CNN-based Respiration Rate Estimation in Indoor

Environments via MIMO FMCW Radar

Kohei Yamamoto Kentaroh Toyoda Tomoaki Ohtsuki
Graduate School of A*STAR, SIMTech, Singapore and Department of Information and
Science and Technology Faculty of Science and Technology Computer Science
Keio University Keio University Keio University,
Kanagawa 223-8522, Japan Kanagawa 223-8522, Japan Kanagawa 223-8522, Japan
Email: [email protected] Email: [email protected] Email: [email protected]

Abstract— Non-contact respiration rate estimation technique method, the respiration rate is estimated based on (i) RSS
in indoor environments is receiving more and more attention in (Received Signal Strength) [3], the CSI (Channel State In-
various fields, e.g., health care and smart home, since respiration formation) amplitude [4], and the CSI phase [5]. However,
is known to reflect our health condition. Hence, various radar-
based respiration rate estimation methods have been proposed the RSS and the CSI amplitude are not very sensitive to the
so far. However, these conventional methods do not work, when chest movements due to respiration [3][4]. In addition, the
a subject is not right in front of the radar. In this paper, we respiration rate estimation accuracies of these methods get
propose a novel CNN (Convolutional Neural Network)-based degraded, when a subject is not in LOS (Line of Sight) [3]-
respiration rate estimation method in indoor environments via [5]. On the other hand, A. Ahmad et al. have shown the
a MIMO (Multiple-Input Multiple-Output) FMCW (Frequency
Modulated Continuous Wave) radar. A MIMO FMCW radar feasibility of a MIMO FMCW radar-based respiration rate
can estimate DoA (Direction of Arrival) and the distance between estimation [6]. A MIMO FMCW radar can estimate DoA
a MIMO FMCW radar and an object. Thus, respiration can be (Direction of Arrival) and receive microwaves from the es-
captured based on the phase variation at a subject’s location. timated DoA by the beamforming technique. In addition, at
However, even when the advanced signal processing, e.g., MUSIC the estimated angle, a MIMO FMCW radar can estimate the
(MUltiple SIgnal Classification) algorithm, is used, it is difficult
to estimate DoA and the distance in indoor environments due distance between an object and a MIMO FMCW radar by
to the large effect of multipath. To deal with this problem, the analyzing the beamformed signal. Thus, when a subject is in
proposed method calculates spectrograms from phase variations the detectable range of a MIMO FMCW radar, the respiration
against various locations, and then estimates the respiration rate can be estimated by capturing the relative subject’s chest
rate by inputting each spectrogram into CNN that outputs the movements due to respiration at a subject’s location. However,
respiration rates, e.g., 0.1 Hz, 0.2 Hz, and non-respiration, i.e.,
a spectrogram without the effect of respiration. We observed in indoor environments, it is difficult to estimate DoA even
respiration in three situations where a subject was lying on his by leveraging the advanced DoA estimation technique such
(i) back, (ii) face, and (iii) side at various indoor locations. We as MUSIC (MUltiple Signal Classification) algorithm widely
confirmed that except for when microwaves were not transmit- used in the field of the DoA estimation. This is because the
ted directly toward a subject’s chest, our method accurately effect of multipath is large in indoor environments, compared
estimated the respiration rate, regardless of the situation.
with in outdoor ones, which also affects the estimation
accuracy of the distance between a subject and a MIMO
I. I NTRODUCTION
FMCW radar.
Respiration is one of major vital signs that reflect our In this paper, to estimate the respiration rate in indoor
health condition [1][2]. Hence, demands for the respiration environments with high accuracy, we propose a novel CNN
rate estimation in indoor environments have increased in (Convolutional Neural Network)-based respiration rate esti-
various fields, e.g., health care and smart home. To monitor mation method in indoor environments via a MIMO FMCW
our health condition in our daily life, it is better to estimate radar. In the proposed method, the phase variations against
the respiration rate without any wearable devices such as multiple locations specified with an angle and a distance
respiration belt, since the device attachment is not preferred from a MIMO FMCW radar are firstly calculated based on
in real application. In addition, it is demanded to estimate the received signals of a MIMO FMCW radar. STFT (Short
respiration rate even in indoor environments where a subject Time Fourier Transform) is then applied to the calculated
is not right in front of the measurement device. phase variations, which results in spectrograms. To detect a
These background has motivated researchers to develop spectrogram with the effect of the respiration and estimate the
the respiration rate estimation method with (i) Wi-Fi [3]-[5] respiration rate based on the detected spectrograms, CNN is
and (ii) a MIMO (Multiple-Input Multiple-Output) FMCW used. CNN is one of the deep learning techniques that have
(Frequency Modulated Continuous Wave) radar [6]. On the been successfully applied to the image recognition [7] and
one hand, in the Wi-Fi-based respiration rate estimation the activity recognition [8]. Specifically, in our method, the

