biosensors

Article
Research on Non-Contact Monitoring System for
Human Physiological Signal and Body Movement
Qiancheng Liang 1 , Lisheng Xu 1,2, *, Nan Bao 1 , Lin Qi 1 , Jingjing Shi 1 , Yicheng Yang 1
and Yudong Yao 1
1 School of Sino-Dutch Biomedical and Information Engineering, Northeastern University,
Shenyang 110819, China; vin_liangqc@foxmail.com (Q.L.); baonan@bmie.neu.edu.cn (N.B.);
qilin@bmie.neu.edu.cn (L.Q.); shijj@bmie.neu.edu.cn (J.S.); y.ang-lin@163.com (Y.Y.);
yyao@bmie.neu.edu.cn (Y.Y.)
2 Neusoft Research of Intelligent Healthcare Technology, Co. Ltd., Shenyang 110167, China
* Correspondence: xuls@bmie.neu.edu.cn




Received: 28 February 2019; Accepted: 16 April 2019; Published: 19 April 2019


Abstract: With the rapid increase in the development of miniaturized sensors and embedded devices for
vital signs monitoring, personal physiological signal monitoring devices are becoming popular. However,
physiological monitoring devices which are worn on the body normally affect the daily activities of
people. This problem can be avoided by using a non-contact measuring device like the Doppler radar
system, which is more convenient, is private compared to video monitoring, infrared monitoring and
other non-contact methods. Additionally real-time physiological monitoring with the Doppler radar
system can also obtain signal changes caused by motion changes. As a result, the Doppler radar system
not only obtains the information of respiratory and cardiac signals, but also obtains information about
body movement. The relevant RF technology could eliminate some interference from body motion
with a small amplitude. However, the motion recognition method can also be used to classify related
body motion signals. In this paper, a vital sign and body movement monitoring system worked at
2.4 GHz was proposed. It can measure various physiological signs of the human body in a non-contact
manner. The accuracy of the non-contact physiological signal monitoring system was analyzed. First,
the working distance of the system was tested. Then, the algorithm of mining collective motion signal
was classified, and the accuracy was 88%, which could be further improved in the system. In addition,
the mean absolute error values of heart rate and respiratory rate were 0.8 beats/min and 3.5 beats/min,
respectively, and the reliability of the system was verified by comparing the respiratory waveforms with
the contact equipment at different distances.

Keywords: doppler bio-radar; non-contact monitoring system; body movement classify;

physiological signals

1. Introduction
Doppler bio-radar has the ability to monitor physiological signals, which depend on the Doppler
Effect. Electromagnetic waves can travel through clothing to the surface of the chest cavity and
reflect. The antenna receives the reflected waveform containing information on chest displacement.
This movement is mainly caused by breathing and the heartbeat [1,2], when the body is still. Otherwise,
it contains information about body movement.
Channel State Information (CSI) [3] used a WiFi signal to do research on action recognition [4],
indoor localization [5], fall detection [6] and physiological signal detection [7]. Those kinds of
studies did not need hardware by using commercial routers and computers. Using self-made devices,

Biosensors 2019, 9, 58; doi:10.3390/bios9020058 www.mdpi.com/journal/biosensors

Biosensors 2019, 9, x FOR PEER REVIEW 2 of 13

Biosensors 2019, 9, 58
devices, Van Dorp [8] judged walking posture from received signals, and Balleri [9] 2used of 13

micro-Doppler signatures to recognize different people by their micro-Doppler signatures.

Van Dorp Self-made devices
[8] judged are also
walking used to
posture frommonitor
receiveda physiological
signals, andsignal,Balleriand subjects
[9] used should also
micro-Doppler
remain stationary
signatures as they
to recognize use the
different system
people byin the micro-Doppler
their radiation direction of the antenna. In practical use,
signatures.
however, there will be weak signals caused by body movement
Self-made devices are also used to monitor a physiological signal, and subjects called Randomshould Body also Movement
remain
(RBM) in the received signal, even though the subject keeps still. The
stationary as they use the system in the radiation direction of the antenna. In practical use, however, frequency characteristic of the
RBM will
there is complicated,
be weak signals whichcaused
makesby thebody
signal hard to be
movement filtered.
called RandomThe common way to deal
Body Movement (RBM) with inthat
the
is to use two identical systems on both sides of the body for compensation.
received signal, even though the subject keeps still. The frequency characteristic of the RBM is complicated, Giovanni [10] monitored
users’ makes
which heart and respiratory
the signal hard torates as they
be filtered. Theperformed
common way different
to dealactivities.
with that Liis toand
useLintwo[11] used
identical
arctangent
systems demodulation,
on both and the
sides of the body other way wasGiovanni
for compensation. to use the [10]subject’s
monitored location information,
users’ heart which
and respiratory
was tracked by amplitude recording to eliminate effects of body movement
rates as they performed different activities. Li and Lin [11] used arctangent demodulation, and the other [12,13]. The existing
methods
way was to canuseonly
the work on location
subject’s body movement
information, within
which a range of 6 cmby
was tracked atamplitude
a speed ofrecording
about 0.4to mm/s [14].
eliminate
effects of body movement [12,13]. The existing methods can only work on body movement withinbe
In the application of monitoring physical information in real life, however, body movement can a
much of
range more6 cm strenuous,
at a speedwhen they0.4
of about are sitting
mm/s down
[14]. In theorapplication
working. Inofthis way, these
monitoring movements
physical could
information
affect
in real the
life, accuracy
however, of body physiological
movement can monitoring
be much strongly.
more strenuous,When when designing
they the Doppler
are sitting down radaror
physiological monitoring system in real time, Khan and Ho Cho [15]
working. In this way, these movements could affect the accuracy of physiological monitoring strongly. set two pre-judgment steps to
make sure the user was in the effective distance without
When designing the Doppler radar physiological monitoring system in real time, Khan and Ho Cho [15] body movement. They used the
auto-correlation
set two pre-judgment width to distinguish
steps to make surebetween the user a wasstate of rest
in the and distance
effective motion. without
Wenda body [16] designed
movement. a
Doppler radar system for indoor use, which could detect breath
They used the auto-correlation width to distinguish between a state of rest and motion. Wenda [16] and movement of standing, sitting,
walking, arunning
designed Dopplerand jumping.
radar system The breathuse,
for indoor monitoring
which could could workbreath
detect withinand1m with an accuracy
movement of standing, of
87%. The classification accuracy using SVM reached round 85%.
sitting, walking, running and jumping. The breath monitoring could work within 1 m with an accuracy of
87%. In Thethis article, a non-contact
classification accuracy using monitoring
SVM reached system for 85%.
round the human physiological signal and body
movementIn thiswas proposed,
article, which was
a non-contact to monitor
monitoring the health
system for thecondition of a person who
human physiological is sitting
signal and body still
for a long time, such as drivers or white-collar workers. Unlike
movement was proposed, which was to monitor the health condition of a person who is sitting still forother systems [17], the proposed
asystem
long time,worked such atas
2.4drivers
GHz with only one antenna,
or white-collar workers. since it could
Unlike other send and receive
systems [17], the signals
proposed in the same
system
time with
worked at the
2.4 GHztransceiver
with only whichone included
antenna, sincetransmitter
it couldcircuit,
send and receiver
receivecircuit and
signals inothers.
the same In time
addition,
with
the antenna did not only work as a sensor, but also transmitted signals
the transceiver which included transmitter circuit, receiver circuit and others. In addition, the antenna to a serial port module
connected
did not only with
work theasupper computer.
a sensor, but alsoThe system could
transmitted signalsgettorespiration
a serial portand heart rates
module after an
connected action
with the
state classification, which could distinguish six kinds of different states.
upper computer. The system could get respiration and heart rates after an action state classification,
which could distinguish six kinds of different states.
2. System Overview
2. System Overview
2.1. Hardware System
2.1. Hardware System
2.1.1. Antenna
2.1.1. Antenna
An antenna plays an important part in a Doppler system, as the performance of the antenna will
affectAntheantenna plays
system. Inanother
important partthe
words, in aapplication
Doppler system, as the
of the performance
Doppler systemof the antenna will
determines the
affect the system. In other words, the application of the Doppler system determines
characteristics of the selected antenna. For monitoring the human physiological signal and body the characteristics of
the selected antenna. For monitoring the human physiological signal and body movement,
movement, the antenna should not be placed too far from people, however, it does require enough the antenna
should
quality not be signal
of the placedtotoo far from
classify andpeople,
extract however, it does
physiological require
signals [18].enough quality
Otherwise, theof the signal
antenna to
needs
classify and extract physiological signals [18]. Otherwise, the antenna needs to
to be focused on only one direction to reduce the influence of other people in the environment. A be focused on only
one direction
dipole to reduce
array antenna thechosen
was influence
for of other
those people in
demands theaenvironment.
with gain of 10.8 dBiA dipole
and thearray antennabeam
half-power was
chosen for those demands with a gain of ◦
width was 49°. The antenna is shown in10.8 dBi 1,
Figure and theitshalf-power
and beam width
three-dimensional was 49pattern
radiation . The antenna
shown
is shown in Figure
in Figure 2 [19]. 1, and its three-dimensional radiation pattern shown in Figure 2 [19].

