Digital Signal Processing 74 (2018) 72–93

Contents lists available at ScienceDirect

Digital Signal Processing

[Link]/locate/dsp

Improved denoising method for through-wall vital sign detection using

UWB impulse radar
Xiaolin Liang a,b,∗ , Hao Zhang a,c , Shengbo Ye b , Guangyou Fang b , T. Aaron Gulliver c
a
Department of Electronic Engineering, Ocean University of China, Laoshan District Song Ling Road 238th, Qing Dao, People’s Republic of China
b
The Key Laboratory of Electromagnetic Radiation and Sensing Technology, Institute of Electronics, Chinese Academy of Science, Haidian District North 4th Ring
West Road 19th, Beijing, People’s Republic of China
c
Department of Electrical Computer Engineering, University of Victoria, 3800 Finnerty Road, Victoria, BC V8P 5C2, Canada

a r t i c l e i n f o a b s t r a c t

Article history: Noncontact vital sign detection is widely used in finding victims in post-disaster search and rescue,
Available online 21 December 2017 through-wall surveillance (TWS), and medical diagnosis and monitoring. Human subject can be remotely
sensed by extracting life activities such as respiration and heartbeat. However, the signal to noise
Keywords:
ratio (SNR) is often low, particularly in complex environments, which results in errors in both range
Vital sign
Impulse ultra-wideband radar (UWB)
and respiration frequency (RF) estimation. To improve the accuracy, an improved system for the vital
Respiration frequency (RF) sign detection is presented which is based on impulse ultra-wideband (UWB) radar. The range is
Short-time Fourier transform (STFT) determined based on the short-time Fourier transform (STFT) of the standard deviation of the received
Ensemble empirical mode decomposition signals. Further, the ensemble empirical mode decomposition (EEMD) based frequency accumulation (FA)
(EEMD) technique is presented to determine RF. Performance results are presented which show that the proposed
Frequency accumulation (FA) method is superior to several well-known techniques.
© 2017 Elsevier Inc. All rights reserved.

1. Introduction ments such as long range and through-wall conditions. Further,

they do not consider a variety of signal characteristics such as sta-
Detecting survivors trapped in ruins after natural disasters such tionary clutter, non-stationary clutter, respiration frequency (RF),
as earthquakes is an important research problem based on impulse and heartbeat rate (HR). The time–frequency characteristics of hu-
ultra-wideband (UWB) radar recently [1–5]. Impulse UWB radar man respiration were analyzed based on the fast Fourier transform
has been employed in many areas such as vital sign detection [6, (FFT) and Hilbert–Huang transform (HHT) in [29,30]. In [32], a
7], moving subject detection [8,9], through-wall imaging [10,11], dual frequency model was designed and the respiration-like clut-
rescue [12,13], indoor subject localization [14,15], and public secu- ter was suppressed using an adaptive clutter cancellation method.
rity [16] due to its excellent performance in penetrating through The method for weak signal enhancement is developed in [33] to
obstacles such as walls and building ruins [17–20]. In particular, improve SNR in the radar system. The RF was estimated in [34]
it can be used to obtain heartbeat and respiration signals. How- using a low complexity maximum likelihood estimator considering
ever, this is a challenging task as thorax motion is typically only the additive white Gaussian noise (AWGN) environment. The sin-
a few millimeters and there can be severe signal attenuation. Sev- gular value decomposition (SVD) method was employed in [35] to
detect respiration signals in low signal to noise and clutter ratio
eral methods have been proposed to detect vital signs [21–25], but
(SNCR) conditions. In [40], a tracing method was used to extract
they have poor performance over long ranges. As a result, further
vital signs. However, it is effective only in short range and high
research is needed to effectively employ UWB radar for vital sign
signal to noise ratio (SNR) conditions. The linear trend subtraction
detection [26–28].
(LTS) was employed in [41] to suppress non-stationary clutter and
Many vital sign detection techniques have appeared in the lit-
the linear trend. In [43], the vital signs were extracted using the
eratures [29–47], but most are not effective in complex environ-
multiple higher order cumulant (HOC) method based on the fact
that the higher order cumulants of Gaussian noise are zero. The
ensemble empirical mode decomposition (EEMD) technique was
*Corresponding author at: Department of Electronic Engineering, Ocean Univer-
used in [44] to detect the human HR [44]. The arctangent de-
sity of China, Laoshan District Song Ling Road 238th, Qing Dao, People’s Republic of
China. modulation method was used in [46] to improve the accuracy of
E-mail address: iamxiaolin2016@[Link] (X. Liang). human HR estimation [46]. In [47], a state-space method (SSM)