978-1-7281-0962-6/19/$31.00 ©2019 IEEE

 Transmitted signal
radar sweeps the frequency of the transmitted signal linearly
Frequency [Hz]
𝑡" Received signal as shown in Fig. 1. Now, let fc be the minimum sweep
frequency, and let B and Tc be the bandwidth of the chirp
and sweep duration, respectively. In the FMCW radar system
𝐵
𝑓'
with one transmit antenna and one receive antenna, a signal
x(t) = A cos(2πfc t + π TBc t2 + ϕ(t)) is transmitted by the
𝑓$ Time [us]
transmit antenna, and the transmitted signal is reflected by
𝑇$ objects, e.g., a subject, furnitures, and walls. The reflected
signal r(t) is then received by the receive antenna.
Fig. 1. The transmitted and received signals of a MIMO FMCW radar.
B
r(t) = A{cos(2πfc (t − td ) + π (t − td )2
Tc (1)
+ϕ(t − td ))},
𝑑# sin 𝜃
Reflected signal
𝜃 where A is the power of the received signal, and td is given
as eq. (2).
𝑑$
𝑑# R(t)
𝑑"
td = 2 , (2)
・・・ Reference point
c
・・・
𝑦$(𝑡) 𝑦#(𝑡) 𝑦"(𝑡) where R(t) is the distance between the FMCW radar and
𝑤$ 𝑤# 𝑤" the object, and c is the speed of the electro-magnetic wave.
Subsequently, a quadrature mixer is applied to the received
signal, which results in two output signals with the phase
𝑦(𝑡) difference of π/2, i.e., in-phase and quadrature signals. These
two signals can be expressed as a complex signal y(t).
Fig. 2. K receive antennas arranged linearly.
y(t) = Aej(2πfb t+Φ(t)+∆ϕ(t)) , (3)
where fb , Φ(t), and ∆ϕ(t) are given as follows.
calculated spectrogram for each location is fed into the CNN 2BR(t)
that outputs labels: the respiration rates, e.g., 0.1 Hz, 0.2 Hz, fb = , (4)
cTc
and non-respiration, i.e., a spectrogram without the effect of
respiration. However, some respiration rates could be simul- B 2
taneously detected at several locations, since subject’s chest Φ(t) = 2πfc td + π t , (5)
Tc d
could be over some locations depending on each location size.
Thus, in our proposed method, by using the output of the ( )
2R
CNN as the weight, the respiration rate is estimated as the ∆ϕ(t) = ϕ(t) − ϕ t − . (6)
c
weighted average of the output respiration rates.
We conducted the experiments to observe the respiration Also, when a object is at Rconst away from the FMCW radar,
rate in three situations where a subject was lying on his R(t) can be also expressed approximately using Rconst and
(i) back, (ii) face, and (iii) side at various indoor locations. the relative movement of the object Rl , e.g., chest movements
Experimental results showed that except for when microwaves due to respiration, as
were not transmitted directly toward a subject’s chest due to R(t) ≈ Rconst + Rl . (7)
furniture, our proposed method estimated the respiration rate
with high accuracy, regardless of the situation. Based on the above equation, by specifying Rconst , a phase
The rest of this paper is organized as follows. In Section variation due to the relative displacement of the object at
II, we describe the principle of a MIMO FMCW radar. We Rconst can be obtained.
then explain some related work and our proposed method in In contrast, in the MIMO FMCW radar system where
Sections III and IV, respectively. In Section V, we evaluate the multiple transmit and receive antennas are linearly arranged,
performance of our method. Finally, we conclude this paper an additional phase shift occurs. Fig. 2 shows K receive
in Section VI. antennas arranged linearly. Here, let θ be the incident angle
of the reflected signal, the received signal at k th receiver,
II. P RINCIPLE OF MIMO FMCW R ADAR yk (t), is given as
In this section, for the better understanding of the proposed 2π
yk (t) = Aej(2πfb t+Φ(t)+∆ϕ(t)+ λ dk sin θ) , (8)
respiration rate estimation method in indoor environments via
a MIMO FMCW radar, we explain the principle of a MIMO where λ is the wavelength of the carrier, and dk is the
FMCW radar. Fig. 1 shows the transmitted and received relative distance between the receiver and the reference point
signals of a used MIMO FMCW radar. The MIMO FMCW as shown in Fig. 2. By calculating beamforming weights wk ,
the received signal against a specific angle can be calculated Received signals 𝑦"
as eq. (9).
Calculation of the phase variations against various locations 𝜑$ 𝑡
∑
K
Y (t) = yk (t)wk . (9) 𝜑& 𝑡 ・・・ 𝜑' 𝑡 ・・・ 𝜑$ 𝑡 ・・・ 𝜑() 𝑡
k=1
STFT STFT STFT STFT
Based on the above description, in the MIMO FMCW radar
system, a phase variation due to the relative movement of the Spectrograms
Pre-processing
object at a location can be extracted by specifying an angle
θ and Rconst . Offline training of CNN