Figure 1. Geometry model of the printed dipole array antenna.

Figure 1. Geometry model of the printed dipole array antenna.
 Biosensors 2019, 9, x FOR PEER REVIEW 3 of 13

2.1.2. Transceiver
Biosensors 2019, 9, 58 3 of 13
Biosensors 2019, circuit
9, x FOR in
PEER REVIEW 3 of 13
BiosensorsThe the
2019, 9, x FOR PEER Doppler
REVIEW radar system shown below includes a radio frequency 3 of chip,
13
transmitter circuit, receiver circuit and signal processing circuit. The chip handbook provides a
2.1.2. Transceiver
2.1.2. Transceiver
2.1.2. Transceiver
circuit diagram of minimum system and Barron circuit. The signal transmission process is shown in
Figure
The
The 3, after reaching
circuit in the the radar
Dopplerhuman radarbody it echoed
system shown a signal
below with a 2.4
includesGHz ahigh
radiofrequency
frequency pass and a
chip,
The circuit
circuitinin thethe
Doppler
Doppler system
radar shown
system below
shown includes
below includesradio frequency
a radio chip, transmitter
frequency chip,
transmitter
detector
circuit, receivercircuit,
tube wasreceiver
circuit used
andto circuit
remove
signal and signal
the high
processing processing
frequency
circuit. Thenoisecircuit.
chip first. The
handbookTherechip
washandbook
a circulator
provides provides
a circuit diagramaathe
keeping
transmitter circuit, receiver circuit and signal processing circuit. The chip handbook provides
circuit
of signaldiagram
minimum going inof
system minimum
the right system
direction.
and Barron and
Since
circuit. Barron
the circuit.
echo
The signal signalThe signal
which was transmission
amplified process
by is shown
amplification in
circuit
circuit diagram of minimum system and Barron circuit.transmission process is shown
The signal transmission process in Figure
is shown 3,
in
Figure
after 3,
contained
reachingafter a reaching
thelow
human the
frequency
body human
it body
physiological
echoed a it echoed
signal signal
with a signal
and
2.4 GHz with
bodyhigh 2.4 GHz
movement
frequency high frequency
signal
pass together
and a pass and
with
detector tube a
high
Figure 3, after reaching the human body it echoed a signal with 2.4 GHz high frequency pass and a
detector
wasfrequencytube
usedtube was used
noise,
to remove athe to remove
9 Hz
high low thefilter
pass
frequency high
noise frequency
was used
first. innoise
There the first.
wassystem. ThereThe was
a circulator a circulator
circuit
keeping diagram keeping
of the
the signal going the
filter
in is
detector was used to remove the high frequency noise first. There was a circulator keeping the
signal
theshown
rightgoingin in the
Figure
direction. right
4 withdirection.
the
Since direction. Since
Gain-Frequency
the echo signal the echo signal
image. which was amplified by amplification circuit
signal going in the right Since which
the echo was amplified
signal whichby wasamplification
amplified by circuit containedcircuit
amplification a low
contained physiological
frequency a low frequency signal physiological
and body signal and
movement signalbody movement
together with signal
high together
frequency witha 9high
noise, Hz
contained a low frequency physiological signal and body movement signal together with high
frequency
low pass noise,
filter was a used
9 Hz in lowthe pass filterThe
system. wascircuit
used diagram
in the system.
of the The circuit
filter is showndiagram
in of the
Figure 4 filterthe
with is
frequency noise, a 9 Hz low pass filter was used in the system. The circuit diagram of the filter is
shown in
Gain-Frequency Figure 4 with the Gain-Frequency image.
shown in Figure image. 4 with the Gain-Frequency image.

Figure 2. The three-dimensional radiation pattern of the dipole array antenna used in the system.

Figure 2.
2. The three-dimensional radiation
radiation pattern
pattern of
of the
the dipole
dipole array antenna
antenna used in
in the system.
system.
Figure
Figure 2. The three-dimensional radiation pattern of the dipole array
array antenna used
used in the
the system.

Figure 3. Schematic diagram of hardware system.

Figure 3. Schematic diagram of hardware

hardware system.
Figure 3. Schematic diagram of hardware system.
Figure 3. system.

(a)

Figure 4. Cont.