[Link]
1051-2004/© 2017 Elsevier Inc. All rights reserved.
 X. Liang et al. / Digital Signal Processing 74 (2018) 72–93 73

was employed to detect human respiration and heartbeat, but the

performance is poor in long range and through-wall conditions.
In this paper, a new method is proposed to accurately esti-
mate vital signs in through-wall and long range conditions. It is
designed to remove the stationary and non-stationary clutter, lin-
ear trend, and other clutter that may exist. Further, the products
of the respiration and heartbeat signals, as well as harmonics,
are suppressed using an ensemble empirical mode decomposition
(EEMD) based frequency accumulation (FA) technique. Results are
presented which show that this method outperforms several well-
known algorithms in the literatures. A new approach to determin-
ing the range of a subject is presented which is based on the Fig. 1. The received human respiration signal.
statistical characteristics of the respiration and heartbeat signals.
The short-time Fourier transform (STFT) of the standard deviation
Equation (4) can be expressed in discrete form as [43]
is employed. The performance of the proposed detection method is
  
evaluated using experimental results obtained with the UWB radar R [m, n] = a v s mδ T − τ v (nT s ) + ai s(mδ T − τi )
designed at the Key Laboratory of Electromagnetic Radiation and
i
Sensing Technology, Institute of Electronics, Chinese Academy of   
Sciences. = a v s mδ R − v τ v (nT s ) + ai s(mδ R − v τi /2)
The remainder of this paper is organized as follows. Section 2 i
presents the mathematical model for vital signs detection based = h[m, n] + c [m], (5)
on impulse UWB radar. The proposed detection method is given
where the pulse repetition frequency (PRF) is 1/ T s , t = nT s , n =
in Section 3. Section 4 presents the approach to experimental data
0, . . . , N − 1, δ T is the sampling interval in fast time, N is the
acquisition. The performance of this method is evaluated and com-
number of fast time samples, δ R = v δ T /2 is the sampling inter-
pared with several well-known techniques in Section 5. Finally,
val in slow time, m = 0, . . . , M − 1, M is the number of slow time
some conclusions are given in Section 6.
samples, h[m, n] is the respiration and heartbeat signals, and c [m]
is the stationary clutter reflected by static targets.
2. Mathematical model of vital sign
To avoid the ambiguity in the range and frequency aliasing, T s
must satisfy [43]
The vital signs can be detected based on changes in the prop-  
agation delay of the UWB radar signal [43]. The distance between 1/ T s ≥ 2 max( f r , f h ) . (6)
the antenna and human chest can be expressed as [39]
In a real detection environment, the received signal can be ex-
pressed as [39]
d(t ) = d0 + r (t ) = d0 + A r sin(2π f r t ) + A h sin(2π f h t ), (1)
R [m, n] = h[m, n] + c [m] + a[m, n] + w [m, n]
where d0 is the distance between the antenna and the center of
the chest, f r is the RF, A r is the respiration amplitude, f h is the HR, + q[m, n] + g [m, n] + z[m, n] (7)
and A h is the heartbeat amplitude. If only one subject is present, where a[m, n] is the linear trend introduced by the instability of
and all other objects are stationary, the impulse response is [39] the time base, w [m, n] is AWGN, q[m, n] is non-stationary clutter
   reflected by non-static targets, g [m, n] is unknown clutter caused
h(τ , t ) = a v δ τ − τ v (t ) + ai δ(τ − τi ), (2) by the radar receiver, and z[m, n] is clutter introduced by moving
i objects with larger amplitude than life activities. The signal is ob-
where ai δ(τ − τi ) is the response of the ith stationary object with tained within one time duration [43]
amplitude ai and propagation delay τi , a v δ(τ − τ v (t )) is the re-    
T w + max τ v (t ) − min τ v (t ) < T s , (8)
sponse due to human respiration and heartbeat with amplitude a v
and propagation delay τ v (t ), which is given by [43] where T w is the −6 dB pulse width.
In a stationary environment, the received signal is given by (4).
2d(t ) The stationary clutter can be suppressed by subtracting the mean
τ v (t ) = = τ0 + τr sin(2π f r t ) + τh sin(2π f h t ), (3)
of (4). The signal without stationary clutter is
v
 
where v = 3 × 108 m /s is the speed of light, τ0 = 2d0 / v, τr = R (τ , t ) = a v s τ − τ v (t ) . (9)
2 A r / v, and τh = 2 A h / v.
The RF and HR can be obtained using the Fourier transform (FT)
The received signal is given by [39,43]
of R (mδ T , t ) in slow time which gives
  