Trained weights
A. Related Work Offline training

In this section, we explain researches on the respiration Online respiration rate estimation
rate estimation. To monitor the respiration rate in our daily
Respiration rate 𝑅𝑅+,-
life, it is highly demanded to estimate it without any wearable Online respiration rate estimation
devices, which motivates researchers to investigate the non-
contact respiration rate estimation with Wi-Fi [3]-[5] and a Fig. 3. The flowchart of the proposed method.
radar [6][9]-[14]. On the one hand, the Wi-Fi-based respi-
ration rate estimation methods have been realized with the
RSS [3], the CSI amplitude [4], and the CSI phase [5]. In FMCW radar because of the antenna directivity. To deal with
these methods, the respiration rate is estimated based on the this problem, the use of a MIMO FMCW radar has been also
peak detection and the frequency analysis of the RSS, the CSI investigated [6]. As aforementioned in Section II, a MIMO
amplitude, and the CSI phase. However, the RSS and the CSI FMCW radar can estimate DoA and receive microwaves
amplitude are not very sensitive to chest movements due to from the estimated DoA by the beamforming technique.
respiration [3][4]. Furthermore, when a subject is not in LOS, In addition, at the estimated angle, a MIMO FMCW radar
the respiration rate estimation accuracies of these methods get can estimate the distance between a subject and a MIMO
degraded [3]-[5]. FMCW radar by analyzing the beamformed signal. Thus,
On the other hand, the radar-based respiration rate es- once a subject’s location is estimated, the respiration rate can
timation has been extensively investigated with a Doppler be estimated based on the phase variation at the estimated
radar [9]-[11] and FMCW radar [12]-[14]. A Doppler radar location. However, in indoor environments, it is difficult
can observe target’s velocity and direction by transmitting to estimate DoA and the distance even by leveraging the
microwaves toward the target and then analyzing a reflected advanced DoA estimation technique such as MUSIC algo-
Doppler-shifted microwaves. Specifically, the Doppler radar- rithm. This is because the effect of the multipath is large
based respiration rate estimation methods can be classified in indoor environments. Therefore, it is necessary to develop
into two methods based on (i) the feature detection and a more accurate MIMO FMCW radar-based respiration rate
(ii) the frequency analysis. In the feature detection-based estimation method in indoor environments.
method, the respiration rate is estimated by detecting features
of the pre-processed signal, e.g., peaks [9] and zero-crossing III. P ROPOSED M ETHOD
points [10]. Also, in the frequency analysis-based method, In this section, we propose a novel CNN-based respiration
the respiration rate is estimated by detecting a peak due to rate estimation method in indoor environments via a MIMO
respiration over the spectrum calculated with the frequency FMCW radar. Fig. 3 shows the flowchart of the proposed
analysis [11]. In contrast, the FMCW radar-based respiration method. The proposed method consists of three steps: (i) the
rate estimation methods estimate the respiration rate based on pre-processing step, (ii) the offline training step, and (iii) the
the phase variation due to the relative chest displacement due online respiration rate estimation step. In the pre-processing
to respiration. Specifically, in the conventional methods [12]- step, phase variations φm against various locations are firstly
[14], the frequency analysis is firstly applied to estimate calculated by specifying θ and Rconst , where 1 ≤ m ≤ ML ,
the distance between a subject and a FMCW radar. The and ML denotes the number of the locations. STFT based
respiration rate is then estimated by (i) the feature detection on a 20 s-window with a 0.5 s-step size is subsequently
and (ii) the frequency analysis of the phase variation at the applied to φm , which results in ML spectrograms. Here, the
estimated distance. Through the experiments, these conven- window size and the step size are enough to analyze the
tional Doppler radar and FMCW radar-based respiration rate respiration periodicity. Each spectrogram is then segmented
estimation methods have been shown to provide the accurate by a 20 s-window with a 10 s-step size, and the segmented
respiration rate estimation [9]-[14]. However, these methods spectrogram within [0.1, 1] Hz, i.e., [6, 60] bpm (breath per
work in the situation where a subject is right in front of minute), is extracted as an input to the CNN, which results
a Doppler radar and FMCW radar, but not in the situation in a spectrogram with the size of 10 × 40 as shown in Figs. 4
where a subject is not right in front of a Doppler radar and and 5.
 Input spectrogram
32 Conv. 3×3 Dropout 32 Conv. 3×3
Pool. 1×3