(a)
(a)
Biosensors 2019, 9, x FOR PEER REVIEW 4 of 13
Biosensors 2019, 9, 58 4 of 13

Frequency/Hz

(b)

Low pass
Figure 4. Low
Figure pass filter.
filter. (a)
(a) Circuit
Circuit diagram
diagram of
of low
low pass
pass filter. (b) Bode diagram of low pass filter.

The chip
The chip used
used ininthe
thetransceiver
transceiverhad hadseveral
severalfunctions.
functions. First,
First, it needed
it needed to generate
to generate a signal
a signal at
at 2.4
2.4 GHz.
GHz. Second,
Second, the the analogue
analogue signal
signal hadhadto to
bebe digitized
digitized bybyit.it.Finally,
Finally,the
thedigital
digitalsignal
signal had
had to
to be
be
loaded on the continuous wave sent to the human and the serial port module.
loaded on the continuous wave sent to the human and the serial port module. Thus, nRf24Le01 Thus, nRf24Le01 which
worked at 2.4–2.45
which worked GHz was
at 2.4–2.45 GHz chosen for its for
was chosen 8-bit
itsprocessor, 1KB in-chip
8-bit processor, RAM RAM
1KB in-chip and 14-channel ADC
and 14-channel
channel
ADC up to the
channel up12-bit
to theresolution it had. nRf24Le01
12-bit resolution was also cheap
it had. nRf24Le01 was as a lower
also cheappower
as a processor.
lower power
processor.
2.1.3. Serial Port Module
2.1.3.Previous
Serial Port Module
systems [20] have tended to use cable to connect itself with the upper computer, somehow
wasting the transmit
Previous ability
systems [20]ofhave
the RF signal.toThus,
tended a serialtoport
use cable with radio
connect itself frequency
with the receiving function
upper computer,
was used in this system. It received the digital signal sent from the antenna and
somehow wasting the transmit ability of the RF signal. Thus, a serial port with radio frequency passed it to the host
computer.function
receiving This device enabled
was used the system.
in this system to have a better
It received compatibility
the digital signal sentwith the upper
from computer
the antenna and
though Universal Serial Bus. The digital signal was used by the algorithm in the
passed it to the host computer. This device enabled the system to have a better compatibility withhost computer for
classifying and extracting physiological information.
the upper computer though Universal Serial Bus. The digital signal was used by the algorithm in
the host computer for classifying and extracting physiological information.
2.2. Algorithm Design

2.2. Algorithm
2.2.1. DesignAlgorithm
Classification

2.2.1.To other sensors,

Classification signals produced by human activities can be used by pattern recognition helping
Algorithm
action recognition. Basic sensing like the acceleration sensor has the ability to reach 90 percent accuracy
To other sensors,
in distinguishing signals
sleeping, produced
sitting, standing, byworking
human and activities
running can[21].
be used by pattern recognition
helping
The action
proposed recognition. Basic on
system focused sensing likethe
removing theinfluence
acceleration sensor
of large bodyhas the ability
movement to reach 90
on physiological
percent accuracy in distinguishing sleeping, sitting, standing, working and running
signal monitoring by classifying them, therefore, the states which were chosen for classification were [21].
normalThemovement
proposedorsystem focused
conditions such as:on Static
removing the influence
measurement (A), limbof moving
large body(B), a movement
second person on
physiological signal monitoring by classifying them, therefore, the states which
entering (C), unmanned environment (D) standing up (E) and sitting down (F). The signal of these six were chosen for
classification were normal
different movements movement
collected or conditions
by the system is shown such as: Static
in Figure measurement
5, and the recorded (A), limb moving
distance change
(B), a second person entering (C), unmanned environment (D) standing up
between the antennas to human when they were in a different state. Limb movement, for example, (E) and sitting down (F).
The signal of these six different movements collected by the system is shown
changed the distance between the antenna and participant frequently, which caused dramatic changes in Figure 5, and the
recorded distance change between the antennas to human when they were
in the signal. It also happened when the participant was sitting down or standing up, the drasticin a different state. Limb
movement, for example,when
fluctuation disappeared changed the distancefinished.
the movement betweenAs thea second
antennaperson
and participant
entered thefrequently,
detection
which
zone, his movement would cause the signal change as well. Only when the participant was
caused dramatic changes in the signal. It also happened when the participant sitting
was sitting
down orlarge
without standing up, the
amplitude bodydrastic fluctuation
movement, disappearedsignal
the high-quality whencould
the bemovement
gatheredfinished. As a
by the system.
second person entered the detection zone, his movement would cause the signal
Additionally, action classification may have helped develop other functions of the system, for example, change as well.
Only when the participant was sitting without large amplitude body movement,
the time of sitting could also be measured while every instance of “standing up” and “sitting down” the high-quality
signal could be gathered by the system. Additionally, action classification may have helped develop
was detected.
other functions of the system, for example, the time of sitting could also be measured while every
instance of “standing up” and “sitting down” was detected.
Biosensors 2019, 9, 58 5 of 13
Biosensors 2019, 9, x FOR PEER REVIEW 5 of 13

Amplitude/V

Time/s

(a)
Amplitude/V

Time/s

(b)
Amplitude/V

Time/s

(c)
Amplitude/V

Time/s
(d)
Amplitude/V

Time/s

(e)
Amplitude/V

Time/s
(f)

Figure 5. Signals corresponding to different movement states. (a) Limb moving. (b) Unmanned
environment. (c) Static measurement. (d) A second person entering. (e) Sitting down. (f) Standing up.
Biosensors 2019, 9, 58 6 of 13