R (τ , t ) = s(τ ) ∗ h(t , τ ) = a v s τ − τ v (t ) + ai s(τ − τi ), (4) +∞
i Y (mδ T , f ) = R (mδ T , t )e − j2π f t dt , (10)
where s(τ ) is the transmitted pulse. −∞
To explain fast time and slow time clearly, the sketch map of which can be obtained using the two-dimension FT of R (mδ T , t ),
the reflected pulses only with one respiration movement and no Y (υ , f ) which gives
stationary targets is shown in Fig. 1. The fast time bin is given
by the dashed line. The location of human chest changes due to +∞
human respiration, as a result the time delay of the reflected pulse Y (mδ T , f ) = Y (υ , f )e j2π υτ dυ , (11)
changes. −∞
74 X. Liang et al. / Digital Signal Processing 74 (2018) 72–93

Fig. 2. The resulting matrix (a) without clutter; (b) under AWGN.

where υ denotes FT of the fast time and +∞

C k0 = S (υ ) J k (βr υ ) J 0 (βh υ )dυ (20)
+∞ +∞
−∞
Y (υ , f ) = R (mδ T , t )e − j2π f t e − j2π υτ dtdτ , (12)
when βr f c  1
−∞ −∞
+∞ +∞
− j2π f t − j2π υτ v (t )
Y (υ , f ) = a v S (υ )e e dt , C k0 ≈ S (υ ) J k (βr υ )dυ , (21)
−∞ −∞
+∞ where f c is the center frequency of the transmitted signal. The
− j2π υτ0
= a v S (υ )e e − j2π υ mb sin(2π f r t ) harmonics of the heartbeat signal can then be obtained by setting
−∞
k = 0 to obtain

×e − j2π υ mh sin(2π f h t ) − j2π f t

e dt , (13) +∞ +∞
C 0l = S (υ ) J 0 (βb υ ) J l (βh υ )dυ ≈ S (υ ) J l (βh υ )dυ (22)
where S (υ ) is the FT of the transmitted pulses. Using Bessel func-
−∞ −∞
tions, (13) can be represented as
when βh f c  1.
+∞ 
+∞
Y (υ , f ) = a v S (υ )e − j2π υτ0 J k (βr υ )e − j2π kf r t +∞
k=−∞
C 0l ≈ S (υ ) J l (βh υ )dυ . (23)
−∞
 +∞ −∞

× J l (βh υ )e − j2π lf b t e − j2π f t dt , (14) In this paper, the goal is to estimate the range of a human sub-
l=−∞ ject and the corresponding RF. However, as shown in (6), Gaussian
+∞ noise, non-stationary clutter, linear trend, harmonics, and other

e − jz sin(2π f 0 t ) = J k ( z)e − j2π kf 0 t , (15) clutter will exist in a real environment. Thus, extracting this in-
formation is difficult, particularly in a complex environment such
k=−∞
as after a natural disaster. Fig. 2(a) shows the ideal resulting ma-
where βr = 2π A r and βh = 2π A h . Then (10) can be expressed as trix, while the received pulses under AWGN is shown in Fig. 2(b).
+∞ 
+∞ It can be seen that it is challenging to extract vital signs in low

Y (mδ T , f ) = a v G kl (τ )δ( f − kf r − lf h ), (16) SNR.
k=−∞ l=−∞
3. Detection algorithm
where
+∞ In this section, the proposed detection algorithm is presented
j2π υ (τ −τ0 ) including the clutter suppression, SNR and vital sign improvement,
G kl (τ ) = S (υ ) J k (βr υ ) J l (βh υ )e dυ . (17)
range and RF estimations. This algorithm has following main steps:
−∞

The maximum value of (17) is (1) stationary clutter, non-stationary clutter, and linear trend sup-
pression;
+∞ (2) SNR improvement via filtering in fast time and slow time;
C kl = G kl (τ0 ) = S (υ ) J k (βr υ ) J l (βh υ )dυ (18) (3) range estimation using a STFT on the characteristics of the
−∞ data in slow time;
+∞ +∞ (4) SNR improvement via slow time windowing and EEMD-based
  FA in the frequency domain;
Y (τ0 , f ) = a v C kl δ( f − kf r − lf h ). (19)
(5) RF estimation.
k=−∞ l=−∞

From (19), the respiration harmonics can be obtained by setting The flow chart of the proposed human subject detection
l = 0 as method showing the individual steps is shown in Fig. 3.
 X. Liang et al. / Digital Signal Processing 74 (2018) 72–93 75

3.1.3. Non-stationary clutter suppression

In high SNR conditions, human respiration and heartbeat sig-
nals can clearly be determined in the absence of non-stationary
clutter. However, the non-stationary clutter q[m, n] can signifi-
cantly degrade signal estimation. In this paper, the SVD is used to
reduce the effects of q[m, n]. With this method, W is decomposed
into orthonormal matrices given by [35]