Fully-connected layer 32 Conv. 2×2

32 Conv. 2×2

Fig. 4. Examples of the input spectrogram with the effect of respiration. Output

Pool. 2×2
Dropout Dropout

Fig. 6. The architecture of the CNN in the proposed method. “Conv.” and
“Pool.” denote the convolutional and max pooling operations, respectively.

Fig. 5. Examples of the input spectrogram without the effect of respiration.

where the pooling area slides without the overlap, meaning
that the dimension of the feature map in the layer l + 1 is
A. Offline training half of that in the layer l.
In this step, to estimate the respiration rate based on the In the fully-connected layer, the output data obtained
spectrogram with the effect of respiration, CNN is trained. through the convolutional and pooling operations are trained
CNN is one of the deep learning techniques that have been in a neural network. In this neural network, the difference
successfully applied to the image recognition [7] and the between the desired output of the true label and the out-
activity recognition [8], and is typically consists of three put of the CNN is calculated with a loss function. Here,
layers: (i) the convolutional layer, (ii) the pooling layer, and spectrograms are labeled with 11 labels in our method: 10
(iii) the fully-connected layer. Here, let xli,j be the element of possible respiration rates, i.e., 0.1 Hz, 0.2 Hz, . . . , 1.0 Hz,
the feature map with the size of M × N in the layer l, where and non-respiration, i.e., a spectrogram without the effect
i = 0, 1, . . . , M −1 and j = 0, 1, . . . , N −1. Also, let wp,q be of respiration. Specifically, by minimizing the values of
the element of the convolutional filter with the size of H ×H, the loss function with the back propagation algorithm, the
where p = 0, 1, . . . , H − 1 and q = 0, 1, . . . , H − 1. In the weights between units in adjacent layers are updated with the
convolutional layer, when the step size of the convolutional stochastic gradient descent method. Also, in our CNN, the
filter is denoted by s, the convolutional calculation is operated squared error loss function E is used as the loss function,
as and is given as
∑ H−1
H−1 ∑
1∑
N
ui,j = xsi+p,sj+q wp,q , (10)
p=0 q=0
E= (gn − on )2 , (13)
2 n=1
where the size of the matrix U with the elements ui,j is
where N is the number of training spectrograms, gn and on
(M − 2⌊H/2⌋) × (N − 2⌊H/2⌋). Also, though the training
are the desired output and the output of the CNN, respectively.
time can be reduced by increasing s, s is set as 1 in our
method, since the size of the input spectrogram is not so Fig. 6 shows the architecture of the CNN in the proposed
large. To enhance the feature extracted by the convolutional method. In our CNN, the convolutional operation is applied
operation, the activation function σ(·) is then used, which to each input spectrogram by 32 convolutional filters with
results in the element of the feature map in the next layer the size of 3 × 3. To prevent the over-learning, the dropout
xl+1
i,j as operation is then applied with the dropout rate of 25 %.
The convolutional operation is subsequently applied by 32
xl+1
i,j = σ(ui,j ), (11)
convolutional filters with the size of 3 × 3, the max pooling
where the activation function is ReLU (Rectified Linear Units) operation is applied with the pooling size 1 × 3, and the
in our method. dropout operation is applied with the dropout rate of 25 %.
The pooling layer reduces the resolution of the feature map The convolutional operation is then applied by 32 convolu-
by the downsampling algorithm, which brings the training tional filters with the size of 2 × 2, the dropout operation is
time reduction. In our method, the max pooling is used as applied with the dropout rate of 25 %, and the convolutional
the downsampling algorithm. Now, let P be the pooling area operation is applied by 32 convolutional filters with the size
over the feature map in the layer l, the max pooling in P is of 2 × 2. After the max pooling operation is applied with
operated as the pooling size 2 × 2, the weights between units in adjacent
layers are updated in the fully-connected layer, and finally the
xl+1 l
i,j = max xp,q , (12) output is obtained.
(p,q)∈P
 4
Lying on his back
Lying on his face
Lying on his side