In the application of monitoring human movement, artificial neural networks [22,23], naive
Bayes [24] and support vector machines [25] are the most commonly used. In this paper, two different
motion recognition algorithms were designed, which used Support Vector Machine (SVM) and deep
learning to classify motion. Among them, when SVM was used, waveform extraction features were
selected manually, while when deep learning was used, a bidirectional long and short- term memory
network was used for feature extraction of actions. Figure 5 shows the six types of movements that
needed to be classified, and there were 300 samples for each movement. The data was five seconds
long with a sampling rate of 100 Hz.
The features used in SVM were time duration, maximal upper Doppler shift, maximal low
Doppler shift, peak-to-peak bandwidth, mean power of Doppler and standard deviation of Doppler
power. Since one SVM classifier could only distinguish between two types, six classifiers were used
for classification.
Convolution Neural network (CNN) can make the arbitrary dimension input, but often the
input data of the network is two-dimensional. The main idea of CNN is to use the kernel and filter
for multi-scale analysis of data, and the data means of this analysis was convolution. Similarly,
the image processing algorithm for two-dimensional convolution was used in the neural network.
One-dimensional convolution of a one-dimensional signal was to use the convolution function as a
filter to filter the signal once and find the characteristics of the waveform at this frequency.
The signal of the six different movements collected by the system is shown in Figure 4, and it
recorded the distance change between the antenna and the human when they were in different states.
Limb moving, for example, changed the distance between the antenna and participant frequently,
which caused dramatic changes in signal. It also happened when the participant was sitting down or
standing up, while the drastic fluctuation disappeared when the movement finished. As a second person
was entering the detection zone, his movement would cause the signal change as well. Only when
the participant was sitting without large amplitude body movement, the high-quality signal could be
gathered by the system.
By setting the channel to the number of convolutions, the feature dimension extracted in the
convolutional neural network could be changed. The data features in the convolutional neural network
appeared in two-dimensional form as a continuous time series, and the other dimension was the
time of the sequence. However, in the convolutional neural network, the time series was not taken
as continuous signals for information extraction, but each point was taken as a separate unit for
feature extraction.
Different from the image information, there was a strong correlation between the adjacent points
of the one-dimensional time-domain signal. For convolutional neural networks, the elements are
independent of each other, and there is no correlation. Therefore, a special Cyclic Neural Network
(RNN) Long Short-Term Memory (LSTM) network became the first choice to deal with the time domain
signal problem.
When designing the classification network setting of deep learning, the network model of CNN
was used to process the signal at first, because the original signal that was collected contained noise and
waveform information of different frequencies. In addition, due to the difference in action duration,
different action signals had different characteristics in the time domain. Therefore, a CNN network
was required to carry out multi-scale convolution processing on the signal to extract as much as
possible the signal information in each frequency and each time domain. In the CNN signal feature
extraction part, the algorithm used the vgg-16 model and the network structure between conv1_1 and
conv3_3 [26]. In order to ensure the network input format, the network changed to a one-dimensional
input. When passing through the three-layer network of the model, 128 different convolution kernels
were used for each convolution layer, and 128 convolution times were carried out, and 128-dimensional
features were obtained. Then the data was compressed by a pooling layer.
After CNN processing, the signal input bidirectional LSTM was obtained, and the network outputs
with 128 channels in each direction. In the actual process of signal processing, the system mainly
Biosensors 2019,
Biosensors 9, 58
2019, 9, x FOR PEER REVIEW 7 of7 13
of 13
Biosensors 2019, 9, x FOR PEER REVIEW 7 of 13

mainly focused on the peak state of the signal, so a two-way LSTM system was used to ensure that
mainly on
focused focused on the
the peak peak
state of state of the signal,
the signal, so LSTM so a two-way
a two-way LSTM LSTM
system system
was was
usedused to ensure
to ensure thatthat
more
more information was collected. After the network, a CNN network was needed to adjust the
more information was collected. After the LSTM network, a CNN network was needed to adjust the
information was collected. After the LSTM network, a CNN network was needed
output of the LSTM. The algorithm used three full connection layers to extract semantic features, to adjust the output
output of the LSTM. The algorithm used three full connection layers to extract semantic features,
of the
and LSTM. The of
the number algorithm
channelsused three
in their full units
output connection
was 512,layers
512 and to extract semantic features, and the
6, respectively.
and the number of channels in their output units was 512, 512 and 6, respectively.
numberInofthe channels
processinoftheir outputsetting
parameter units was 512, 512 andlong
of bidirectional 6, respectively.
and short memory network, a batch
In the process of parameter setting of bidirectional long and short memory network, a batch
was set process
In the and 64 of images were setting
parameter input. The initial learning
of bidirectional long rate
and was
short0.001.
memoryThenetwork,
learning aratebatchwaswas
was set and 64 images were input. The initial learning rate was 0.001. The learning rate was
multiplied
setmultiplied
and 64 imagesby 0.1werewhen 24, The
input. 36 and 48 learning
initial epochs were iterated.
rate was 0.001.SGD The was usedrate
learning for was
the multiplied
gradient
by 0.1 when 24, 36 and 48 epochs were iterated. SGD was used for the gradient
calculation.
bycalculation.
0.1 when 24, Two
36 0.4 drop
and 48 out values
epochs were were added
iterated. SGD between
was the for
used whole
the connection
gradient layer to prevent
calculation. Two 0.4
Two 0.4 drop out values were added between the whole connection layer to prevent
over-fitting,
drop out values and
were the final output of six values represented the judgment scores of six categories.
over-fitting, and the added between
final output of the whole connection
six values representedlayer to preventscores
the judgment over-fitting, and the final
of six categories.
Soft-max
output and crossrepresented
entropy were used as loss scores
functions to conduct the gradient adjustment of the
Soft-max and cross entropy were used as loss functions to conduct the gradient adjustment ofentropy
of six values the judgment of six categories. Soft-max and cross the
network.
were used as loss functions to conduct the gradient adjustment of the network.
network.
2.2.2. PhysiologicalSignal
2.2.2. Signal Algorithm
2.2.2.Physiological
Physiological Signal Algorithm
Algorithm
After
After removing the signal from other states, the next step was to extract the physiological
Afterremoving
removingthe thesignal from
signal from other states,
other thethe
states, next step
next waswas
step to extract the physiological
to extract the physiological signal
signal from state C. The frequency of the respiratory signal and heartbeat signal had its special
from state C. The frequency of the respiratory signal and heartbeat signal had its
signal from state C. The frequency of the respiratory signal and heartbeat signal had its special special range, thus,
range, thus, wavelet transform was used to rebuild the signal at a particular frequency and filter
wavelet
range, transform
thus, wavelet wastransform
used to rebuild
was used thetosignal
rebuildat the
a particular
signal at frequency
a particularand filter noise
frequency and in other
filter
noise in other frequency bands. For normal people, the number of breaths and heart beats are 15–20
frequency bands.
noise in other For normal
frequency bands.people, the number
For normal of breaths
people, the numberand heart beats
of breaths are 15–20
and heart and15–20
beats are 60–100,
and 60–100, respectively. Thus, the frequency of respiratory was 0.15 Hz–0.5 Hz and the frequency
and 60–100,Thus,
respectively. respectively. Thus, the
the frequency frequency of
of respiratory respiratory
was wasHz
0.15 Hz–0.5 0.15andHz–0.5 Hz and the
the frequency frequency
of the heartbeat
of the heartbeat was 1 H–2 Hz. The contact devices, which were the pressure sensor for respiratory
of the
was 1 H–2heartbeat
Hz. The wascontact
1 H–2 devices,
Hz. The contact
which devices,
were thewhich weresensor
pressure the pressure sensor forand
for respiratory respiratory
heart rate
and heart rate monitoring was used with the non-contact monitoring system together. The
and heartwas
monitoring rateused
monitoring
with the was used with
non-contact the non-contact
monitoring system and monitoring
together. system together. The
waveform of two sensors and the reconstructed respiratory heartThe waveform
movement of two
signal fromsensors
the
waveform
andnon-contact of two
the reconstructed sensors and
respiratory the
andreconstructed
heart movementrespiratory
signal and
from heart
the movement
non-contact signal
system from
is the in
shown
system is shown in Figures 6 and 7.
non-contact
Figures 6 andsystem
7. is shown in Figures 6 and 7.
Amplitude/V
Amplitude/V

