diferentiate, even when considered as future work to solve this limitation. An ad-
judged by human eyes. Tis result was considered a ditional limitation was that an ultrasonic signal was
Journal of Healthcare Engineering 19

transmitted from a node as a triangular shape, as observed [3] “World Health Organization,” Global report on falls preven-
from the top view. Terefore, small areas between neigh- tion in older age, https://www.who.int/ageing/publications/
boring nodes were regarded as blind areas in which falls were Falls%20prevention7March.pdf, 2020.
impossible to detect. System designers must carefully design [4] L. Little and P. Briggs, “Pervasive healthcare: the elderly
the number of ultrasonic nodes in the array based on the perspective,” 2020, https://dl.acm.org/doi/abs/10.1145/%
201579114.1579185?__cf_chl_
blind zone to be smaller than the minimum human width, as
tk=uOx0n4OElbTG1m5dsAyUYYz3osl_
recommended above. Tis may guarantee protection against ZBNu0BPtsaVuxzw-1653921275-0-gaNycGzNC70.
not detecting a human fall. [5] E. Edfors and A. Westergren, “Home-living elderly people’s
views on food and meals,” Journal of aging research, vol. 2012,
7. Conclusions Article ID 761291, , 2012.
[6] C. Nadee and K. Chamnongthai, “Ultrasonic Array Sensors
To prevent serious risks to elderly individuals after falling in for Monitoring of Human Fall Detection,” in Proceedings of
a room, it is necessary to simultaneously monitor elderly the 2015 12th International Con-Ference on Electrical Engi-
behaviors without intruding on their privacy, detect falls, neering/Electronics, Computer, Telecommunications And In-
and immediately inform caregivers, when a fall occurs so formation Technology (ECTI-CON), pp. 1–4, Hua Hin,
they can provide urgent assistance. A design method for an Tailand, June 2015.
ultrasonic sensor-based system is proposed in this study for [7] C. Nadee and K. Chamnongthai, “Multi Sensor System for
Automatic Fall Detection,” in Proceedings of the 2015 Asia-
elderly fall monitoring in a smart room. In this design,
Pacifc Signal And Information Processing Association Annual
ultrasonic sensors are installed as a sensor array on a wall Summit and Conference, pp. 930–933, APSIPA), Hong Kong,
under the condition that the ultrasonic signal covers the area China, Dec. 2015.
of the whole room with a limited blind zone. Te blind zone [8] C. Nadee and K. Chamnongthai, “Octagonal Formation of
is determined in advance to be smaller than the width of the Ultrasonic Array Sensors for Fall Detection,” in Proceedings of
human, and the determined blind zone and average human the 2016 Asia-Pacifc Signal And Information Processing As-
fall duration are used to calculate the distance between sociation Annual Summit and Conference (APSIPA), pp. 1–4,
neighboring ultrasonic nodes and the total number of ul- Jeju, Korea (South), Dec2016.
trasonic nodes on a wall. Ten, activated ultrasonic nodes [9] R. Igual, C. Medrano, and I. Plaza, “Challenges, issues and
are transmitted in a one-by-one manner without interfer- trends in fall detection systems,” BioMedical Engineering
ence in a zigzag scanning line, and the ultrasonic signals, Online, vol. 12, no. 66, p. 66, 2013.
which are time-independent, are classifed as a fall or a [10] L. Ren and Y. Peng, “Research of fall detection and fall
prevention technologies: a systematic review,” IEEE Access,
nonfall by a time-independent-based classifer, such as
vol. 7, pp. 77702–77722, 2019.
LSTM. Te performance of the proposed method is con- [11] S. D. Hernandez, Y. DeLaHoz, and M. Labrador, “Dynamic
frmed to be efective. Background Subtraction for Fall Detection System Using a 2D
Camera,” in Proceedings of the IEEE Latin-America Confer-
Data Availability ence on Communications (LATINCOM), pp. 1–6, Cartagena,
Colombia, Nov2014.
Data are available at https://livermutlac-my.sharepoint.com/: [12] C. Lin, S. Wang, J. Hong, L. Kang, and C. Huang, “Vision-
f:/g/personal/eleccmk_rmutl_ac_th/EsYB9mvw2ZpFiXJ9zer Based Fall Detection through Shape Features,” in Proceedings
vDcIBSBdMa3enOdK17fhf0M2-DA?e�Znch33. of the 2016 IEEE Second International Conference on Multi-
media Big Data (BigMM), pp. 237–240, Taichung, Taiwan,
Oct2016.