Average AE [bpm]
3

5
n

n
tio

tio

tio

tio

tio
ca

ca

ca

ca

ca
Lo

Lo

Lo

Lo

Lo
Fig. 7. The experimental environment for the testing data collection. the
shaded area denotes the detectable area of a MIMO FMCW radar.
Fig. 8. Average AE against 2 subjects.

B. Online respiration rate estimation

subject was lying on his (i) back, (ii) face, and (iii) side at five
In this step, the respiration rate is estimated based on the
locations as shown in Fig. 7. The distance between the MIMO
output data of the CNN. Although the output label of the
FMCW radar and the subject ranged from 1.2 m to 2.6 m.
CNN is the respiration rate, some respiration rates could
The height of the MIMO FMCW radar was about 1 m, and
be detected simultaneously, since a subject could be over
the detectable area was as shown in Fig. 7. The observation
some locations depending on each location size. Thus, in
was performed on 2 subjects, and the observation duration
our proposed method, the respiration rate is estimated as the
ranged from 70 s to 80 s for each situation. In contrast, in
weighted average of the output respiration rate. Specifically,
the experiment for the training data collection, the respiration
by using the output of the CNN, od , as the weight, the
rate was observed from 4 subjects in same three situations.
respiration rate RR est is estimated as
The experimental room was different from that in the other
∑D
od zd experiment, and had no furniture. The distance between the
RR est = ∑d=1 D
, (14) MIMO FMCW radar and the subject ranged from 1.2 m to
d=1 od
2.5 m, and the height of the MIMO FMCW radar was about
where D is the number of the output respiration rate labels, 1 m. In the both experiments, to obtain the ground truth value
and zd is the output respiration rate label. of the respiration rate, the subjects were asked to synchronize
IV. E XPERIMENTAL E VALUATION their breaths with the bell, breathe naturally, and count their
breaths by theirselves.
A. Experimental specification
In the experiments, θ in the proposed method was specified
To evaluate the estimation accuracy of the proposed with 9 angles: −40◦ , −30◦ , . . . , and 40◦ . Rconst was also
method, we conducted experiments to observe the respiration specified with 20 ranges: 1.1 m, 1.2 m, . . . , and 3.0 m.
rate in two rooms for the testing and training data collection. Thus, the number of locations used for phase variation
As the performance metric, the AE (Absolute Error) between extraction, NL , was 180.
the estimated respiration rate and the ground truth value of
the respiration rate is calculated as eq. (15). B. Results
1 ∑
V Through the experiments, as the testing data, we collected
AE = |RR est (a) − RR ref (a)|, (15) 205 and 335 spectrograms with and without the effect of
Nr a=1 respiration, respectively. In contrast, as the training data, we
where Nr denotes the number of the respiration rates esti- collected 300 and 500 spectrograms with and without the
mated in each observation, RR est (a) and RR ref (a) denote effect of respiration, respectively. Fig. 8 shows the average
the a th estimated respiration rate and the ground truth value AE against 2 subjects. As can be seen from this figure, our
of the respiration rate, respectively. proposed method estimates the respiration rate with the high
In this experiments, the numbers of transmit and receive estimation accuracy at the locations 1, 2, 4, and 5, regardless
antennas of a MIMO FMCW were 2 and 4, respectively, of the situation; Our method achieves the average AEs of
and the distance between the adjacent receive antennas was 0.34 bpm, 0.72 bpm, 1.18 bpm, and 0.50 bpm at the locations
λ/2. The center frequency of the MIMO FMCW radar was 1, 2, 4, and 5, respectively. Considering that the conventional
24.15 GHz, the bandwidth of the chirp B and the sweep radar-based respiration rate estimation methods do not work
duration Tc were 2 GHz and 1,024 ms, respectively. The in the situation where a subject is not right in front of
sampling rate was 1.8 MHz, and the antenna directivity the radar, it is worth noting that our method estimates the
ranged from −45◦ to 45◦ . Fig. 7 shows the experimental respiration rate with such estimation accuracy. From Fig. 8,
environment for the testing data collection. In this experiment, it can be also seen that the average AEs at the location 3 get
the respiration rate was observed in three situations where a degraded, compared with the other locations. This is because
 TABLE I TABLE II
T HE CLASSIFICATION RESULTS OF THE INPUT SPECTROGRAMS . T HE CLASSIFICATION RESULTS OF 180 SPECTROGRAMS COLLECTED IN
THE UNINHABITED ROOM .
True label \ Classified label With respiration Without respiration
Method \ Classified label With respiration Without respiration
With respiration 205 0
Without respiration 16 319 With training 2 178

