15154 IEEE SENSORS JOURNAL, VOL. 22, NO.

15, 1 AUGUST 2022

A Three-Stage Low-Complexity Human Fall

Detection Method Using IR-UWB Radar
Mengxia Chen , Zhaocheng Yang , Member, IEEE, Jialei Lai , Ping Chu , and Jinghong Lin

Abstract —This paper proposes a novel three-stage low-

complexity human fall detection method using an impulse
radio ultra-wideband (IR-UWB) radar. The core idea lies in
the three cascaded stages, namely large-motion detection,
rough fall detection and enhanced fall detection. For the large-
motion detection, we assume the fall is a very sparse event in
daily life and achieve this by a checking of the high Doppler
frequency energy. For the rough fall detection, we do not
intuitively determine the fall events, but propose six time-
frequency features and two position features, and use the
support vector data description (SVDD) detector to divide
the large-motions into non-fall and fall-like events. For the
enhanced fall detection, we add a new feature and use a Maha-
lanobis distance classifier to finally determine whether a fall
happened. The reasons that two classifiers cascaded instead
of one classifier are that (1) we can reduce the difficulty of
identifying falls directly from daily events by using a large
number of non-fall samples to train the SVDD model for anomaly detection, and allowing a certain false alarm rate; and
(2) we can achieve a higher fall detection accuracy in a much smaller searching space by identifying falls only from the
fall-like events. Additionally, a real-time edge fall detection system with a commonly used micro control unit is developed.
Experiment results show that the proposed method exhibits a low computational complexity, and a relative robustness
and high fall detection accuracy under a low false alarm rate.
Index Terms — Fall detection, IR-UWB radar, low-complexity, robustness.

I. I NTRODUCTION for nursing staff to always pay attention to whether the elderly
fall or not. Therefore, intelligent fall detection has become an
A CCORDING to recent demographic statistics, in 2018,
the global population over the age of 64 surpassed the
number of children under the age of 5 for the first time [1].
active area of both the research and industry. The intelligent
fall detection system can be used to: (1) monitor the elderly
As the global population over the age of 64 will continue to living alone and inform their families or relevant personnel
grow, it is clear that we are moving towards an aging world. at the first time after the elderly falling down; (2) reduce the
In daily life, falls are the main cause of injury to the elderly. work intensity of nurses and further ensure the safety of the
If help and treatment are not provided in time after a fall, elderly; and (3) avoid serious consequences after accidental
it may cause various other diseases and even indirectly lead falls, especially in the bathroom.
to the death of the elderly [2], [3]. At present, a large number Nowadays, fall detection technologies are developing
of the elderly live alone. Even in nursing homes, it is difficult rapidly, and a variety of fall detection devices have appeared
on the market. These devices can be divided into two cat-
egories: wearable and non-wearable. Wearable sensors are
Manuscript received 3 May 2022; accepted 7 June 2022. Date of
publication 29 June 2022; date of current version 1 August 2022. mainly micro electro mechanical systems (MEMS), includ-
This work was supported in part by the Guangdong Basic and ing accelerometer, gyroscope, magnetometer and barome-
Applied Basic Research Foundation under Grant 2019A1515011517, ter [4]. Researchers exploit one or multiple MEMS sensors
in part by the Science and Technology Project of Shenzhen under
Grant JCYJ20190808142803565, and in part by the National Nat- in devices, such as belt, watch, wristband, and smartphone,
ural Science Foundation of China under Grant 61771317 and Grant etc., to obtain the posture angle and motion information for
62101207. The associate editor coordinating the review of this article and the fall event [4]–[9]. Lee et al. [9] embedded accelerometers
approving it for publication was Dr. Ying Zhang. (Corresponding author:
Zhaocheng Yang.) into a smartphone, and effectively distinguished falls in four
The authors are with the Guangdong Key Laboratory of Intelli- directions from 11 types of daily activities by analyzing the
gent Information Processing, College of Electronics and Information triaxial acceleration characteristics and using three thresh-
Engineering, Shenzhen University, Shenzhen 518060, China (e-mail:
[email protected]; [email protected]). olds. Norudin et al. [7] embedded a heart rate sensor and
Digital Object Identifier 10.1109/JSEN.2022.3184513 an acceleration sensor into a smartwatch, and proposed a

1558-1748 © 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on October 15,2024 at 16:07:32 UTC from IEEE Xplore. Restrictions apply.
 CHEN et al.: THREE-STAGE LOW-COMPLEXITY HUMAN FALL DETECTION METHOD USING IR-UWB RADAR 15155

novel generative adversarial network (GAN) based method TABLE I

for fall detection by heart rate features and accelerometer R ADAR PARAMETERS C ONFIGURED
features. Since MEMS sensors are small in size, low in power
consumption, and easy to integrate, they are currently a hot
spot in the field of fall detection. However, the sensor needs to
be carried by the human, and many application scenarios are
limited. Non-wearable devices mainly include vision sensors,
infrared sensors, ultrasonic array sensors, vibration sensors and
radar sensors. However, high computational cost, sensitivity
to the light and privacy concerns limit the application of
vision sensors [10]–[12]. Infrared sensors are easily affected
by strong light and temperature [13]. Ultrasonic arrays are
susceptible to environmental interference, leading to high error
rates [14]. Vibration sensors are not practical because they can
only be uesd to detect a fall on specific types of floor [15]. support vector data description (SVDD) detector to divide
Due to the advantages of strong penetrability, free of the light the large-motions into non-fall and fall-like events. Here, the
and good personal privacy protection, radar-based human fall fall-like events include fall, waving one’s hand down, squatting
detection has been extensively studied in recent years and down quickly and sitting down quickly, and non-fall events
becomes an alternative for the non-contact technology indoor include the other large-motions. In stage three, we add a new
monitoring [16]–[18]. feature and use a Mahalanobis distance classifier to finally
Currently, radar-based fall detection methods can be mainly determine whether a fall happened. It is worth noting that
divided into two strategies: deep neural networks (DNN) based the merits of the SVDD classifier in dealing with the binary
methods [19]–[26] and traditional machine learning (TML) classification problem with imbalanced data are demonstrated
based methods [27]–[35]. For the DNN-based methods, there through the comparison of other TML methods. Additionally,
are usually convolutional neural networks (CNN) [20]–[23], the detection accuracy, robustness and computational com-
long-short term memory (LSTM) networks [23], and cap- plexity of the proposed method are also analyzed. The main
sule networks etc. [24]. However, the computational com- contributions of our paper are:
plexity of this strategy is relatively high, and it is hard (1) A novel three-stage low-complexity fall detection strat-
to be implemented in an edge detection system with a egy using an IR-UWB radar is proposed.
low computational resource. For the TML-based methods, (2) A rough fall classifier cascaded with an enhanced fall
most researchers perform time-frequency analysis [27]–[33] classifier is developed instead of one fall classifier, which can
or wavelet transform analysis [34], [35] on the collected facilitate the fall detection under a limited fall sample support
signals to extract relevant features, and use methods, such as and computational resource.
Mahalanobis distance [27], hidden Markov model (HMM) [29] (3) A real-time edge fall detection system with a commonly
and support vector machine (SVM) [28], [30]–[33], to calssify used micro control unit is developed.
human motions. However, most researchers only consider a (4) A study and comparative analysis of the proposed
calssification between falls and a few daily activities, such method based on offline dataset and real-time system is
as picking things up, sitting down, and walking. Currently, conducted.
although a number of radar-based methods have been pro- The rest of this paper is organized as follows. Section II will
posed for fall detection, there are still three main challenges: describe the radar signal model. In section III, the proposed
(1) it is difficult to obtain a large dataset of people fall events; three-stage fall detection method is detailed. It consists of
(2) the motion characteristies obsevered by radar are sensitive human large-motion detection, rough fall detection, enhanced
to the attitude angle of motions, which makes it easy to fall detection and real-time edge fall detection system. Exper-
confuse a fall with other motions, such as fast sitting down iment results will be presented in section IV. Finally, the
and fast squatting down; and (3) the robustness of the existing conclusion will be drawn in section V.
fall detection methods is required to improve.
In this paper, a three-stage low-complexity human fall detec- II. R ADAR S IGNAL M ODEL
tion method using an impulse radio ultra-wideband (IR-UWB) An IR-UWB radar, named X4M03 radar [36], is considered
radar under limited fall sample support is presented. In stage in this paper for fall detection. The power consumption of the
one, to reduce the computational complexity, we assume the radar is only 118 mW with low radiation. The frames per
fall is a sparse event in daily life and extract a high-frequency second (FPS) is set as fr =300 to prevent speed ambiguity.
energy feature to determine whether a large-motion occurs. The detection range is set to 1-6m, and the sampling resolution
In the experiments, we mainly consider 9 types of large- is 0.0064m. Therefore, there are R=782 range bins in one
motions, including fall, waving, squatting down, stretching frame of the received echo. The detailed radar configuration
exercises, picking up things, sitting down, sitting up, drinking parameters are depicted in Table I.
and walking. In stage two, we propose to extract six time- The IR-UWB radar is composed of a single transmitter
frequency features and two position features, and use the and a single receiver, and emits an Gaussian impulse signal,

Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on October 15,2024 at 16:07:32 UTC from IEEE Xplore. Restrictions apply.
 15156 IEEE SENSORS JOURNAL, VOL. 22, NO. 15, 1 AUGUST 2022

Fig. 1. The received signal and the clutter removed signal of one frame.

defined by
t2
s(t) = VT X exp(− ) cos(2π f c t), (1)
2τ 2
where VT X denotes the pulse amplitude, t means the fast-time,
f c denotes the carrier frequency, and τ is described by

τ = 2π f B lg(e). (2)
Here, f B is bandwidth, and e is Euler’s number.
After sampling in X4M03, the received signal can be
defined by r (m, n), where m denotes the frame index and Fig. 2. The workflow of the proposed three-stage human fall detection
n denotes the range bin. The r (m, n) contains a variety method.
of background cultters. In order to remove the background
clutters, the running average method [37] is applied because
instantaneous high Doppler frequency spectrum components.
of the low computational complexity. It is described as
Generally, the fall detection isn’t an easy task and requires a
x(m, n) = r (m, n) − c(m, n), (3) large computational resource. Besides, fall is a sparse event in
c(m + 1, n) = αc(m, n) + (1 − α)r (m, n), (4) daily life. Herein, we do not need to compute the classification
procedure all the time, but only need to conduct that when a
where c(m, n) is the background clutter signal, x(m, n) large-motion happened. Specifically, a high Doppler frequency
denotes the signal after subtracting clutter and α is an update energy feature based method can be used to achieve this aim.
coefficient. We set α as 0.983, meaning that micro-Doppler Firstly, the distance units of z(m, n) are united because we
features arise rapidly while the clutter remains longer. only consider whether a large-motion happened, not where it
Fig. 1 shows the received signal and the clutter removed happened. Secondly, for real-time detection, a sliding window
signal of one frame. It can be seen that after clutter suppres- of 5/6 second overlap is applied to calculate fast Fourier
sion, the scattering amplitude of static target is greatly reduced, transform with fixed length of window (1 second). For the
but the human target signal still keeps a strong scattering kth sliding window, we compute the kth spectrum vector
amplitude. described by
After clutter suppression, in order to facilitate the extraction
+∞ 
 R
of target Doppler information, digital frequency mixing and − j 2π
fd
m
low-pass filtering are performed to the signal x(m, n) to X (k, f d ) = z k (m, n)e fr , (5)
m=−∞ n=1
remove the carrier signal and obtain the baseband signal
z(m, n). where z k (m, n) is the baseband signal z(m, n) at the kth
sliding window, and f d denotes the Doppler frequency. Finally,
III. P ROPOSED T HREE -S TAGE H UMAN we extract a high Doppler frequency energy feature E high from
FALL D ETECTION M ETHOD the spectrum vector and set a threshold γ1 to decide whether
This section will detail the proposed method by three a large-motion occurs, namely
stages: human large-motion detection, rough fall detection and H1
enhanced fall detection, as shown in Fig. 2. E high (k) >
γ1 , (6)
<
H0
A. Stage 1: Human Large-Motion Detection where H1 means that a large-motion occurs and H0 means
Here, we mainly consider 9 types of large-motions, includ- oppositely, and
ing fall, waving, squatting down, stretching exercises, picking 
up things, sitting down, sitting up, drinking and walking. E high (k) = |X (k, f d )|2 ,  = { f 1 < | fd | < f 2 } . (7)
From the view of radar, these events will produce large f =

Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on October 15,2024 at 16:07:32 UTC from IEEE Xplore. Restrictions apply.
 CHEN et al.: THREE-STAGE LOW-COMPLEXITY HUMAN FALL DETECTION METHOD USING IR-UWB RADAR 15157