Conflicts of Interest [13] G. M. Basavaraj and A. Kusagur, “Vision based surveillance
Te authors declare that they have no conficts of interest. system for detection of human fall,” in Procedings of the 2nd
IEEE International Conference on Recent Trends in Electronics
Information & Communication Technology (RTEICT),
Acknowledgments pp. 1516–1520, Bangalore, India, May2017.
[14] X. ShanShan and C. Xi, “Fall Detection Method Based on
Te authors are thankful for PhD scholarship supported by Semi-contour Distances,” in Proceedings of the 2018 14th IEEE
the Rajamangala University of Technology Lanna (RMUTL), International Conference on Signal Processing (ICSP),
Chiangmai, Tailand (195/2557). pp. 785–788, Beijing, China, Aug. 2018.
[15] Y. Chen, X. Kong, L. Chen, L. Meng, and H. Tomiyama, “A
References Dynamic Height Analysis on Vision Based Fall Detection
System,” in Proceedings of the International Conference on
[1] S. Afairs and P. Division, “World Population Prospects 2019: Advanced Mechatronic Systems, pp. 148–151, ICAMechS),
Highlights,” United Nations Department of Economic, https:// Kusatsu, Japan, Aug. 2019.
population.un.org/wpp/Publications/Files/WPP2019_ [16] X. Cai, S. Li, X. Liu, and G. Han, “A Novel Method Based on
Highlights.pdf, 2019. Optical Flow Combining with Wide Residual Network for Fall
[2] S. Afairs and P. Division, “World Population Ageing 2019: Detection,” in Proceedings of the IEEE 19th International
Highlights,” United Nations Department of Economic, https:// Conference on Communication Technology (ICCT), pp. 715–
www.un.org/en/development/desa/population/publications/ 718, Oct2019, https://researchr.org/publication/icct-2019.
%20pdf/ageing/WorldPopulationAgeing2019-Highlights.pdf, [17] A. Núñez-Marcos, G. Azkune, and I. Arganda-Carreras,
2019. “Vision-based fall detection with convolutional neural
20 Journal of Healthcare Engineering

networks,” Wireless Communications and Mobile Computing, [32] A. Jahanjoo, M. N. Tahan, and M. J. Rashti, “Accurate Fall
vol. 16, p. 2017, 2017. Detection Using 3-axis Accelerometer Sensor and MLF Al-
[18] N. Lu, Y. Wu, L. Feng, and J. Song, “Deep learning for fall gorithm,” in Proceedings of the 3rd International Conference
detection: three-dimensional CNN combined with LSTM on On Pattern Recognition And Image Analysis (IPRIA),
video kinematic data,” IEEE Journal of Biomedical and Health pp. 90–95, Shahrekord, Iran, April2017.
Informatics, vol. 23, no. 1, pp. 314–323, 2019. [33] A. Mao, X. Ma, Y. He, and J. Luo, “Highly portable, sensor-
[19] K. de Miguel, A. Brunete, M. Hernando, and E. Gambao, based system for human fall monitoring,” Sensors, vol. 17,
“Home camera-based fall detection system for the elderly,” no. 9, p. 2096, 2017.
Sensors, vol. 17, no. 12, p. 2864, 2017. [34] C. Wang, W. Lu, S. J. Redmond, M. C. Stevens, S. R. Lord, and
[20] A. Lotf, S. Albawendi, H. Powell, K. Appiah, and N. H. Lovell, “A low power fall detector balancing sensitivity
C. Langensiepen, “Supporting independent living for older and false alarm rate,” IEEE Journal of Biomedical and Health
adults; employing a visual based fall detection through ana- Informatics, vol. 22, no. 6, pp. 1929–1937, 2018.
lysing the motion and shape of the human body,” IEEE Access, [35] A. Purushothaman, K. V. Vineetha, and D. Kurup, “Fall
vol. 6, pp. 70272–70282, 2018. Detection System Using Artifcial Neural Network,” in Pro-
[21] R. Freitas, M. Terroso, M. Marques, J. Gabriel, A. Torres ceedings of the 2018 Second International Conference on In-
Marques, and R. Simoes, “Wearable sensor networks sup- ventive Communication and Computational Technologies
ported by mobile devices for fall detection,” Sensors, , (ICICCT), pp. 1146–1149, Coimbatore, India, April2018.
pp. 2246–2249, IEEE, 2014. [36] S. B Khojasteh, “Improving fall detection using an on-wrist
[22] P. Pierleoni, A. Belli, L. Palma, M. Pellegrini, L. Pernini, and wearable accelerometer,” Sensors, vol. 5, 2018.