Time/s
Time/s
(a)
(a)
Amplitude
Amplitude

Time/s
Time/s
(b)
(b)
Figure 6. Respiratory signal. (a) The original signal from the contact band sensor; (b) Reconstructed
Figure6.6.Respiratory
Figure Respiratorysignal.
signal. (a) The original
originalsignal
signalfrom
fromthe
thecontact
contactband
band sensor;
sensor; (b)(b) Reconstructed
Reconstructed
respiratory signal.
respiratorysignal.
respiratory signal.
Amplitude/
Amplitude/

Time/s
Time/s
(a)
(a)
Figure 7. Cont.
 Biosensors 2019, 9, x FOR PEER REVIEW 8 of 13
Biosensors 2019, 9, 58 8 of 13
Biosensors 2019, 9, x FOR PEER REVIEW 8 of 13

Amplitude/V
Amplitude/V

Time/s

Time/s(b)
(b)
Figure 7. Heartbeat signal. (a) The original signal of the pulse sensor. (b) Reconstructed body
movement
Figure 7. signal caused
Heartbeat
Figure 7. Heartbeat signal.by
signal. heartbeat.
(a) The original signal of the
The original signal of the pulse
pulse sensor.
sensor.(b)(b)Reconstructed
Reconstructedbody
body
movement signal caused by heartbeat.
movement signal caused by heartbeat.
3. Experiment
3. 3.
Experiment
Experiment
The experimental scenario is shown in Figure 8. The experimental equipment included contact
The
type The experimental
experimental
breathing scenario
scenario
monitoring is shown
is shown
equipment in in
Figure an8. electronic
Figure
and The experimental
8. The experimentalequipment
equipment
sphygmomanometer, included
pulsecontact
included type
contact
sensor and
breathing
type monitoring
breathing equipment
monitoring and
equipment an electronic
and an sphygmomanometer,
electronic pulse
sphygmomanometer,
hardware used in the system. In the data collection of motion classification, the system had the sensor
pulse and hardware
sensor and
hardware
used in the
ability used insignals
system.
to classify Inthe
thesystem.
data In the
collection
collected data
from ofsix collection
motion of motion the
classification,
different situations classification,
in hadthe
system 2.2.
Section thesystem
The ability had the
to classify
participator sat in
ability
signals to classify
collected signals
from six collected
different from six
situations different
in Section situations
2.2. The in Section 2.2.
participator
front of the antenna 20 cm, 30 cm and 40 cm to sampling signals (except the “Unmanned satThe
in participator
front of the sat in
antenna
front of cm
theand
20environment”).
cm, 30 antenna
40 cm
Each of20to cm,
the 30 cmsignals
sampling
datapoints and
was40(except
cmseconds,
five to the
sampling signals
“Unmanned
and there (except the points
environment”).
was 300 data “Unmanned
Eachforofevery
the
environment”).
datapoints Each of the datapoints was five seconds, and there was 300 data points for every
situation. was five seconds, and there was 300 data points for every situation.
situation.

Figure 8. Theexperiment
experiment scene and
and devices (transceiver, serial port module and antenna).
Figure8.8.The
Figure The experimentscene
scene anddevices
devices(transceiver, serial
(transceiver, port
serial module
port and
module andantenna).
antenna).

There
There were
were 1010people
peopleinvolved
involvedin insignal
signal collection,
collection, including
including66males malesand and4 4females.
females. During
During thethe
There were 10 people involved in signal collection, including 6 males and 4 females. During the
verification
verification testtest of cardiac
of cardiac signalsignal
and and respiratory
respiratory signalsignal extraction,
extraction, the pressure
the pressure sensorsensor
of the of the
fingertip
verification test of cardiac signal and respiratory signal extraction, the pressure sensor of the
wasfingertip
tied towasthe tied to theoffingertip
fingertip the indexof the index
finger finger
of the of thetosubject
subject collecttopulse
collect pulse
data, anddata,
theand the
belt-type
fingertip was tied to the fingertip of the index finger of the subject to collect pulse data, and the
belt-type respiratory
respiratory sensor wassensortied towasthe tied to the abdomen
abdomen diaphragm diaphragm of thetosubject
of the subject collecttobreath
collectdata.
breath data.
belt-type
The respiratory
three parts sensor
of the was tiedhave
hardware to the abdomen
been zoomed diaphragm
in on as of the
shown in subject
Figure toThe
8. collect
two breath
circuits data.
The three parts of the hardware have been zoomed in on as shown in Figure 8. The two circuits
were The three
transceiver parts
and of the
serial hardware
portmodule. have
module.AAside been
sideview zoomed
viewof ofthe in on as
the experiment shown
experiment scene in
scene is Figure 8. The two circuits
were transceiver and serial port is also
also shown
shownin inFigure
Figure 8,
were
8, andtransceiver
the distance and serial port
between module.
human and A side view
antenna (d) wasof the experiment
changed from 20scene
cm to is100
alsocm shown
during inthe
Figure
and the distance between human and antenna (d) was changed from 20 cm to 100 cm during the
8,experiment.
and the distance between
The distance human
between and antenna
human and Doppler(d) was
radarchanged
ranged fromfrom 20 cm to
20 cm to 80
100 cmand
cm, during
two the
experiment. The distance between human and Doppler radar ranged from 20 cm to 80 cm, and two
experiment. The distance
sets of data were measuredbetween
every 10human
cm, with and Doppler
each radar ranged
measurement time offrom1 min.20Acm to of
total 8010cm, and two
people
sets of data were measured every 10 cm, with each measurement time of 1 min. A total of 10 people
sets of data were
participated measured
in signal everyincluding
collection, 10 cm, with six each
malesmeasurement
and four females. time of 1 min.
This A total different
is because of 10 people
participated in signal collection, including six males and four females. This is because different people
participated
people breathe in in
signal collection,
different including
ways, more testerssixcanmales and the
improve four females.
accuracy of This is because
the system. different
Attention
breathe in different ways, more testers can improve the accuracy of the system. Attention should be
people
should breathe
be paid to inthe
different
following ways, more
aspects testers
in the can improve the accuracy of the system. Attention
laboratory:
paid to the following aspects in the laboratory:
should be paid to the following aspects in the laboratory:
Biosensors 2019, 9, 58 9 of 13