Here, f 1 and f2 represent the low bound and high bound of

the high Doppler frequency, respectively. In the experiments,
f 1 and f 2 can be set to 50 Hz and 75 Hz, respectively.
Because the fall events will inevitably exhibit a significant
amount of energy in high Doppler frequency component,
a small threshold γ1 will not lead to missed detection of a
fall. If a large-motion occurs, we need to find the abnormal
time tbegin (which is the start time of a large-motion) to extract
the basedband signal for the later fall detection. Here, we set
a smaller threshold γ2 to find the tbegin in one second data Fig. 3. The trajectory acquisition process of a fall event.
E high (k1 ), where k1 is {k − 6, k − 5, . . . , k − 1}. The index k1
is the first value of E high (k1 ) that is greater than γ2 , and is
denoted by tbegin . Both γ1 and γ2 are experience values. Thirdly, for each frame of data, the ordered statistics con-
stant false alarm rate (OS-CFAR) detector is used to judge
B. Stage 2: Rough Fall Detection whether it is the target from the first range bin. If consec-
If a large-motion occurs, the subsequent t1 seconds data utive five range bins are detected as the target, we stop the
after tbegin of z(m, n) are used for fall detection. From a large detection and take the first of the five range bins as the target
number of experiments, 4 seconds is the most appropriate position. Thus, we denote the obtained trajectory as L(m 1 ).
value for t1 . This is beacuse that data during this time can Furthermore, we perform median filtering to L(m 1 ) and obtain
not only contain the information of the fall, but also ensure a smooth trajectory T (m 1 ), as shown in Fig. 3.
the real-time detection. The details are in section V. For futher processing, we first detail the differences of
For the t1 seconds data, the short-time Fourier transform time-frequency spectrums and trajectories of different motions.
(STFT) [38] is used to generate time-frequency spectrum From a large number of fall experiments, we find that the time-
frequency spectrums when the radar is placed on the side
+∞ 
 R fd
− j 2π l are much more sensitive to the people falling angle relative
Z (m, fd ) = z(m + l, n)h(l)e fr , (8)
to radar than those when the radar is placed on the top.
l=−∞ n=1
In particular, in the case of the radar top-mounted, when a
where h(l) is a Gaussian window function. The length of person falls directly under the radar, it is obvious that the
Gaussian window is set to 100, and 80-point overlap is applied. radar is no longer sensitive to the attitude angle of falls. Thus,
To obtain the target Doppler freqency components, we perform in the following, we consider the case of radar placed on
zero-forcing processing to Z (m, f d ) and remove the Doppler the top. However, although the the Doppler shift sensitive
components of the clutter [39]. to the attitude angle of motions can be largely reduced by
Additionally, we use the following steps to obtain the the installation of the radar sensor, this issue has not been
trajectory of the human body relative to the radar. Firstly, completely solved. This can be seen from the plots of Fig. 4.
we integrate every k1 range bins of x(m, n) into an interval The time-frequency spectrum of falling towarding 90 ◦ is
along the fast-time sequence. Thus, we divide the fast-time slightly different from others.
axis of x(m, n) into R1 (R1 = R/k1  and · indicates Fig. 5 shows time-frequency spectrums and trajectories
rounding down the value) intervals. We then fill each interval obtained by a person at a position with different large-
with the sum of the absolute values of the k1 related range motions. Fig. 6 show time-frequency spectrums and trajecto-
bins. The propcessed signals can be described by ries obtained by a person standing at a horizontal distance
1 ∗n 1
k of 1.2m from the radar and falling towards different direc-

x (m, n 1 ) = |x(m, n)|, (9) tions. Fig. 16 shows time-frequency spectrums and trajectories
n=k1 ∗(n 1 −1)+1 obtained by two people at different positions falling from the
where n 1 = {1, 2, . . . , R1 }. Here, k1 can be set to 10, radar to a certain direction. From the figures, we observed
ensuring that not only the integrated signal has an enough that: (1) the fall and most large-motions are separable in the
range resolution, but also the computational complexity can time-frequency spectrums because the high Doppler frequency
be reduced obviously. spectrum components produced by a fall are greater than
Secondly, we reduce the time resolution of the sig- most large-motions, such as stretching exercises, picking up
x (m, n 1 ) along the slow-time sequence similarly. The
nal  things, slow sitting down, sitting up, drinking and walking; and
propcessed signals can be given by (2) the time-frequency characteristics observed by radar are not
particularly sensitive to the posture angle of human body, but
2 ∗m 1
k
more sensitive to the position of motions, which results that

x (m 1 , n 1 ) = |
x (m, n 1 )|, (10) time-frequency spectrums of some fall events are very similar
m=k2 ∗(m 1 −1)+1 with those of some large-motions, such as waving one’s hand
where m 1 = {1, 2, . . . , t1 ∗ F P S/k2 }. In the experiments, k2 down, squatting down quickly and quick sitting down quickly.
can be set to 10, ensuring that not only the integrated signal However, these events are highly separable in trajectory.
has an enough time resolution, but also the computational Based on the above observations, it is not a good choice
complexity can be reduced obviously. to distinguish falls from all other large-motions by using

Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on October 15,2024 at 16:07:32 UTC from IEEE Xplore. Restrictions apply.
 15158 IEEE SENSORS JOURNAL, VOL. 22, NO. 15, 1 AUGUST 2022

Fig. 4. The time-frequency spectrums obtained by a person standing at a position and falling towards different directions (0◦ , 30◦ , 60◦ and 90◦ )
when the radar is placed on the side and top respectively.

a classifier with the same inputs. Therefore, we would like where

to achieve the fall detection by two stages, one is the rough 
fall detection which devides these 9 large-motions into non- Elow = |Z (m, f d )|2 ,  = { f 5 < | f d | < f6 } . (15)
m f =
fall and fall-like events, and the other is the enhanced fall
detection which finally determines whether a fall happens. Here, f 5 and f6 represent the low bound and high bound of
In the following, we detail the proposed rough fall detection the low Doppler frequency. In the experiments, f 5 and f 6 can
approach. be set to 25 Hz and 50 Hz, respectively.
1) Feature Extraction: To extract the differences among non- The fourth feature is the length of event F4 , given by
fall and fall-like events, the following six time-frequency
features are first extracted. F4 = tend − tbegin (16)
The first feature is the energy of high Doppler frequency where tend represents the end time of the event. The method
components F1 , computed by of obtaining tend is the same as that of tbegin . To a certain

F1 = |Z (m, f d )|2 ,  = { f3 < | f d | < f 4 } , (11) extent, a fall and some lagre-motions are separable based on
m f = the feature F4 , as shown in Fig. 5 (a).
where f 3 and f4 represent the low bound and high bound The time-frequency spectrum may consist of weak compo-
of the high Doppler frequency. In the experiments, f 3 and nents, particularly around the extreme frequencies, which may
f 4 can be set to 75 Hz and 100 Hz, respectively. As shown yield broken segments after zero-forcing processing. To obtain
in Fig. 5 (a), compared with non-fall evnets, a fall-like event the extreme Doppler frequency, we perform morphologial
often exhibits a significant amount of energy of high Doppler processing to Z (m, f d ) [27]. Morphological operation can
frequency components. remove disconnected regions and bridge closely related broken
The second feature is the high Doppler frequency informa- segments together to keep the significant time-frequency spec-
tion entropy F2 , given by trum representing the target activities [40]. Then, we extract
the fifth feature, namely the extreme Doppler frequency F5 ,
F2 = −Plog2 (P), (12) defined by
where P is the ratio of F1 to total energy, and described by
F5 = max( f +max , f−max ), (17)
F1
P=   . (13) and the sixth feature, namely the extreme Doppler frequency
f d |Z (m, f d )|
2
m
ratio F6 , given by
The information entropy is a quantitative index of information    
 f −max   f +max 
content of a system. F6 = max   ,   , (18)
The third feature is the ratio of F1 to the energy of the low f +max   f −max 
Doppler frequency components, given by where f+max and f −max represent the maximum frequency
F1 in the positive frequency and the minimum frequency in the
F3 = , (14)
Elow negative frequency range, respectively.

Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on October 15,2024 at 16:07:32 UTC from IEEE Xplore. Restrictions apply.
 CHEN et al.: THREE-STAGE LOW-COMPLEXITY HUMAN FALL DETECTION METHOD USING IR-UWB RADAR 15159

Fig. 5. The time-frequency spectrums and trajectories of a person at a position with different large-motions.

Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on October 15,2024 at 16:07:32 UTC from IEEE Xplore. Restrictions apply.
 15160 IEEE SENSORS JOURNAL, VOL. 22, NO. 15, 1 AUGUST 2022

Fig. 6. The time-frequency spectrums and trajectories obtained by a person standing at a horizontal distance of 1.2m from the radar and falling
towards different directions (0 ◦ , 30 ◦ , 60 ◦ and 90 ◦ ).

Fig. 7. The time-frequency spectrums and trajectories of two people at different positions (0m, 0.6m, 1.2m) from the radar and falling in a direction.

Because the displacement relative to the radar caused by the The seventh feature F7 is the displacement relative to the
fall is greater than that of other motions and the target usually radar caused by motions, given by
stays on the ground after a fall, we extract the following two
features from trajectories. F7 = d1 − d0 , (19)

Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on October 15,2024 at 16:07:32 UTC from IEEE Xplore. Restrictions apply.
 CHEN et al.: THREE-STAGE LOW-COMPLEXITY HUMAN FALL DETECTION METHOD USING IR-UWB RADAR 15161

where d1 denotes the farthest position of the target from the

radar, and d0 represents the position of the target closest to
the radar, as shown in Fig. 5 (b).
The eighth feature F8 is the median of the positions in the
last 1/2s of the t1 seconds data, and can be computed by
F8 = median(T (m 1 )), m 1 ∈ {106, 107, . . . , 120}. (20)
where median(T (m 1 )) is the median of the vector T (m 1 ).
2) SVDD-Based Rough Fall Detection: Here, we select
SVDD for the rough fall detection. SVDD is a spherical
data description algorithm based on the theory of SVM and
can be used for outlier detection to detect non-characteristic
objects in the dataset [41]. Given a training set y ∈ R kk×dd ,
where kk is the number of samples and dd is the feature Fig. 8. Three-feature distributions of falls and other fall-like events.
dimension. First, the training sample data is mapped from
the original space to the feature space through nonlinear
transformation function. There are multiple kernel functions number of non-fall events. SVDD is suitable for dealing with
for choosing. Usually, the Gaussian kernel function is a good the binary classification problem with imbalanced data [42],
choice. Then, we look for the smallest volume hypersphere which can be proved by the experimental results of offline
that contains most of the mapped samples in the feature dataset in section IV. Secondly, the computational complexity
space as the SVDD model. In the experiments, the feature of SVDD is relatively low and can be used in an embedded
set y = [F1 , F2 , F3 , F4 , F5 , F6 , F7 , F8 ] extracted from a large system, and the details are in section IV. Thirdly, we are
number of non-fall events is used to train a SVDD model. surprised to find that the SVDD detector can distinguish fall-
Here, we add a few negative samples (60 random fall-like like events from non-fall events and other arbitrary daily
events) to the positive training set to prevent over fitting. activities, even though those other daily activities events are
Finally, assuming that the sample set belonging to the not used to train the SVDD model.
support vector in the training data set is Sv, the radius of
the hypersphere r and the distance d from the test sample yt C. Stage 3: Enhanced Fall Detection
to the center of the hypersphere are calculated to determine
If a fall-like motion is detected, we have to finally determine
whether a fall-like motion occurs
whether it is a fall. Through the above analysis, we know that
H2
> fall and other fall-like events are indistinguishable only using
d < r, (21)
the high Dopper frequency spectrum components. Therefore,
H3
to extract the differences between a fall and other fall-like
where, H2 means that a fall-like event occurs, H3 means a motions, the ninth feature, namely the average velocity of the
non-fall event occurs, human torso, is proposed and given by

kk 
kk 
kk
F9 = mean (center(m)) , (24)
r= K (yv , yv )−2 αi K (yv , yi )+ αi α j K (yi , y j ),
i=1 i=1 j =1 where, mean(center(m)) is the mean of the vector center(m),
(22) and center(m) is the centroid of the time-frequency vector at
each time m and defined as
and 
f d f d ∗ Z (m, f d )

kk 
kk 
kk center(m) =  . (25)
d= K (yt , yt )−2 αi K (yt , yi )+ αi α j K (yi , y j ). f d Z (m, f d )

i=1 i=1 j =1 The feature F9 is extracted to catch the difference between

(23) trunk motions and limb motions, such as waving.
(y −y ) Fig. 8 shows the three-feature distributions of falls and
− i 2j
Here, yv ∈ Sv, K (yi , y j ) = e is the Gaussian kernel
2σ other fall-like events. From the plots, we observed that F7 ,
function, σ is the variance of the Gaussian kernel function, F8 , and F9 can be uesd to classify a fall and other fall-like
αi ∈ [0, C1 ] and α j ∈ [0, C2 ] are Lagrangian coefficients motions in a certain. By considering that the Mahalanobis
corresponding to the samples yi and y j respectively, C1 is distance classifier is effective to calculate the similarity of two
the penalty positive-cost parameter, and C2 is the penalty unknown sample sets and its computational complexity is very
negtive-cost parameter. In the experiments, the Gaussian kernel low, we use it to finally determine whether a fall happened.
function with σ =21.8 is used, C1 and C2 are set as 0.5 and Thus the final inputs of Mahalanobis distance classifier is
0.4 respectively. y = [F7 , F8 , F9 ]. By using the training dataset, incluing

It is noted that there are three reasons for choosing SVDD fall and other fall-like events, such as waving one’s hand
as the rough fall detector. Firstly, it is very difficult to obtain down, squatting down, sitting down and some other motions,
q q q
a large number of fall events, but it is simple to obtain a large we compute the corresponding mean μq = [μ F7 , μ F8 , μ F9 ]

Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on October 15,2024 at 16:07:32 UTC from IEEE Xplore. Restrictions apply.
 15162 IEEE SENSORS JOURNAL, VOL. 22, NO. 15, 1 AUGUST 2022

Fig. 10. Real-time edge fall detection system.