when the subject is at the location 3, the subject’s chest the output respiration rates. Through the experiments in three
is covered by the table, which makes it difficult to capture situations where a subject was lying on his (i) back, (ii) face,
the phase variation due to respiration. Also, TABLE I lists and (iii) side, we confirmed that except for when microwaves
the classification results of the input spectrograms. In this were not transmitted directly toward the subject’s chest, our
table, the labels “With respiration” and “Without respiration” method estimated the respiration rate with the average AE of
denote the cases where a spectrogram is classified into the 0.74, regardless of the situation. In our future work, to realize
one with and without the effect of respiration, respectively. As a more accurate respiration estimation, it is necessary to
can be seen from this table, our method accurately classifies improve the classification accuracy of a spectrogram without
a spectrogram with the effect of respiration. In contrast, in the effect of respiration in the uninhabited room.
terms of the classification of a spectrogram without the effect
R EFERENCES
of respiration, some spectrograms are classified incorrectly.
This is because microwaves reflected by the chest could be [1] M. A. Cretikos, R. Bellomo, K. Hillman, J. Chen, S. Finfer, and A.
Flabouris, “Respiratory rate: The neglected vital sign,” Med. J. Austral.,
received by the receive antennas after the microwaves are vol. 188, no. 11, pp. 657-659, 2008.
reflected again by other objects, e.g., wall and furniture, but [2] D. J. Cooper and M. D. Buist, “Vitalness of vital signs, and medical
it is worth mentioning that such incorrect classification does emergency teams,” Med. J. Austral., vol. 188, no. 11, pp. 630-631,
2008.
not degrade the estimation accuracy of the respiration rate. [3] H. Abdelnasser, K. A. Harras, and M.Youssef,
In addition to the above evaluation, to clarify whether “UbiBreathe:AUbiquitous non-Invasive WiFi-based Breathing
the proposed method classifies spectrograms accurately in Estimator”, ACM MobiHoc, 2015.
[4] X. Liu, J. Cao, S. Tang, J. Wen, and P. Guo, “Contactless Respiration
the uninhabited room, we collected 360 spectrograms in the Monitoring Via Off-the-Shelf WiFi Devices,” IEEE Trans. Mobile
uninhabited room for the testing data collection, and then Computing, vol. 15, no. 10, pp. 2466-2479, Oct. 2016.
evaluated our method based on 180 testing and 180 training [5] X. Wang, C. Yang and S. Mao, “PhaseBeat: Exploiting CSI phase data
for vital sign monitoring with commodity WiFi devices,” in IEEE Int.
data, respectively. Here, note that the CNN was trained with Conf. on Distributed Computing Systems (ICDCS), 2017.
not only the 180 training data but also the training data [6] A. Ahmad, J. C. Roh, D. Wang, and A. Dubey, “Vital signs monitoring
used in the above evaluation, i.e., 300 and 500 spectrograms of multiple people using a FMCW millimeter-wave sensor,” IEEE Int.
Conf. on Radar Conference (RadarConf), pp. 1450-1455, Apr. 2018.