S. Valenti, “A high reliability wearable device for elderly fall [37] T. de Quadros, A. E. Lazzaretti, and F. K. Schneider, “A
detection,” IEEE Sensors Journal, vol. 15, no. 8, pp. 4544–4553, movement decomposition and machine learning-based fall
2015. detection system using wrist wearable device,” IEEE Sensors
[23] N. Otanasap, “Pre-Impact Fall Detection Based on Wearable Journal, vol. 18, no. 12, pp. 5082–5089, 2018.
Device Using Dynamic Treshold Model,” in Proceedings of [38] M. Saleh and R. L. B. Jeannes, “Elderly fall detection using
the 2016 17th International Conference On Parallel And wearable sensors: a low cost highly accurate algorithm,” IEEE
Distributed Computing, Applications And Technologies Sensors Journal, vol. 19, no. 8, pp. 3156–3164, 2019.
(PDCAT), pp. 362–365, Guangzhou, China, Dec. 2016. [39] H.-W. Tzeng, M.-Y. Chen, and J. Chen, “Design of Fall
[24] A. Kurniawan, A. R. Hermawan, and I. K. E. Purnama, “A Detection System with Floor Pressure and Infrared image,” in
Wearable Device for Fall Detection Elderly People Using Tri
Proceedings of the International Conference on System Science
Dimensional Accelerometer,” in Proceedings of the 2016 In-
and Engineering, pp. 131–135, Taipei, July2010.
ternational Seminar On Intelligent Technology And its Ap-
[40] Q. Guan, C. Li, X. Guo, and B. Shen, “Infrared Signal Based
plications (ISITIA), pp. 671–674, Lombok, Indonesia,
Elderly Fall Detection for in-Home Monitoring,” in Pro-
July2016.
ceedings of the 9th International Conference on Intelligent
[25] C. Chu, H. Chiang, C. Chang, H. Li, and T. Chang, “A Ex-
Human-Machine Systems and Cybernetics (IHMSC),
ponential Smoothing gray Prediction Fall Detection Signal
pp. 373–376, Oct2017.
Analysis in Wearable device,” in Proceedings of the 6th In-
[41] Y. Ogawa and K. Naito, “Fall Detection Scheme Based on
ternational Symposium on Next Generation Electronics (ISNE),
Temperature Distribution with IR Array Sensor,” in Pro-
pp. 1–4, Keelung, Taiwan, May2017.
[26] K. M. Shahiduzzaman, X. Hei, C. Guo, and W. Cheng, ceedings of the 2020 IEEE International Conference on Con-
“Enhancing fall detection for elderly with smart helmet in a sumer Electronics (ICCE), pp. 1–5, Las Vegas, NV, USA,
cloud-network-edge architecture,” in Proceedings of the IEEE Jan2020.
International Conference on Consumer Electronics - Taiwan [42] D. Asbjørn and T. Jim, “Recognizing bedside events using
(ICCE-TW), pp. 1-2, Yilan, Taiwan, May2019. thermal and ultrasonic readings,” Sensors, vol. 17, no. 6, 2017.
[27] M. I. Nari, S. S. Suprapto, I. H. Kusumah, and W. Adiprawita, [43] Z. Chen and Y. Wang, “Infrared-ultrasonic sensor fusion for
“A Simple Design of Wearable Device for Fall Detection with support vector machine-based fall detection,” Journal of In-
Accelerometer and gyroscope,” in Proceedings of the Inter- telligent Material Systems and Structures, vol. 29, no. 9,
national Symposium on Electronics and Smart Devices pp. 2027–2039, 2018.
(ISESD), pp. 88–91, Bandung, Indonesia, Nov2016. [44] M. G. Amin, Y. D. Zhang, F. Ahmad, and K. C. D. Ho, “Radar
[28] Y.-H. Nho, J. G. Lim, and D.-S. Kwon, “Cluster-analysis-based signal processing for elderly fall detection: the future for in-
user-adaptive fall detection using fusion of heart rate sensor home monitoring,” IEEE Signal Processing Magazine, vol. 33,
and accelerometer in a wearable device,” IEEE Access, vol. 8, no. 2, pp. 71–80, 2016.
pp. 40389–40401, 2020. [45] L. Liu, M. Popescu, M. Skubic, M. Rantz, T. Yardibi, and
[29] H. Chen, “An Accurate Crowdsourcing-Based Adaptive Fall P. Cuddihy, “Automatic Fall Detection Based on Doppler
Detection Approach Using Smart Devices,” in Proceedings of Radar Motion signature,” in Proceedings of the 5th Interna-
the 2016 IEEE International Conference on Healthcare In- tional Conference on Pervasive Computing Technologies for
formatics (ICHI), p. 304, Chicago, IL, USA, Oct2016. Healthcare, pp. 222–225, Dublin, Republic of Ireland,
[30] Y. Tang, “iPrevent: A Novel Wearable Radio Frequency Range May2011.