Fig. 9. Three-feature distributions of falls and other events.

q q q
and variance σ q = [σ F7 , σ F8 , σ F9 ]. Here, the superscript q
denotes the class of events, i.e., q ∈{fall, others}. After we
obtain 
y from the real-time data, the fall detection can be
conducted by
H4 Fig. 11. Real-time fall monitoring interface.
> f all
Dyot hers < Dy , (26)
H5
where, H4 means that a fall occurs and H5 means oppositely,
q
Dy = yT − μq )T ( q )−1 (
( y − σ q ), (27)

where, (·)T denotes transpose, and

q q q
 q = diag[σ F7 , σ F8 , σ F9 ]. (28)
Here, diag(σ q )
stands for a diagonal matrix with the main
diagonal taken from the elements of the vector σ q . It should
be noted that it is not feasible to use only the Mahalanobis Fig. 12. The three real-time test scenes.
distance classifier for fall detection based on  y without using
the SVDD detector. To illustrate this, Fig. 9 shows the three-
feature distributions of falls and other events. From the results, access memory (SRAM). The proposed method is imple-
we can see that falls and other events are seriously aliased. mented in XTMCU02 by C language without cloud servers.
It should be noted that the main contribution of the proposed The ESP32 is the wireless transmission module, and it receives
method lies that a rough fall classifier (SVDD) cascaded with the outputs of XTMCU02 through the serial port and sends
an enhanced fall classifier (Mahalanobis distance classifier) data to the computer by WiFi. The computer receives the
instead of one fall classifier. By using the cascaded classifiers, results of the embeded system and displays them in a designed
we can: (1) reduce the difficulty of identifying falls directly monitoring interface. As shown in Fig. 11, the monitoring
from daily events by using a large number of non-fall samples interface includes the fall alarm interface, the daily activities
to train the SVDD model for anomaly detection, and allowing interface, the motion trajectory and the fall log.
a certain false alarm rate; and (2) achieve a higher fall Finally, as shown in Fig. 12, three real-time test scenes,
detection accuracy in a much smaller searching space by including office, bathroom and bedroom, are selected to verify
identifying falls only from the fall-like events. the real-time edge fall detection system. The results will be
described in detail in section IV.
D. Real-Time Edge Fall Detection System
Fig. 10 shows the proposed real-time edge fall detection IV. E XPERIMENTS AND R ESULTS
system. The system consists of the signal transmitting and The experimental scene is a 3.5m×5m office with tables,
receiving module, the acquisition module, the wireless trans- chairs and a anti-fall cushion, as shown in Fig. 13. The
mission module and mobile interaction module. The X4A02 is radar is installed at the top with a height of three meters.
the signal transmitting and receiving module, and X4SIP02 is There are eleven subjects, including eight males and three
the acquisition module. The core of XTMCU02 is advanced females, participating in the experiments, and all participants
RISC machine (ARM) Cortex M7, works at 300MHz, and are healthy adults. Table II shows the physical information
includes 2048kb flash memory and 384Kb sramstatic random of the participants. It is worth mentioning that the data

Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on October 15,2024 at 16:07:32 UTC from IEEE Xplore. Restrictions apply.
 CHEN et al.: THREE-STAGE LOW-COMPLEXITY HUMAN FALL DETECTION METHOD USING IR-UWB RADAR 15163

Fig. 14. The comparison of different time duration.

Fig. 13. The experimental scene.

TABLE II
T HE P HYSICAL I NFORMATION OF PARTICIPANTS

acquisition was completed in more than five months. In the Fig. 15. TPR and FPR curves of human fall detection obtained by
different values of r of the SVDD model.
experiments, the perticipants perform 9 types of large-motions.
Each motion lasts 10 seconds, corresponding to the time of the
data acquisition for each sample. Participants simulate a fall Secondly, we analyze the influence of the radius parame-
toward 0◦ , 30◦ , 60◦ , 90◦, 120◦ ,150◦ or 180◦. Thus, there are ter (r ) of the SVDD model on the detection performance of
a total of seven types of fall. Each type of falls is repeated the proposed method. Fig. 15 shows the TPR and FPR curves
2 or 3 times at three positions, including 0m, 0.6m, and 1.2m. of human fall detection obtained by different values of r of the
For sittng down and sitting up, there are five chairs around the SVDD model. From the figure, we can observe that: (1) when
office, and each participant performs sitting down and sitting r is set as 0.8726, FPR is only 0.79%, and TPR is up to
up 9 or 10 times on each chair, respectively. For the other 99.09%; (2) when r is less than 0.8726, FPR decreases slowly
large-motions, each type of them is repeated 15 or 20 times with the decrease of r , while TPR decreases very quickly; and
at three positions, including 0m, 0.6m, and 1.2m. In this (3) when r is greater than 0.8726, TPR increases slowly with
experiment, the number of experimental samples obtained by the increase of r , while FPR increases very fast. Therefore,
each participant is different because some participants don’t 0.8726 is a good choice for the parameter r .
have much free time. In total, we collect 464 fall samples and Here, we use the five-fold cross-validation strategy to verify
4432 other large-motion samples, as shown in Table III. This the accuracy of the proposed method using the offline dataset.
paper addresses a binary classification problem, where the fall The offline dataset is randomly divided into five parts, four of
represents the positive class and other motions represent the which are used for training and the fifth is used for testing.
negative class. Three metrics, including the global accuracy, Additionally, in order to verify the superiority of SVDD in
false positive rate (FPR) and true positive rate (TPR), are used dealing with the binary classification problem with imbalanced
to evaluate the performance of the proposed method. data, we use the k-Nearest Neighbor (KNN) or SVM detector
for rough fall detection, and a Mahalanobis distance classifier
A. Results of Offline Dataset for enhanced fall detection. As shown in Table IV, we can
In this example, we first evaluate the inputs of the time observe that: (1) the average global accuracy of the proposed
duration of spectrums on the detection performance of the method is up to 99.4%, the true positive rate is up to 97.6% and
proposed method, as shown in Fig. 14. From the figure, we can false positive rate is only 0.4%; (2) compared with other TML
see that a short time duration may causes poor performance methods, the superiority of SVDD in dealing with the binary
and a long time duration leads to a larger time delay and a classification problem with imbalanced data is illustrated; and
higher computational complexity. With the increasing of time (3) if we only use one classifier with the same inputs (the nine
duration, some information such as noise and other motions features mentioned) to distinguish falls from other motions, the
may reduce the performance of fall detection. Thus, the time performance of fall detection degrades. The true positive rate
duration of 4 seconds is a good choice. of using only one SVDD detector or one Mahalanobis distance

Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on October 15,2024 at 16:07:32 UTC from IEEE Xplore. Restrictions apply.
 15164 IEEE SENSORS JOURNAL, VOL. 22, NO. 15, 1 AUGUST 2022

TABLE III
T HE N UMBER OF S AMPLES OF E ACH M OTION T YPE

TABLE IV experimental samples and the number of misclassified samples

E XPERIMENTAL R ESULTS OF O FFLINE D ATASET U SING for each participant. It can be seen from Table V that the
F IVE -F OLD C ROSS -VALIDATION S TRATEGY
average global accuracy of LOSO is up to 99.2% and the
average true positive rate is also up to 97.6%. This verifies
the strong robustness of the proposed method to different
people.