with and without the effect of respiration. TABLE II lists [7] A. Krizhevsky, I. Sutskever, and G. Hinton,“ImageNet classification
the classification results of the spectrograms collected in the with deep convolutional neural networks,” Int. conf. on Advances in
uninhabited room. From this table, it can be seen that our neural information processing systems, pp. 1106-1114, Dec. 2012.
[8] M. Zeng, L. T. Nguyen, B. Yu, O. J. Mengshoel, J. Zhu, P. Wu, and J.
method provides the accurate classification of the spectro- Zhang, “Convolutional neural networks for human activity recognition
grams collected in the uninhabited room. However, we can using mobile sensors,” Int. Conf. on Mobile Computing, Applications
also say that two spectrograms are classified incorrectly, and and Services, pp. 197-205, Nov. 2014.
[9] A. Rahman, LV. M. ubecke, O. Boric-Lubecke, J. H. Prins, and
hence it is necessary to improve our method that it does not T. Sakamoto, “Doppler Radar Techniques for Accurate Respiration
detect respiration in the uninhabited room. Characterization and Subject Identification,” IEEE Journal on Emerging
and Selected Topics in Circuits and Systems, vol. 8, pp. 350-359, Feb.
V. C ONCLUSION 2018.
[10] W. Hu , Z. Zhao, Y. Wang, H. Zhang, and F. Lin, “Noncontact accurate
In this paper, we proposed a novel CNN (Convolutional measurement of cardiopulmonary activity using a compact quadrature
Doppler radar sensor,” IEEE Trans. Biomedical Engineering, vol. 61,
Neural Network)-based respiration rate estimation method in no. 3, pp. 725-735, Mar. 2014.
indoor environments via a MIMO (Multiple-Input Multiple- [11] A. Dell’Aversano, A. Natale, A., A. Buonanno, and R. Solimene,
Output) FMCW (Frequency Modulated Continuous Wave) “Through the wall breathing detection by means of a doppler radar
and MUSIC algorithm,” IEEE sensors letters, vol. 1, pp. 1-4, June.
radar. In our method, phase variations against various loca- 2017.
tions are calculated from the received signals of a MIMO [12] K. Van Loon, M. J. M. Breteler, L. Van Wolfwinkel, A. R. Leyssius,
FMCW radar. Each spectrogram calculated from the phase S. Kossen, C. J. Kalkman, and L. M. Peelen, “Wireless non-invasive
continuous respiratory monitoring with FMCW radar: a clinical vali-
variation is then fed into the CNN that outputs the labels: dation study,” Journal of clinical monitoring and computing, vol. 30,
the respiration rates and non-respiration, i.e., a spectrogram pp. 797-805, June 2016.
without the effect of respiration. However, some respiration [13] M. He, N. Yongjian, and G. Yushun, “Novel signal processing method
for vital sign monitoring using FMCW radar,” Biomedical Signal
rates could be detected simultaneously, since a subject could Processing and Control vol. 33, pp. 335-345, 2017.
be over some locations depending on each location size. Thus, [14] A. Prat, S. Blanch, A. Aguasca, J. Romeu, and A. Broquetas, “Col-
in our method, by using the output of the CNN as the weight, limated beam FMCW radar for vital sign patient monitoring,” IEEE
Transactions on Antennas and Propagation, Dec. 2018.
the respiration rate is estimated as the weighted average of