(1) According to experimental physiology and electromagnetism, it could be known that others
will greatly
Biosensors 2019, 9,affect the accuracy
x FOR PEER REVIEW of the experiment when measuring the equipment. 9Therefore, of 13
the experimental environment with less affected should be a better option.
(2) (1)
When According
the systemto experimental
is turned on physiology
and off, and electromagnetism,
it will it couldwaveform
affect the generated be known to thatsome
others
extent.
will greatly affect the accuracy of the experiment when measuring the equipment.
In addition, when the system is filtered in the frequency domain, it will also affect the waveform at Therefore,
the experimental environment with less affected should be a better option.
the edge. Therefore, during the verification test of cardiac signal and respiratory signal extraction,
(2) When the system is turned on and off, it will affect the generated waveform to some extent. In
more than one minute of data should be collected in the experiment, and the middle one minute
addition, when the system is filtered in the frequency domain, it will also affect the waveform at
of data
the should be used for
edge. Therefore, calculation,
during so as to avoid
the verification test ofthe influence
cardiac of the
signal and above phenomenon
respiratory signal on
the experimental results.
extraction, more than one minute of data should be collected in the experiment, and the middle
(3) Try tooneensure
minutethe of diversity
data shouldof samples
be used forwhen collecting
calculation, so classified
as to avoidactions, that is,ofeach
the influence action has
the above
dataphenomenon
collected byon 10the experimental results.
subjects.
(4) (3)
In theTryverification
to ensure thetestdiversity of samples
of cardiac signalwhen collecting classified
and respiratory signal actions, that is,
extraction, tryeach action has
to ensure that the
data collected by 10 subjects.
range of each breath of the subject is the same. This verification can be generally judged by the
(4) In the verification test of cardiac signal and respiratory signal extraction, try to ensure that the
waveform form collected by the contact device.
range of each breath of the subject is the same. This verification can be generally judged by the
waveform form collected by the contact device.
4. Results and Discussion
4. Results and Discussion
4.1. Effective Distance
4.1.aEffective
As Distance
non-contact device for monitoring a physical signal, it was imperative to design an experiment
for distance testing, where
As a non-contact devicethe performance
for monitoring of each distance
a physical could
signal, be judged
it was by the
imperative energyanof the
to design
signal. The experiment
experiment involved
for distance testing,ten participants,
where including
the performance six males
of each distanceand fourbe
could females.
judged A byfingertip
the
energy
pressure of theand
sensor signal. The experiment
a breath sensor were involved
placedtenon
participants, including
the participant’s six males
body and four
together withfemales.
an antenna
whichA was
fingertip pressure
a part of thesensor
Dopplerand radar
a breath sensorlocated
system were placed on the
in front of participant’s body The
the participant. together withrange
length
an antenna which was a part of the Doppler radar system located in front
between them was 20 cm to 80 cm with seven equidistant collection points. The measurement of the participant. The time
length range between them was 20 cm to 80 cm with seven equidistant collection points. The
was 1 min for each experience, and every participant needed to do two experiments in one point.
measurement time was 1 min for each experience, and every participant needed to do two
After extracting the waveform of breath and heartbeat, the energy of the original signal, the breath
experiments in one point.
signal and the heartbeat
After signal
extracting the were calculated
waveform by the
of breath and formulathe
heartbeat, below:
energy of the original signal, the
breath signal and the heartbeat signal were calculated
n by the formula below:
X
E= S(i)2 (1)
E 𝑆i=𝑖1 (1)

The The
histogram
histogramdiagram
diagramofofenergy anddistance
energy and distancewaswasmademade according
according to thetoaverage
the average
energy energy
of 20 of
20 1-min
1-minantenna
antennareceiving waveformscollected
receiving waveforms collectedat at each
each position
position fromfrom 20 to
20 cm cm80tocm.
80 cm. Atsame
At the the same
time,time,
the respiration rate rate
the respiration and and
the heart rate rate
the heart werewere
extracted, andand
extracted, thethe
energy waswas
energy calculated and
calculated andis is
shown
in Figure 9. It can be seen that the energy received by the antenna dropped suddenly at
shown in Figure 9. It can be seen that the energy received by the antenna dropped suddenly at 40 cm and
cm
and
almost almost disappeared
disappeared at 60 cm. at However,
60 cm. However, the energy
the energy of theofcardiac
the cardiac signal
signal basically
basically disappearedatat50 cm.
disappeared
50 cm.the
Therefore, Therefore,
maximumthe maximum normal working
normal working distancedistance
of this of this antenna
antenna was about
was about 40 cm.
40 cm.

Figure 9. The
Figure relation
9. The between
relation betweenelectromagnetic waveenergy
electromagnetic wave energyand
and distance
distance received
received by antenna.
by antenna.
Biosensors 2019, 9, 58 10 of 13

4.2. Action Classification

When the SVM carried out 100 iterations, the accuracy was stable at around 80.72%. The confusion
matrix at this time is shown in Table 1. When using the deep learning network to classify six different
types of motion, a 4-fold cross-validation method was selected to verify the accuracy of the algorithm.
The idea of folded cross validations was to divide all the data into four equal parts and use one of
them as a test set to calculate the accuracy of the model, and the other three as a training set to train the
model. After adjusting the learning rate and the loss function in the training, the confusion matrix of
the classification calculation algorithm shown in Table 2 could be obtained.

Table 1. Support Vector Machine classification accuracy after genetic selection algorithm optimization.

A B C D E F
A 0.75 0.04 0.03 0.12 0.02 0.06
B 0.11 0.79 0.02 0.07 0.03 0.05
C 0.02 0.04 0.83 0.04 0.01 0.04
D 0.09 0.04 0.03 0.76 0 0.01
E 0.01 0.06 0.05 0.01 0.82 0.10
F 0.02 0.03 0.04 0 0.12 0.74

Table 2. Classification accuracy of the deep learning algorithm.

A B C D E F
A 0.91 0 0.01 0 0 0.04
B 0.01 0.96 0 0 0.01 0
C 0.08 0.02 0.99 0 0.01 0.01
D 0 0 0 1 0 0.01
E 0 0.02 0 0 0.76 0.30
F 0 0 0 0 0.22 0.64

It can be seen from Table 2 that the accuracy reached 88% when the deep learning network was
used to classify the states of movement. In addition to the low degree of differentiation between the
sudden changes of standing and sitting, the rest of the states could be basically judged accurately and
the misjudgment of standing and sitting often occurred in mutual discrimination. In this case, a simple
method to improve the accuracy of the algorithm was completed, that is, by judging the movements
before and after the standing and sitting of the algorithm, so as to correctly distinguish the standing
and sitting.
Figure 10 shows the comparison of the total accuracy of the two classification methods and the
accuracy of the six somatic classification methods. It can be seen that the classification method based
on SVM had an average accuracy of judgment for each action, indicating that the features extracted
manually mainly focused on the morphological features of waveform. The total accuracy of the deep
learning algorithm was relatively high, and the main reason that it affected the accuracy was that it
made more errors in the judgment of the interaction between body movements with a sudden change
in the states of standing and sitting. The accuracy of classifying them with other actions was 96.0%
showing in Table 2. One way to improve the classification algorithm is to use it in the real time system
which can record human’s previous movement state. For example, if there had already existed a
“standing up” the next “standing up” should have been “sitting down.”
Biosensors 2019, 9, x FOR PEER REVIEW 11 of 13
Biosensors 2019,9,
Biosensors 2019, 9,58
x FOR PEER REVIEW 11 of 13
11 13