B. Results of Edge Real-Time System

In this section, we test the edge real-time system in three
scenes, i.e., office, bathroom and bedroom, shown in Fig. 12.
In particular, in order to verify the robustness of the proposed
method to different rooms, we only use the dataset from the
office to train the classifier model and verify the robustness
of the proposed method in the bathroom and bedroom. In the
office, the participants A-K simulated a fall while standing
or sitting, rolling off the bed, and doing other arbitrary daily
activities. Ultimately, 80 falls and 814 other motions were
captured in one month, where 2 falls and 5 other large-motions
are misclassified (2 false negatives and 5 false positives). The
5 false positives come from participants simulated squatting
down very fastly. This is because these motions are very
similar to falls from the point view of radar. In the bathroom,
the participants J-K simulated a person sliding to the floor,
washing hands, squatting down and picking up things. In one
day, 36 falls were tested and 162 other motions were captured,
where 3 falls are misclassified (3 false negatives) and other
motions are all classified correctly (0 false positives). The
reason for the misclassification of the 3 falls is that the
top of bathroom is not three meters high enough, which
affects the features we extract. In the bedroom, the participants
I-K simulated a person falling between the closet and the
bed, rolling off the bed, falling when getting up, and doing
classifier are 91.5% and 92.6%, respectively. Meanwhile, the other arbitrary daily activities. In one day, 54 falls were
false positive rates of using only one SVDD detector or one tested and 240 other motions were captured, where 5 falls
Mahalanobis distance classifier are much higher than that of and 7 other motions are misclassified (5 false negatives and
the proposed method. This verifies the superior classification 7 false positives). The environment of bedroom is relatively
accuracy of the proposed method. However, although the complicated, and the motion of participants falling between
Doppler shift sensitive to the attitude angle of motions can the closet and the bed may be misclassified. Although there
be largely reduced by the installation of the radar sensor, this are some misclassifications, the performance of the proposed
issue has not been completely solved. From the experimental method is excellent as shown in Table VI. The results of the
results, we found that the events detected by mistake are edge real-time system effectively show that the robustness of
basically the fall samples of people standing at a distance of the proposed method to different rooms.
1.2m and falling towards 90◦ . In the second example, we evaluate the impact of radar
Additionaly, in order to verify the robustness of the pro- installation height on the fall detection performance of the
posed method to different people, we use the leave-one- proposed method. We install the radar on the top of the office
subject-out (LOSO) strategy to divide the dataset. The idea of with heights of 3.2m, 3m, 2.9m, 2.7m, 2.6m and 2.5m for fall
LOSO is to select M −1 persons represent the training dataset detection, respectively. It is worth noting that we only use the
and the Mth person as the testing dataset. In the experiments, samples obtained when the radar is installed at the top with a
M is equal to 11. Since the number of experimental samples height of 3m to train the model. We tested a total of six people,
obtained by each participant is different, we give the number of three of whom did not participate in the training of the model.

Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on October 15,2024 at 16:07:32 UTC from IEEE Xplore. Restrictions apply.
 CHEN et al.: THREE-STAGE LOW-COMPLEXITY HUMAN FALL DETECTION METHOD USING IR-UWB RADAR 15165

TABLE V
E XPERIMENTAL R ESULTS OF THE P ROPOSED M ETHOD U SING LOSO S TRATEGY

TABLE VI
E XPERIMENTAL R ESULTS OF E DGE R EAL -T IME
S YSTEM IN T HREE S CENES

TABLE VII
E XPERIMENTAL R ESULTS OF E DGE R EAL -T IME S YSTEM AT
D IFFERENT R ADAR I NSTALLATION H EIGHTS

At each height, 80 falls and more than 800 other motions were
captured in two weeks. The results are shown in Table VII.
From the results, we can observe that: (1) when the radar is
installed at a height of 3m, the global accuracy rate and the true
positive rate of fall detection are the highest; (2) when the radar
is installed in the height range of 2.7m to 3.2m, the global Fig. 16. The time-frequency spectrums and trajectories obtained by one
accuracy of the proposed method is higher than 98.6% and the people at a position and falling in a direction when the radar is installed
at different heights (3.2m, 3m, 2.9m, 2.7m, 2.6m and 2.5m).
true positive rate of the proposed method is higher than 91.3%;
and (3) if the radar is installed at a height below 2.7m, it will
greatly affect the true positive rate of the proposed method.
relative distance between the human body and the radar
The true positive rates of the proposed method are only 78.8%
remains unchanged (for example, when we place the radar
and 58.8% when the radar is installed at a height of 2.6m
at 2.7m, the radar detection range can be set as 0.7-5.7m).
and 2.5m, respectively. This shows that for a certain range
(2) Fig. 16(b) shows the time-frequency spectrums obtained
of installation height, the proposed method exhibits relative
by one people at a position and falling in a direction when the
robustness. The reason for the above phenomenon is that when
radar is installed at different heights, and we can see that the
we change the installation height of the radar with a certain
change of time-frequency characteristics observed by the radar
range, the difference between the characteristics observed by
is small. However, different radar installation heights will lead
the radar and the characteristics of the training samples is still
to changes in the signal-to-noise ratio, which will affect the
relatively small. For better understanding, two points of view
time-frequency features. It can be seen from Table VII that
are provided.
when the installation height of radar is lower than 2.6m, the
(1) Fig. 16(a) shows the trajectories obtained by one people
difference between the test sample and the training sample
at a position and falling in a direction when the radar is
reaches a certain extent, and the accuracy of human fall
installed at different heights. It is obvious that the posi-
detection significantly degrades.
tion features (F7 and F8 ) observed by the radar are basi-
cally unchanged, because in the radar parameter configuration
(Table I), the detection range of the radar is set as 1-6m (the C. Comparison With Other State-of-the-Art Methods
radar is placed at the height of 3m). When we change the Here, we compare the performance of our work with other
installation height of the radar, we can adjust the detection state-of-the-art methods in Table VIII. It can be seen that
range of the radar accordingly, which can ensure that the although we only use one UWB radar, the accuracy rates

Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on October 15,2024 at 16:07:32 UTC from IEEE Xplore. Restrictions apply.
 15166 IEEE SENSORS JOURNAL, VOL. 22, NO. 15, 1 AUGUST 2022