N
W = U Σ VT = ui σii vTi (30)
i =1

where U = [u1 , u2 , . . . , u M ] is a unitary matrix with columns

uk ∈ C M called the left singular vectors, V = [v1 , v2 , . . . , v N ] is
a unitary matrix with columns vk ∈ C N which are called right
singular vectors, and a matrix Σ ∈ R M × N with nonzero values
σii ∈ {σ11 ≥ σ22 ≥, . . . , ≥ σN N }. Since there are N degrees of free-
dom in W, rank (W) = N. The energy in q[m, n]W M × N reduces
the SNR. Usually, this energy can be suppressed by removing the
largest values σii , i = 1, . . . , n < N, which gives [35]


N
 = U Σ VT = ui σii vTi (31)
i =n

Fig. 3. Flowchart of the proposed detection algorithm. 3.2. SNR improvement

3.1. Clutter suppression Filtering is now employed in slow time and fast time to im-
prove the SNR. Meanwhile, to improve vital signs, one advanced
3.1.1. Stationary clutter suppression gain method is performed on the fast time.
The stationary clutter in slow time is due to the rubble, and can
be approximated as a DC component. The time mean subtraction 3.2.1. Filtering in fast time
(TMS) method is employed to estimate this component as [42] The received signal depends on the detection environment in-
cluding the frequency and azimuth between the subject and the
1 
M 
N
radar antenna. Further, the electromagnetic wave transfer func-
= R [m, n] (24)
M×N tion depends on factors such as the dielectric constant, humidity,
m =1 n =1
polarization, and obstacles, which typically cannot be estimated
and the result after subtraction is accurately. As a result, a matched filter is not suitable for signal
detection [43].
 M × N =  M × N − . (25)
In this paper, the infinite impulse response (IIR) band-pass fil-
The stationary clutter c [m] can have a significant amplitude. Af- ter (BPF) is used to high and low frequency interference of the
ter (25), another effective and straightforward method to reduce acquired radar echo. This filter is used with  M × N in fast time.
this clutter is range profile subtraction (RPS) which takes the dif- The difference equation of the IIR filter is given by [48]
ference of the adjacent samples [42]
Λ[m, n]
Ω[m, n] = Ω[m, n] − Ω[m, n − 1]. (26)
To reduce the dependence on previous samples, the adaptive
= χ1 Φ[m, n] + χ2 Φ[m − 1, n] + · · · + χ Nb +1 Φ[m − N b , n]
background subtraction (ABS) is employed which uses a weighting − κ2 Φ[m − 1, n] − · · · − κ Na +1 Φ[m − N a , n], (32)
coefficient λ giving [42]
where N b = N a is the filter order, κi and χi denote the filter coef-
p [m, n] = λ p [m, n − 1] + (1 − λ)Ω[m, n] (27) ficients.
The Butterworth filter has been widely used in radar signal pro-
so the resulting matrix is [42]
cessing. In frequency domain, one N f order low-pass analogical
T [m, n] = Ω[m, n] − p [m, n] (28) Butterworth filter obeys the follow function [49,50]
where p [m, 1] = Ω[m, 1] and λ = 0.9. 2 1
H ( j ω) = , (33)
3.1.2. Linear trend suppression 1 + (ω/ωc )2N f
In a real detection environment, the acquired data is affected where | H ( j ω)|2 denotes the square magnitude of the frequency
√
by factors such as time jitter and drift introduced by the triggering response of the filter, j = −1, the cutoff frequency ωc is given
unit in the radar system. This results in amplitude fluctuations i.e. by
linear trend in slow time. To remove the linear trend a[m, n], the
ωc
LTS method is employed which gives [41] ωnc = (34)
 − 1 π fs
T T T T
W=T −X X X X T , (29)
where f s = 33 GHz is the sampling frequency in fast time, ωnc is
where X = [x1 , x2 ], x1 = [0, 1, . . . , N − 1] , and x2 = [1, 1, . . . , 1]TN .
T
the normalized cut-off frequency [50].
76 X. Liang et al. / Digital Signal Processing 74 (2018) 72–93

The IIR-BPF used in this paper is composed of one N f = 5 order

low-pass analogical Butterworth filter with the normalized cutoff
frequency 0.1037 and one N f = 5 order high-pass analogical But-
terworth filter with the normalized cutoff frequency 0.0222.