Detector for Fall Prevention,” in Proceedings of the 2016 IEEE [46] S. Tomii and T. Ohtsuki, “Learning Based Falling Detection
International Symposium on Radio-Frequency Integration Using Multiple Doppler sensors,” in Proceedings of the IEEE
Technology (RFIT), pp. 1–3. 14th International Conference on E-Health Networking, Ap-
[31] H. Djelouat, “CS-based Fall Detection for Connected Health plications and Services (Healthcom), pp. 196–201, Beijing,
Applications,” in Proceedings of the 2017 Fourth International China, Oct2012.
Conference on Advances in Biomedical Engineering (ICABME) [47] M. Wu, X. Dai, Y. D. Zhang, B. Davidson, M. G. Amin, and
ICABME, pp. 1–4, Beirut, Lebanon, Oct2017. J. Zhang, “Fall detection based on sequential modeling of
Journal of Healthcare Engineering 21

radar signal time-frequency features,” IEEE International Biology Society, pp. 5867–5870, San Diego, CA, USA,
Conference on Healthcare Informatics, pp. 169–174, 2013. Sept2012.
[48] F. Wang, M. Skubic, M. Rantz, and P. E. Cuddihy, “Quan- [62] M. Chefena, “Fall detection using smartphone audio fea-
titative gait measurement with pulse-Doppler radar for pas- tures,” IEEE Journal of Biomedical and Health Informatics,
sive in-home gait assessment,” IEEE Transactions on vol. 20, no. 4, pp. 1073–1080, 2016.
Biomedical Engineering, vol. 61, no. 9, pp. 2434–2443, 2014. [63] W. Zhi, L. Wang, and X. Hu, “Recent Advances in the Efects
[49] A. Gadde, M. G. Amin, Y. D. Zhang, and F. Ahmad, “Fall of Microwave Radiation on Brains,” Military Medical Re-
detection and classifcations based on time-scale radar signal search, vol. 4, no. (1), pp. 1–14, 2017.
characteristics,” Radar Sensor Technology, vol. XVIII, pp. 1–9, [64] A. Miller, M. Sears, L. Morgan et al., “Risks to health and well-
Article ID 907712, 9077. being from radio-frequency radiation emitted by cell phones
[50] Q. Wu, Y. D. Zhang, W. Tao, and M. G. Amin, “Radar based and other wireless devices,” Frontiers in Public Health,
fall detection based on Doppler time-frequency signatures for pp. 1–10, 2019.
assisted living,” IET Radar, Sonar & Navigation, vol. 9, no. 2, [65] B. Smagowska and M. Pawlaczyk-Łuszczyńska, “Efects of
pp. 164–172, 2015. ultrasonic noise on the human body-A bibliographic review,”
[51] B. Y. Su, K. C. Ho, M. J. Rantz, and M. Skubic, “Doppler radar International Journal of Occupational Safety and Ergonomics,
fall activity detection using the wavelet transform,” IEEE vol. 19, no. 2, pp. 195–202, 2013.
Transactions on Biomedical Engineering, vol. 62, no. 3, [66] H. Toshio and Y. Nishida, Ultrasonic Sensors for the Elderly
pp. 865–875, 2015. and Caregivers in a Nursing Home, ICEIS, New Delhi, Delh,
2005.
[52] H. Wang, D. Zhang, Y. Wang, J. Ma, Y. Wang, and S. Li, “Rt-
[67] Y. Huang and K. Newman, “Improve quality of care with
Fall: RT-fall: a real-time and contactless fall detection system
remote activity and fall detection using ultrasonic sensors,”
with commodity WiFi devices,” IEEE Transactions on Mobile
Annual International Conference of the IEEE Engineering in
Computing, vol. 16, no. 2, pp. 511–526, 2017.
Medicine and Biology Society, pp. 5854–5857, 2012.
[53] U. Mahmood Khan, Z. Kabir, and S. A. Hassan, “Wireless
[68] Y. Chang and T. K. Shih, “Human Fall Detection Based on Event
Health Monitoring Using Passive WiFi Sensing,” in Pro-
Pattern Matching with Ultrasonic Array sensors,” in Proceedings of the
ceedings of the 13th International Wireless Communications 10th International Conference on Ubi-media Computing and Work-
and Mobile Computing Conference, pp. 1771–1776, Valencia, shops, pp. 1–4, Pattaya, Tailand, Aug2017.