TABLE VIII
T HE P HYSICAL I NFORMATION OF S UBJECTS

TABLE IX memory consumption of the proposed method are much lower

T HE C OMPLEXITY OF D IFFERENT M ODELS than those of the DNN-based methods. Furthermore, it is worth
noting that the computational complexity of the proposed
method is a little lower than that of the SVM-based method.
The reasons are that (1) the proposed method only uses non-
fall samples and a few fall-like samples for the training of
the SVDD model, and uses other fall-like samples for the
training of the Mahalanobis distance classifier, but the SVM-
based method use all experimental samples to train the SVM
model; and (2) the computational complexity of Mahalanobis
distance classifier is very low. Additionally, we implemented
the proposed low complexity algorithm on a limited hardware
of proposed method using five-fold cross-validation strategy
resources, and prove this in real applications.
and LOSO strategy are higher than that of other state-of-
the-art methods. Furthermore, we verified the robustness of
the proposed method in different rooms and different radar V. C ONCLUSION
installation heights in real-time. This paper proposed a three-stage low-complexity human
In the experiments, almost all the studies install the radar fall detection method using an IR-UWB radar with limited
on the side of the room. However, from a large number of fall sample support and computational resource. Firstly, con-
fall experiments, we find that the time-frequency spectrums sidering that fall is a sparse event in daily life, we extract the
when radar is placed on the side are much more sensitive to high-frequency energy feature to determine whether a large-
the people falling angle relative to radar than that when the motion occurs. Secondly, we found that it is not a good
radar is placed on the top. Thus, we install the radar on the choice to distinguish falls from all other large-motions by
top of the room. Furthermore, most of the traditional machine using a classifier with the same inputs, and then we achieve
learning-based methods only use time-frequency features for the fall detection by using two stages, namely the rough fall
fall detection. In this paper, we not only use the time-frequency detection and the enhanced fall detection. A large number
features, but also the position features of the human body of experiments with off-line and real-time are conducted.
relative to the radar for fall detection. Meanwhile, most of The results showed that the proposed method exhibits a low
studies only use one classifier with the same inputs for computational complexity and can achieve an average fall
fall detection. It is worth noting that we use two cascaded detection accuracy above 97% with the false alarm rate smaller
classifiers, and each of which uses different features from than 1%. Additionally, the LOSO tests and the real-time
different datasets. Especially, we are surprised to find that tests illustrate that the proposed method is robust to different
the proposed method can distinguish falls from other large- rooms (including office, bathroom and bedroom), the radar
motions and arbitrary daily activities. installation height, the humans without training the classifiers,
Finally, We use the same one sample to compare our work and a large number of daily actions, even the motions without
with other state-of-the-art methods in terms of computational training the classifiers. Since the problem that the Doppler
complexity and memory consumption, as shown in Table IX. shift is sensitive to the attitude angle of motions has not
The computational complexity is mainly described with the been completely solved, we will consider using multistatic
number of multiplications and additions, and the memory radar to obtain the angle information of the target to solve
consumption is described with the number of parameters. the sensitivity of the radar to the attitude angle of falls in the
From the table, we note that the computational complexity and future research.

Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on October 15,2024 at 16:07:32 UTC from IEEE Xplore. Restrictions apply.
 CHEN et al.: THREE-STAGE LOW-COMPLEXITY HUMAN FALL DETECTION METHOD USING IR-UWB RADAR 15167

R EFERENCES [25] H. Sadreazami, M. Bolic, and S. Rajan, “Residual network-based super-