3.2.2. Filtering in slow time

To further improve the SNR, an average extraction filter is used
to reduce the remaining high and low frequency clutter. The out-
put of this filter is
+1)−1
β(k
1
S [k, n] = Λ[m, n], (35)
β
m=β k

where k = 1, . . . , M /β , β = 7 is the filter coefficient, M /β rep-

Fig. 4. The SD values obtained from the data set at a distance of 9 m from the
resents the maximum integer less M /β , and  M × N = 0 when antenna.
M > M /β .
For fast time index m, the SD of (39) is given by [53]
3.2.3. Advance normalization
The advance normalization method is performed on the matrix N
n=1 (Ψ [m, n] − μ)2
S M × N in fast time to enhance the vital signs given by Γ [m] = (40)
N −1
S [i , n ]
Ψ [i , n ] = , i = 1, . . . , τmax (0), (36) where μ is the corresponding mean value.
S [τmax (0), n] The data with one volunteer served as the detection subject at
where S [τmax (0), n] is the first maximum of S [i , n], i = 1, . . . , M. a distance of 9 m from the antenna in through-wall conditions was
If τmax (0) < M, then used to calculate the SD results given in Fig. 4. The detection en-
vironment will be described in detail in the next section. These
S [i , n ]
Ψ [i , n ] = , i = τmax (0) + 1, . . . , τmax (1), (37) results show that the SD in the subject region is much larger than
S [τmax (1), n] that elsewhere and satisfies the periodicity approximately [54],
where S [τmax (1), n] is the second maximum of S [i , n], which provides a basis for vital sign estimation.
i = τmax (0) + 1, . . . , M. Fig. 5 presents the calculated SD values in the target area
This process continues until τmax (k) = M. To explain this and the corresponding spectrum using the Fourier transform. This
method in more detail, the pseudo-code is as follows [33] shows that the SD is periodic approximately based on the spec-
trum using the FFT method as shown in Fig. 5 [54], and indicates
k=0 that the range of human subject can be estimated using the SD
while τmax (k) < M values.
  To estimate the range, a STFT is performed on  to obtain the
τmax (k) = arg max S(m, n)
τmax (k−1)<τ <τend τmax (k−1)<m<τend range (fast time)–frequency characteristics [55,56] given by
 
V max = max S(m, n) 
M
τmax (k−1)<m<τend K [o, p ] = Γ [m]Ξ [o − m]e− j2p π m/P (41)
  S(τmax (k − 1) : τmax (k), n) m =1
 τmax (k − 1) : τmax (k), n =
V max
where p denotes frequency, and Ξ is the Hamming window given
k=k+1 by
end (38)  
2π o
Ξ (o) = α − β cos , o = 0, 1, . . . , O , (42)
3.3. Range determination O
where α = 0.54 and β = 0.46.
Using (24)–(38), the stationary clutter c [m], linear trend a[m, n],
The critical parameter in the STFT-based spectral analysis is
non-stationary clutter q[m, n], and other high and low frequency the window length [57]. There will be excessive bias if it is too
clutter g [m, n] are suppressed. Further, z[m, n] can be eliminated small, while the result tends to zero based on the central limit
by ensuring no one moves in the detection environment. The re- theorem (CLT) as the length increases. Thus, in this paper the win-
sulting signal is then ideally given by dow length is set to O = 512 [58]. As shown in Fig. 6, for the
  obtained time–frequency matrix, the range axis corresponds to the
Ψ [m, n] = a v s mδT − τ v (nT s ) fast time while the frequency axis corresponds to the 1/2 sampling
 
= a v s mδ R − v τ v (nT s ) = h[m, n] (39) frequency of the fast time. The SD and corresponding STFT without
a human subject are shown in Fig. 7.
The standard deviation (SD) has been proposed for random mo- Comparing Figs. 6 and 7(b), it is obvious that the STFT can
tion of the person, not for person’s respiration motion detection in
be used for range estimation. As shown in Fig. 6, in the range-
the continuous-wave radar [51]. In this subsection, a new approach
frequency matrix, the signal amplitude in the target area is
to determining the range of a subject is presented via analyzing
stronger than other areas, which provide the basis for range es-
the statistical characteristics i.e. SD of the respiration and heartbeat
timate, i.e. time of arrival of the UWB pulse reflected by human
signals. For the first time, SD is used to deal with the detection
of respiratory motion of the person in the UWB pulse radar, and subject estimate. The range estimate is
the range between the UWB radar and human subject can be esti- 
τ
mated by performing the STFT method on the SD values [52]. L=v× (43)
2
 X. Liang et al. / Digital Signal Processing 74 (2018) 72–93 77

Fig. 5. (a) The time–domain waveform of the SD in the subject area, and (b) the corresponding FFT.