Spain, June2017. [69] A. Ghosh, A. Chakraborty, D. Chakraborty, M. Saha, and
[54] Y. Gu, Y. Zhang, M. Huang, and F. Ren, “Your WiFi Knows S. Saha, “UltraSense: a non-intrusive approach for human
You Fall: A Channel Data-Driven Device-free Fall Sensing activity identifcation using heterogeneous ultrasonic sensor
System,” in Proceedings of the 5th IEEE International Con- grid for smart home environment,” Journal of Ambient In-
ference on Cloud Computing and Intelligence Systems, telligence and Humanized Computing, 2019.
pp. 943–947, Nanjing, China,, Nov2018. [70] World Population Ageing 2019, “United Nations, New York,
[55] R. Ramezani, Y. Xiao, and A. Naeim, “Sensing-Fi: Wi-Fi CSI Department of economic and social afairs, Population di-
and Accelerometer Fusion System for Fall Detection,” in vision,” 2020, https://www.un.org/en/development/desa/
Proceedings of the IEEE EMBS International Conference on population/publications/pdf/ageing/
Biomedical & Health Informatics (BHI), pp. 402–405, Las WorldPopulationAgeing2019-Highlights.pdf.
Vegas, NV, USA, March2018. [71] J. Ausubel, Older People Are More Likely to Live Alone in the
[56] H. Cheng, J. Zhang, Y. Gao, and X. Hei, “Deep Learning Wi-Fi U.S. Tan Elsewhere in the World, Pew Research Center,
Channel State Information for Fall Detection,” in Proceedings Washington, D.C., U.S., 2020.
of the IEEE International Conference on Consumer Electronics [72] B. Lin, J. Su, H. Chen, and C. Y. Jan, “Fall Detection System
- Taiwan, pp. 1-2, YILAN, Taiwan, June2019. Based on Human Body Silhouette,” in Proceedings of the 2013
[57] Y. Hu, F. Zhang, C. Wu, B. Wang, and K. J. Ray Liu, “A WiFi- Ninth International Conference On Intelligent Information
Based Passive Fall Detection System,” in Proceedings of the Hiding And Multimedia Signal Processing, pp. 49–52, Beijing,
ICASSP 2020 - 2020 IEEE International Conference On China, June2013.
Acoustics, Speech And Signal Processing, pp. 1723–1727, [73] P. A. Hewitt and D. Dobberfuhl, “Te science and art of
ICASSP), Barcelona, Spain, May2020. proportionality,” Science Scope, vol. 27, no. 4, pp. 30-31, 2004.
[58] Y. Zigel, D. Litvak, and I. Gannot, “A method for automatic [74] S. J Ling, J. Sanny, and B. Moebs, “Energy and Power of a
fall detection of elderly people using foor vibrations and Wave,” 2020, https://bit.ly/33SKgqX.
sound-proof of concept on human mimicking doll falls,” IEEE [75] “National Institute of Standards and Technology,” Interna-
Transactions on Biomedical Engineering, vol. 56, no. 12, tional aspects of the SI, http://physics.nist.gov/cuu/Units/
pp. 2858–2867, 2009. international.html, 2022.
[59] Y. Li, Z. Zeng, M. Popescu, and K. C. Ho, “Acoustic fall [76] H. M. Elmehdi, J. H. Page, and M. G. Scanlon, “Ultrasonic
detection using a circular microphone array,” in Proceedings investigation of the efect of mixing under reduced pressure
of the Annual International Conference of the IEEE Engi- on the mechanical properties of bread dough,” Cereal
neering in Medicine and Biology, pp. 2242–2245, Buenos Chemistry Journal, vol. 81, no. 4, pp. 504–510, 2004.
Aires, Argentina, Sept2010. [77] S. Karsoliya, “Approximating number of hidden layer neu-
[60] Y. Yun Li, K. C. Ho, and M. Popescu, “A microphone array rons in multiple hidden layer BPNN architecture,” Interna-
tional Journal of Engineering Trends and Technology, vol. 3,
system for automatic fall detection,” IEEE Transactions on
pp. 714–717, 2012.
Biomedical Engineering, vol. 59, no. 5, pp. 1291–1301, 2012.
[78] J. Wu, X. Y. Chen, H. Zhang, L. D. Xiong, H. Lei, and
[61] Y. Li, M. Popescu, and K. C. Ho, “Improving automatic
S. H. Deng, “Hyperparameter optimization for machine
sound-based fall detection using iVAT clustering and GA-
learning models based on Bayesian optimization,” Journal of
based feature selection,” in Proceedings of the Annual Inter-
Electronic Science and Technology, vol. 1, pp. 26–40, 2019.
national Conference of the IEEE Engineering in Medicine and