vised learning of remotely sensed fall incidents using ultra-wideband
[1] H. Ritchie and M. Roser. (2020). Age Structure. Our World in Data. radar,” in Proc. IEEE Int. Symp. Circuits Syst. (ISCAS), May 2019,
[Online]. Available: https://ourworldindata.org/age-structure pp. 1–4.
[2] S. A. Alshammari et al., “Falls among elderly and its relation with their [26] B. Jokanovic, M. Amin, and F. Ahmad, “Radar fall motion detec-
health problems and surrounding environmental factors in Riyadh,” J. tion using deep learning,” in Proc. IEEE Radar Conf. (RadarConf),
Family Community Med., vol. 23, no. 1, pp. 29–34, 2018. May 2016, pp. 1–6.
[3] C. van Weel, H. Vermeulen, and W. van den Bosch, “Falls, a community [27] L. R. Rivera, E. Ulmer, Y. D. Zhang, W. Tao, and M. G. Amin, “Radar-
care perspective,” Lancet, vol. 345, no. 8964, pp. 1549–1551, Jun. 1995. based fall detection exploiting time-frequency features,” in Proc. IEEE
[4] S. Nooruddin et al., “Sensor-based fall detection systems: A review,” China Summit Int. Conf. Signal Inf. Process. (ChinaSIP), Jul. 2014,
J. Ambient Intell. Hum. Comput., vol. 13, pp. 1–17, Apr. 2021. pp. 713–717.
[5] S. A. Mousavi, F. Heidari, E. Tahami, and M. Azarnoosh, “Fall detection [28] B. Erol, M. G. Amin, B. Boashash, F. Ahmad, and Y. D. Zhang, “Wide-
system via smart phone and send people location,” in Proc. 28th Eur. band radar based fall motion detection for a generic elderly,” in Proc.
Signal Process. Conf. (EUSIPCO), Jan. 2021, pp. 1605–1607. 50th Asilomar Conf. Signals, Syst. Comput., Nov. 2016, pp. 1768–1772.
[6] G.-M. Sung, H.-K. Wang, and W.-T. Su, “Smart home care system with [29] Z. Zhou, J. Zhang, and Y. D. Zhang, “Ultra-wideband radar and vision
fall detection based on the Android platform,” in Proc. IEEE Int. Conf. based human motion classification for assisted living,” in Proc. IEEE
Syst., Man, Cybern. (SMC), Oct. 2020, pp. 3886–3890. Sensor Array Multichannel Signal Process. Workshop (SAM), Jul. 2016,
[7] Y.-H. Nho, S. Ryu, and D.-S. Kwon, “UI-GAN: Generative adversarial pp. 1–5.
network-based anomaly detection using user initial information for [30] A. Shrestha, J. Le Kernec, F. Fioranelli, E. Cippitelli, E. Gambi, and
wearable devices,” IEEE Sensors J., vol. 21, no. 8, pp. 9949–9958, S. Spinsante, “Feature diversity for fall detection and human indoor
Apr. 2021. activities classification using radar systems,” in Proc. Int. Conf. Radar
[8] K. Desai, P. Mane, M. Dsilva, A. Zare, P. Shingala, and D. Ambawade, Syst., 2017, pp. 1–6.
“A novel machine learning based wearable belt for fall detection,” in [31] G. Diraco, A. Leone, and P. Siciliano, “Radar sensing technology for fall
Proc. IEEE Int. Conf. Comput., Power Commun. Technol. (GUCON), detection under near real-life conditions,” in Proc. 2nd IET Int. Conf.
Oct. 2020, pp. 502–505. Technol. Act. Assist. Living (TechAAL), 2016, pp. 1–6.
[9] J.-S. Lee and H.-H. Tseng, “Development of an enhanced threshold- [32] A. Alnaeb et al., “Human activities detection and classification based on
based fall detection system using smartphones with built-in accelerom- micro-Doppler signatures near the baseline of forward scattering radar,”
eters,” IEEE Sensors J., vol. 19, no. 18, pp. 8293–8302, Sep. 2019. in Proc. Int. Conf. Radar (RADAR), Aug. 2018, pp. 1–6.
[10] H. Abdo, K. M. Amin, and A. M. Hamad, “Fall detection based on [33] B. Erol, M. G. Amin, and B. Boashash, “Range-Doppler radar sensor
RetinaNet and MobileNet convolutional neural networks,” in Proc. 15th fusion for fall detection,” in Proc. IEEE Radar Conf. (RadarConf),
Int. Conf. Comput. Eng. Syst. (ICCES), Dec. 2020, pp. 1–7. May 2017, pp. 0819–0824.
[11] Z. Huang, Y. Liu, Y. Fang, and B. K. P. Horn, “Video-based fall [34] A. Gadde et al., “Fall detection and classification based on time-scale
detection for seniors with human pose estimation,” in Proc. 4th Int. radar signal characteristics,” Proc. SPIE, vol. 9077, pp. 1–9, May 2014.
Conf. Universal Village (UV), Oct. 2018, pp. 1–4. [35] B. Y. Su, K. C. Ho, M. J. Rantz, and M. Skubic, “Doppler radar fall
[12] P. S. Sase and S. H. Bhandari, “Human fall detection using depth activity detection using the wavelet transform,” IEEE Trans. Biomed.
videos,” in Proc. 5th Int. Conf. Signal Process. Integr. Netw. (SPIN), Eng., vol. 62, no. 3, pp. 865–875, Mar. 2015.
Feb. 2018, pp. 546–549. [36] (2018). Novelda’s XeThruX4M03. [Online]. Available:
[13] Y. Ogawa and K. Naito, “Fall detection scheme based on temperature https://www.xethru.com/x4-radar-chip.html/
distribution with IR array sensor,” in Proc. IEEE Int. Conf. Consum. [37] I. Setitra and S. Larabi, “Background subtraction algorithms with post-
Electron. (ICCE), Jan. 2020, pp. 1–5. processing: A review,” in Proc. 22nd Int. Conf. Pattern Recognit.,
[14] C. Nadee, P. Yingkayun, and K. Yingkayun, “Fall detection using half Aug. 2014, pp. 2436–2441.
circle ultrasonic array sensors for indoor environment,” in Proc. 18th [38] X. Wang, T. Ying, and W. Tian, “Spectrum representation based on
Int. Conf. Electr. Eng./Electron., Comput., Telecommun. Inf. Technol. STFT,” in Proc. 13th Int. Congr. Image Signal Process., Biomed. Eng.
(ECTI-CON), May 2021, pp. 879–882. Informat. (CISP-BMEI), Oct. 2020, pp. 435–438.
[15] J. Clemente, F. Li, M. Valero, and W. Song, “Smart seismic sensing for [39] J. Lai, Z. Yang, and B. Guo, “A two-stage low-complexity human sleep
indoor fall detection, location, and notification,” IEEE J. Biomed. Health motion classification method using IR-UWB,” IEEE Sensors J., vol. 21,
Informat., vol. 24, no. 2, pp. 524–532, Feb. 2020. no. 18, pp. 20740–20749, Sep. 2021.
[40] R. Gonzalez and R. Woods, Digital Image Processing. New York, NY,
[16] S. Kianoush, S. Savazzi, F. Vicentini, V. Rampa, and M. Giussani,
USA: Oxford, 2001.
“Leveraging RF signals for human sensing: Fall detection and local-
[41] D. M. J. Tax and R. P. W. Duin, “Support vector domain descrip-
ization in human-machine shared workspaces,” in Proc. IEEE 13th Int.
tion,” Pattern Recognit. Lett., vol. 20, nos. 11–13, pp. 1191–1199,
Conf. Ind. Informat. (INDIN), Jul. 2015, pp. 1456–1462.
1999.
[17] M. G. Amin, Y. D. Zhang, F. Ahmad, and K. C. D. Ho, “Radar signal [42] Z. M. Miao et al., “Support vector date description implemented in
processing for elderly fall detection: The future for in-home monitoring,” class-imbalance learning,” J. Appl. Sci., vol. 26, no. 1, pp. 79–84,
IEEE Signal Process. Mag., vol. 33, no. 2, pp. 71–80, Mar. 2016. 2008.
[18] C. W. Ding et al., “Non-contact human motion recognition based on
UWB radar,” IEEE J. Emerg. Sel. Topics Circuits Syst., vol. 8, no. 2,
pp. 306–315, Jun. 2018.
[19] Y. Wang, J. Zhou, J. Tong, and X. Wu, “UWB-radar-based synchronous
motion recognition using time-varying range–Doppler images,” IET
Radar, Sonar Navigat., vol. 13, no. 12, pp. 2131–2139, Dec. 2019.
[20] H. Sadreazami, M. Bolic, and S. Rajan, “Fall detection using standoff
radar-based sensing and deep convolutional neural network,” IEEE
Trans. Circuits Syst. II, Exp. Briefs, vol. 67, no. 1, pp. 197–201,
Jan. 2020.
[21] B. Wang, L. Guo, H. Zhang, and Y.-X. Guo, “A millimetre-wave
radar-based fall detection method using line kernel convolutional neural
Mengxia Chen received the B.E. degree in com-
network,” IEEE Sensors J., vol. 20, no. 22, pp. 13364–13370, Nov. 2020.
munication engineering from Chongqing Uni-
[22] H. Sadreazami, M. Bolic, and S. Rajan, “Contactless fall detection
versity, Chongqing, China, in June 2016. She
using time-frequency analysis and convolutional neural networks,” IEEE
is currently pursuing the M.E. degree in elec-
Trans. Ind. Informat., vol. 17, no. 10, pp. 6842–6851, Oct. 2021.
tronic information engineering with Shenzhen
[23] J. Maitre, K. Bouchard, and S. Gaboury, “Fall detection with UWB University. Her research focuses are radar sig-
radars and CNN-LSTM architecture,” IEEE J. Biomed. Health Informat., nal processing and especially radar-based fall
vol. 25, no. 4, pp. 1273–1283, Apr. 2021. detection.
[24] H. Sadreazami, M. Bolic, and S. Rajan, “Capsfall: Fall detection
using ultra-wideband radar and capsule network,” IEEE Access, vol. 7,
pp. 55336–55343, 2019.

Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on October 15,2024 at 16:07:32 UTC from IEEE Xplore. Restrictions apply.
 15168 IEEE SENSORS JOURNAL, VOL. 22, NO. 15, 1 AUGUST 2022

Zhaocheng Yang (Member, IEEE) received the Ping Chu received the B.E. and Ph.D. degrees
B.E. degree in information engineering from the in information and communication engineer-
Beijing Institute of Technology, Beijing, China, ing from Harbin Engineering University, Harbin,
in 2007, and the Ph.D. degree in information and China, in 2007 and 2011, respectively. Since
communication engineering from the National 2011, she has been with the College of Elec-
University of Defense Technology, Changsha, tronics and Information Engineering, Shenzhen
China, in 2013. University, where she is currently a Lecturer.
From November 2010 to November 2011, Her research interests include the area of signal
he was a Visiting Scholar with the University of processing, including array signal processing,
York, York, U.K. From June 2013 to August 2015, and its applications to radar systems.
he was a Lecturer with the School of Electronics
Science and Engineering, National University of Defense Technology.
He is currently an Associate Professor with the College of Electronics and
Information Engineering, Shenzhen University, Shenzhen, China. His
research interests include the area of signal processing, including array
signal processing, adaptive signal processing, compressive sensing,
machine learning, and its applications to radar systems.

Jialei Lai received the B.E. degree from Jinghong Lin is currently pursuing the B.E.
the School of Electronic Information Engineer- degree from the School of Electronics and Infor-
ing, Shenzhen University, Shenzhen, China, mation Engineering, Shenzhen University. His
in June 2019, where he is currently pursuing research focuses are radar signal processing and
the M.E. degree. His research focuses are radar especially the fall detection based on radar.
signal processing and pattern recognition, espe-
cially about the radar-based human target detec-
tion, and motion recognition.

Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on October 15,2024 at 16:07:32 UTC from IEEE Xplore. Restrictions apply.