 2
O [i ] = Ψ [i , n ] . (45)
i =−100

3.4.1. Ensemble empirical mode decomposition

As an adaptive method, EEMD has been widely used to analyze
the non-stationary and non-linear signals [59], which overcomes
the drawbacks existing in Empirical Mode Decomposition (EMD)
such as the oscillations [60]. Using EEMD, a non-linear signal can
be decomposed into a series of intrinsic mode functions (IMFs)
in a descending order of frequency and a residual trend mode by
adding Gaussian white noise adaptively [61]. The IMFs and residual
trend mode can be used to represent the noise which can then be
deleted, resulting in an improved SNR [62]. To calculate the IMFs
Fig. 6. The STFT results for a human subject at a distance of 9 m from the antenna. of O1× N , the procedures for EEMD are summarized as

where τ represents the time delay, the range estimate can be cal- I) Add noise signals to O1× N ;
culated using (43) i.e. the maximum value obtained from (41) as II) Decompose O1× N into IMFs using EMD as
shown in Fig. 6. Based on (43), the STFT result has a range of 1) Let v = 0 with the index v is used to indicate the vth IMFs;
9.104 m, which has an error of only 0.104 m. 2) To identify all these local extrema values including both lo-
cal maximum and minimum values of the original signal
3.4. Respiration frequency estimation O1× N ;
3) To generate the upper (lower) envelope r u (t ) (rl (t )) of O1× N
In this section, human RF is acquired based on the proposed with a cubic spline function based on the local maximum
EEMD based FA method. (minimum) values;
The index of the fast time estimate 
τ in (39) is 4) To determine the average envelope based on r u (t ) and rl (t ),
which can be expressed as
=
τ /δT (44)
To acquire RF estimation, the accumulation in time domain is r u (t ) + rl (t )
m(t ) = (46)
used to reduce the harmonics given by 2

Fig. 7. The results without a human subject (a) the SD values, and (b) the corresponding STFT.
78 X. Liang et al. / Digital Signal Processing 74 (2018) 72–93

5) To subtract out m(t ) from the original signal O1× N , which 

5

is expressed as À1× N = IMF mean

v (t ) (54)
v =4
h(t ) = O1× N − m(t ) (47)
3.4.2. Frequency accumulation
6) Go to step (2), the h(t ) is processed until an IMF satisfies To improve the SNCR, a window over the frequencies in the
the stoppage criteria; range 0.08 to 0.8 Hz was employed. A rectangular window of
7) v = v + 1, IMF v (t ) = h(t ) and the residue signal is given by length κ is applied in the frequency domain on each slow time
dimension which gives
q v (t ) = O1× N − IMF v (t ) (48)
 
8) Go to step (2) and the residue signal is processed. When [n] = χ [n] FFT À [n] , n = k ∗ , k ∗ + 1, . . . , k ∗ + κ − 1
the amplitude of IMF v (t ) is small enough, to stop the whole
(55)
decomposition.
where FFT{À (n)} is the FFT of À (n), and k∗ is the minimum value
Using EMD, the time domain signal O1× N can be decomposed of n.
into a serious of time domain functions i.e. IMFs, which represent To suppress the clutter which remains in the same frequency
the local variability of the original signal at a particular range of band as the desired signals, the FA method is proposed based on
frequencies, and the residue signal as the technique discrete analysis [65]. The result is


Mv
à[n] = l[n] + jl[n] (56)
O1× N = IMF v (t ) + q M v (t ) (49)
v =1 where
⎧
III) For each different noise signals, O1× N is processed using steps ⎨ 2[n], κ > 0
⎪
(I)–(II) repeatedly; l[n] = [n], κ =0 (57)
⎪
⎩
IV) To obtain the average ensemble values of the IMFs as the final 0, κ <0
result, which is given by
4. Experimental data acquisition