Figure 10. Accuracy comparison of deep learning algorithm and Support Vector Machine algorithm
in RF signal
Figure
Figure 10. state classification.
10. Accuracy
Accuracy comparison
comparisonofofdeep
deeplearning
learningalgorithm
algorithmand
andSupport
SupportVector Machine
Vector algorithm
Machine in
algorithm
RF signal
in RF state
signal classification.
state classification.
Therefore, it can be concluded that the motion classification algorithm of deep learning is very
effective for theititaccurate
Therefore,
Therefore, can
can be judgment of
be concluded
concluded physiological
that
that the
the motion signals and the
motion classification
classification monitoring
algorithm
algorithm of and learning
of deep
deep classification
learning is of
is very
very
signals
effectiveoffor
effective body
for themovement.
the accurate
accurate judgment
judgment of of physiological
physiological signals
signals and
and the
the monitoring
monitoring and and classification
classification of of
signals of body movement.
signals of body movement.
4.3. Physiological Signal Extraction
4.3. Physiological Signal Extraction
4.3. Physiological
The absolute Signal
errorExtraction
in the number of the cycles between the reference signal and the
The absolute
reconstructed error in theonenumber of the wascycles between the reference signal and the reconstructed
The absolute error in theminute
signal within number ofselected
the cyclesto measure
between thethe
extracted
referenceaccurate
signalvalue,
and and the
signal
the within
absolute one
error minute
in averagewas selected
time to
between measure
the the
peaks extracted
within oneaccurate
minute, value,
which and the
could absolute
also be usederror
as
reconstructed signal within one minute was selected to measure the extracted accurate value, and
in
theaverage
absolute time between
index of the the peaksaccurate
average within one minute, which could also be used as the absolute index
difference.
the absolute error in average time between the peaks within one minute, which could also be used as
of theFigure
average accurate difference.of cycles extracted from 20 one-minute data at 20 cm, 30 cm and 40
the absolute11 shows
index of the
the number
average accurate difference.
cm in Figure
each 11 shows
position the
and number
the meanof cycles extracted
valueextracted from 20 error
of the absolute one-minute
of the data at 20 heart
average cm, 30ratecm and
Figure 11 shows the number of cycles from 20 one-minute data at 20 cm, 30 cm and 40
40 cm in each position
respiration of theand the mean value of thethat
absolute error of the average heart rate and the
respiration
cm in eachrate position Doppler
and radar
the mean system
value of thewas simultaneously
absolute error of monitored
the average using
heart contact
rate and
rate
heartof the
rate Doppler
and radar
respiration system that
measuring was simultaneously
equipment as the monitored
control group.using
As the contact
shown in heart
the rate
figure, and
the
respiration rate of the Doppler radar system that was simultaneously monitored using the contact
respiration
accuracy of measuring
the test equipment
decreased as the
with an control
increase group.
in the As shown in
distance. Thethemonitoring
figure, the accuracy
of of the rate
respiration test
heart rate and respiration measuring equipment as the control group. As shown in the figure, the
decreased
was moreof with
accuratean increase theinmonitoring
the distance. The
heartmonitoring of respiration rate was more accurate than
accuracy the testthan
decreased with an of increase rate at the
in the same
distance. distance. Moreover,
The monitoring the accuracy
of respiration of
rate
the monitoring
heart rate accurateof
monitoringheart rate at the same distance. Moreover, the accuracy of heart rate monitoring had
was more thanhadthe amonitoring
larger error of with
heart the
ratechange of distance.
at the same distance.According
Moreover,tothe the previous
accuracy of
aanalysis,
larger error
this with
is the change
caused by the of distance.
decrease of Accordingenergy
waveform to the previous
caused byanalysis,
distance this
and isthe
caused
low by the
energy
heart rate monitoring had a larger error with the change of distance. According to the previous
decrease
of of waveform
the received energy caused by distance and the low energy of the received cardiac signal.
analysis, this iscardiac
causedsignal.
by the decrease of waveform energy caused by distance and the low energy
of the received cardiac signal.

Figure 11. Error analysis of three different distances.

distances.
Figure 11. Error analysis of three different distances.
Biosensors 2019, 9, 58 12 of 13

5. Conclusions
The Doppler radar system proposed in this paper is suitable for classifying six different body
movement signals with the classification method of deep learning and can also monitor physiological
signals in a distance which is less than 40 cm. The accuracy of the classifier was 88% and could be
further improved in the system. While the respiratory waveform, cardiac waveform and contact
devices were compared at 20 cm, 30 cm and 40 cm respectively, and the errors of the respiratory cycle
number, average respiratory rate, cardiac cycle number and average heart rate at the corresponding
positions were obtained. Compared with the contact device, the average error of respiration rate in
the optimal measurement distance was about 0.8 beats/min, and the average error of the heart rate
was about 3.5 beats/min, which verified the accuracy of the algorithm. Compared with the existing
system, this system has a better performance in the classification of action state which is shown in
Table 3. What’s more, it has the ability to monitor the heartbeat. However, the effective distance for the
system is shorter than the previous one [16], this feature makes it more suitable in offices or cars.

Table 3. The comparison between two systems.

[16] Proposed System

Maximum distance 1m 0.4 m
Methods of classification SVM VGG-16 and LSTM
Classification accuracy 85% >88%
Error of respiratory detection 13% 0.779 beats/min
Error of Heart rate detection No experiment 3.49 beats/min