Mv
O1× N = IMF mean
v (t ) + qmean
M v (t ) (50)
4.1. Radar system
v =1

where The UWB impulse radar used in this paper for data acquisition
was designed in the Key Laboratory of Electromagnetic Radiation
1 
Nv
IMF mean (t ) = IMF v (t ) (51) and Sensing Technology, Chinese Academy of Sciences. This radar
v
Nv contains two antennas in a .45 × .22 × .45 m3 box and is operated
v =1
via a wireless personal digital assistant (PDA).
1 
Nv
Table 1 shows the radar parameters used for data acquisition. It
qmean
M v (t ) = q v (t ) (52)
Nv operates at a center frequency of 400 MHz with a PRF of 600 kHz.
v =1
The data sets were obtained simultaneously by six segments with
and N v stands for the times of the added noises. a segment time window of 124 ns and M s = 682 samples per seg-
ment. The number of samples in fast time is M = 4092. To improve
For EEMD algorithm, the variance and times of the added the SNR, N A samples were averaged during data acquisition, so
AWGN are the two key parameters which need to be determined the pulse signals are saved every M s N A /PRF = 0.0341 s. In slow
[63]. When the variance of the added AWGN is less than the am- time, N = 512 pulses are received every 17.6 s. A hybrid sampling
plitude of O1× N , the AWGN cannot achieve better performance on scheme combining equivalent-time sampling [66] and real-time
determining the local extrema values. But, when the variance of [67] sampling was implemented as it outperforms an analog re-
the added AWGN is larger than the amplitude of O1× N , the big- ceiver employing only equivalent-time sampling. Fig. 10 shows a
ger signal errors will be caused [44]. As usual, the relationship typical received signal using the UWB radar.
between the variance of the AWGN and the amplitude of O1× N
4.2. Experimental setup
can be estimated as
ε To validate the performance of the developed method com-
εn = √ (53)
Nv pared with several reference algorithms including the direct FFT
method [30], the constant false alarm rate (CFAR) [39], the ad-
where ε denotes the amplitude of the AWGN, εn is the error be-
vanced method (AM) [42], and the multiple higher order cumulant
tween O1× N and the accumulation IMFs after EEMD.
(HOC) method [43], a series of experiments were conducted and
The residual trend is usually considered as noise, several IMFs
carried out at the Institute of Electronics, Chinese Academy of Sci-
can be used to reconstruct the vital signs. By employing the FFT,
ences in Beijing and the China National Fire Equipment Quality Su-
human RF can be acquired. As mentioned in [64], the RF is typi-
pervision Centre i.e. the national life detector identification centre
cally in the range 0.2 to 0.5 Hz with the amplitude of 5 to 15 mm, in Shanghai. The experimental setup is shown in Fig. 11(a) and (b).
and the heartbeat frequency is in the range 0.8 to 2.5 Hz with The first experiment was conducted outdoors at the Institute
an amplitude of 2 to 3 mm. The time–frequency characteristics of of Electronics, Chinese Academy of Sciences with a male volunteer
O1× N were shown in Fig. 8, while Fig. 9 showed welch power spec- served as human subject as shown in Fig. 11(c). The volunteer was
tral density of the IMFs. Consequently, the IMFs with the power surrounded by several kinds of swaying plants and flowers behind
concentrated on the range 0.08 to 0.8 Hz were used to reconstruct him, which makes vital sign challenging to extract. The volunteer
the vital signs given by stood behind the wall at distances of 3 m, 6 m, 9 m, 11 m, and
 X. Liang et al. / Digital Signal Processing 74 (2018) 72–93 79

Fig. 8. The time–frequency characteristics using EEMD: (a) the original signal; (b) IMF1 ; (c) IMF2 ; (d) IMF3 ; (e) IMF4 ; (f) IMF5 ; (g) IMF6 ; (h) IMF7 ; (i) IMF8 ; (j) IMF9 ;
(k) IMF10 .

16 m from the radar antenna facing the antenna. The wall consists at distances of 4 m, 7 m, 10 m, 12 m, and 18 m facing the antenna
of three common building materials, .3 m of brick, .35 m of con- in the China National Fire Equipment Quality Supervision Center
crete, and .35 m of pine, for a total thickness of 1 m. The radar located in Shanghai as shown in Fig. 11(d).
was installed on a table 1.5 m above the ground. The third experiment was carried out indoors at the China
The second experiment was carried out indoors with another National Fire Equipment Quality Supervision Center using the ac-
male volunteer served as human subject standing behind the wall tuator as shown in Fig. 11(e) to imitate human respiration. The
80 X. Liang et al. / Digital Signal Processing 74 (2018) 72–93

Fig. 9. The welch power spectral density of the IMFs: (a) IMF1 ; (b) IMF2 ; (c) IMF3 ; (d) IMF4 ; (e) IMF5 ; (f) IMF6 ; (g) IMF7 ; (h) IMF8 ; (i) IMF9 ; (j) IMF10 .

actuator was placed on a desk 1.3 m above the ground. The dis- tenna at China National Fire Equipment Quality Supervision Center,
tance between the actuator and antenna was 4 m, 7 m, 10 m, and as shown in Fig. 11(b). The distance between the actuator and the
12 m. The actuator signal has the amplitude of 3 mm and a motion antenna is 6 m. The beam angle of the antenna in the radar system
frequency (MF) of 0.3333 Hz. is 60◦ .
In the fourth experiment, the actuator was placed on a desk In the fifth experiment, the actuator stood at an azimuth an-
at different azimuth angles of 30◦ and 60◦ with respect to the an- gle of 0◦ between the antenna and the actuator at the Institute
 X. Liang et al. / Digital Signal Processing 74 (2018) 72–93 81