Author Contributions: Conceptualization, L.X. and J.S.; methodology, L.X. and Q.L.; hardware and software,
Q.L. and Y.Y.; validation, L.X., Q.L. and Y.Y.; resources, L.X.; data curation, Q.L. and Y.Y.; writing—original draft
preparation, Q.L.; writing—review and editing, L.X.; visualization, Q.L.; supervision, L.X.; project administration,
L.X.; funding acquisition, L.X.
Funding: The work here was, in part, supported by the National Natural Science Foundation of China
(No. 61773110, No. 613740156), and the Fundamental Research Funds for the Central Universities (No. N161904002,
No. N172008008). This work is also supported by the Open Program of Neusoft Research of Intelligent Healthcare
Technology, Co. Ltd., Item number NRIHTOP1801.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Lee, Y.S.; Pathirana, P.N.; Steinfort, C.L.; Caelli, T. Monitoring and Analysis of Respiratory Patterns Using
Microwave Doppler Radar. IEEE J. Transl. Eng. Health Med. 2014, 2, 1–12. [CrossRef] [PubMed]
2. Singh, L. Respiratory Monitoring and Clutter Rejection Using a CW Doppler Radar with Passive RF Tags.
IEEE Sens. J. 2012, 12, 558–565. [CrossRef]
3. Liu, Y.; Draper, S.C.; Sayeed, A.M. A Secret Key Generation System Based on Multipath Channel Randomness,
RSSI vs. CSSI. IEEE Trans. Inf. Forensics Secur. 2011, 3534, 1484–1497.
4. Li, H.; Ota, K.; Dong, M.; Guo, M. Learning Human Activities through Wi-Fi Channel State Information with
Multiple Access Points. IEEE Commun. Mag. 2018, 56, 124–129. [CrossRef]
5. Wang, X.; Gao, L.; Mao, S. CSI Phase Fingerprinting for Indoor Localization with a Deep Learning Approach.
IEEE Internet Things J. 2017, 3, 1113–1123. [CrossRef]
6. Wang, H.; Zhang, D.; Wang, Y.; Ma, J.; Wang, Y.; Li, S. RT-Fall, A Real-time and Contactless Fall Detection
System with Commodity WiFi Devices. IEEE Trans. Mob. Comput. 2017, 16, 511–526. [CrossRef]
7. Wu, C.; Yang, Z.; Zhou, Z.; Liu, X.; Liu, Y.; Cao, J. Non-Invasive Detection of Moving and Stationary Human
with WiFi. IEEE J. Sel. Areas Commun. 2015, 33, 2329–2342. [CrossRef]
8. Van Dorp, P.; Groen, F.C.A. Human Walking Estimation with Radar. IEE Proc. Radar Sonar Navig. 2003, 150,
356–358. [CrossRef]
9. Balleri, A.; Chetty, K.; Woodbridge, K. Classification of Personnel Targets by Acoustic micro-Doppler
signatures. IET Radar Sonar Navig. 2011, 5, 943–951. [CrossRef]
Biosensors 2019, 9, 58 13 of 13

10. Giovanni, D.; Alessandro, L.; Pietro, S. A Radar-Based Smart Sensor for Unobtrusive Elderly Monitoring in
Ambient Assisted Living Applications. Biosensors 2017, 7, 55.
11. Li, C.; Lin, J. Random Body Movement Cancellation in Doppler Radar Vital Sign Detection. IEEE Trans.
Microw. Theory Tech. 2008, 56, 3143–3152.
12. Munoz-Ferreras, J.M.; Peng, Z.; Gomez-Garcia, R.; Li, C. Random Body Movement Mitigation for
FMCW-radar-based Vital-sign Monitoring. In Proceedings of the 2016 IEEE Topical Conference on Biomedical
Wireless Technologies, Networks, and Sensing Systems (BioWireleSS), Austin, TX, USA, 24–27 January 2016;
pp. 22–24.
13. Tang, M.C.; Wang, F.K.; Horng, T.S. Single Self-Injection-Locked Radar with Two Antennas for Monitoring
Vital Signs with Large Body Movement Cancellation. IEEE Trans. Microw. Theory Tech. 2017, 65, 5324–5333.
[CrossRef]
14. Lv, Q.; An, K.; Chen, L.; Wang, J.; Li, H.; Ye, D.; Huangfu, J.; Li, C.; Ran, L. Doppler Vital Signs Detection in
the Presence of Large-Scale Random Body Movements. IEEE Trans. Microw. Theory Tech. 2018, 66, 4261–4270.
[CrossRef]
15. Faheem, K.; Sung, C. A Detailed Algorithm for Vital Sign Monitoring of a Stationary/Non-Stationary Human
through IR-UWB Radar. Sensors 2017, 17, 290.
16. Li, W.; Tan, B.; Piechocki, R. Passive Radar for Opportunistic Monitoring in E-Health Applications. IEEE J.
Transl. Eng. Health Med. 2018, 6, 1–10. [CrossRef] [PubMed]
17. Van Nguyen, T.P.; Tang, L.; Nguyen, D.M.; Hasan, F.; Mukhopadhyay, S. Wide Band Antennae System
for Remote Vital Signs Detecting Doppler Radar Sensor. In Modern Sensing Technologies; Springer:
Cham, Switzerland, 2019; pp. 47–62.
18. Park, Z.; Li, C.; Lin, J. A Broadband Microstrip Antenna with Improved Gain for Noncontact Vital Sign Radar
Detection. IEEE Antennas Wirel. Propag. Lett. 2009, 8, 939–942.
19. Liang, Q.; Mpanda, R.S.; Wang, X.; Shi, J.; Xu, L. A Printed Dipole Array Antenna for Non-contact Monitoring
System. In Proceedings of the 2018 Cross Strait Quad-Regional Radio Science and Wireless Technology
Conference (CSQRWC), Xuzhou, China, 21–24 July 2018; pp. 1–3.
20. Hu, X.; Jin, T. Short-range Vital Signs Sensing Based on EEMD and CWT Using IR-UWB Radar. Sensors 2016,
16, 2025. [CrossRef] [PubMed]
21. Shoaib, M.; Bosch, S.; Incel, O.D.; Scholten, H.; Havinga, P.J.M. Complex Human Activity Recognition Using
Smartphone and Wrist-Worn Motion Sensors. Sensors 2016, 16, 426. [CrossRef]
22. Zhang, M.; Sawchuk, A.A. A Feature Selection-based Framework for Human Activity Recognition Using
Wearable Multimodal Sensors. In Proceedings of the ACM Conference on Ubiquitous Computing,
Beijing, China, 17–21 September 2011; pp. 92–98.
23. Chavarriaga, R.; Millán, J.D.R. Unsupervised Adaptation for Acceleration-based Activity Recognition:
Robustness to Sensor Displacement and Rotation. Pers. Ubiquitous Comput. 2013, 17, 479–490. [CrossRef]
24. Sekine, M.; Tamura, T.; Akay, M.; Fujimoto, T.; Togawa, T.; Fukui, Y. Discrimination of Walking Patterns
Using Wavelet-based Fractal Analysis. IEEE Trans. Neural Syst. Rehabil. Eng. 2002, 10, 188–196. [CrossRef]
25. Hemalatha, C.S.; Vaidehi, V. Frequent Bit Pattern Mining Over Tri-axial Accelerometer Data Streams for
Recognizing Human Activities and Detecting Fall. Procedia Comput. Sci. 2013, 19, 56–63. [CrossRef]
26. Simonyan, K.; Zisserman, A. Very Deep Convolutional Networks for Large-Scale Image Recognition.
Comput. Sci. 2014, 4, 1–15.

© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).