Table 1 of Electronics, Chinese Academy of Sciences. The distance between

Parameters for the UWB impulse radar. the human subject and the antenna is 11 m.
Parameter Value
Center frequency 400 MHz
5. Results
Transmitted signal amplitude 50 V
Pulse Repeat Frequency (PRF) 600 kHz
Number of averaged values (N A ) 30
Time window 124 ns
5.1. Intuitive detection performance
Number of samples (M) 4092
Input bandwidth of the Analog to Digital Converter (ADC) 2.3 GHz
ADC sampling rate 500 MHz The detection performance of the steps for clutter suppression
ADC sample size 12 bits is discussed in this section based on the data from the first ex-
Receiver dynamic range 72 dB periment with the human subject at a distance of 6 m from the
antenna.
Fig. 12(b) shows the result matrix after performing TMS method
stated in Fig. 3 based on the acquired experimental data given in
Fig. 12(a). As shown in Figs. 12(c)–(d), the stationary clutter can be
removed using the ABS method, and the linear trend is suppressed
by employing the LTS method. It is obvious that the respiration sig-
nals are too weak to be identified due to the non-stationary clutter.
To show the excellent performance of the developed method, the
SVD method is employed on the acquired range-frequency data as
shown in Fig. 12(d). By employing the SVD method, the acquired
result matrix is shown in Fig. 12(e). It can be seen that the respira-
tion signal is easy to see the oscillations compared with Fig. 12(d).
The result matrixes after filtering in fast and slow time are given in
Figs. 12(f) and (g), respectively. As shown in Fig. 12(h), the respira-
tion signals are improved and the noises are suppressed effectively.
Obviously, it can be seen that the vital sign signal is improved
Fig. 10. A normalized received signal obtained using the UWB radar. compared with Fig. 12(e).

Fig. 11. The experimental setup for data acquisition: (a) the experiment model for range detection; (b) the experiment model for angle detection; (c) through-wall detection
outdoors; (d) through-wall detection indoors; and (e) the actuator experiment.
82 X. Liang et al. / Digital Signal Processing 74 (2018) 72–93

Fig. 12. The resulting matrix (a) original signal; (b) performing TMS method; (c) removing stationary clutter; (d) performing LTS method; (e) performing SVD method;
(f) filtering in fast time; (g) filtering in the slow time; (h) performing the advanced gain method.

5.2. Detection performance with different distances outdoors where  represents the frequency index containing the target.
The calculated SD using the acquired data sets at different dis-
In this section, the data obtained for human subject at the In- tances is given in Fig. 13. Results indicate that the SD varies signifi-
stitute of Electronics is used to evaluate the detection performance. cantly in the target area compared to the non-target area. The STFT
was used on the SD values to obtain the range estimates given
The capability of improving SNR of the developed method is ana-
in Fig. 14, and the corresponding RF estimations are given in Ta-
lyzed by employing the data acquired at different distances from 4
ble 2. The errors in these estimates are 0.15 m, 0.15 m, 0.17 m,
to 11 m outdoors quantitatively. By employing the CFAR method,
and 0.24 m, respectively. Table 2 presents the RF estimates, range
AM, MHOC, and the new method, both the range and RF of the
errors, and SNRs for other three detection methods including the
human subject are detected. Typically, the SNR decreases with in- CFAR, AM, and MHOC method. The corresponding RF estimations
creasing distance between the antenna and subject due to large are respectively 0.31 Hz, 0.24 Hz, 0.27 Hz, and 0.22 Hz using the
scale attenuation [43]. Consequently, the detection precision and proposed method, which are quite similar, and much closer than
the detectability of human subject will be improved when the SNR with the other techniques. These results show that the proposed
is improved. As mentioned in [43], the SNR can be estimated as method provides superior performance. In addition, it results in
the highest SNR improvements such as the SNR based on the data
 +1 
n=−1 |Ψ [, n]|
τ sets at 11 m is 3.25 dB using the proposed method, while it is
SNRresult = 20. log10 −2 N −15.26 dB using the CFAR method, which has a 18.51 dB im-
n=1 |Ψ [, n]| +
τ n=+2 |Ψ [, n]| τ
provement. Fig. 15 presents the detection results using the CFAR
(58) method. This indicates that the range can only be estimated over
 X. Liang et al. / Digital Signal Processing 74 (2018) 72–93 83

Fig. 13. The SD calculated outdoors at (a) 3 m; (b) 6 m; (c) 9 m; and (d) 11 m.

Fig. 14. The range estimations using STFT at (a) 3 m; (b) 6 m; (c) 9 m; and (d) 11 m.
84 X. Liang et al. / Digital Signal Processing 74 (2018) 72–93

Fig. 15. The range estimations using the CFAR method at (a) 3 m; (b) 6 m; (c) 9 m; and (d) 11 m.

Fig. 16. The SD acquired indoors at (a) 4 m; (b) 7 m; (c) 10 m; and (d) 12 m.
 X. Liang et al. / Digital Signal Processing 74 (2018) 72–93 85

Table 2 Table 3
Detection performance using different methods. RF estimations with different methods.

Method 3m 6m 9m 11 m Method 4m 7m 10 m 12 m
CFAR Error (m) 0.27 4.36 6.72 9.54 CFAR Error (m) 0.32 4.79 5.83 9.51
RF (Hz) 0.18 0.1 0.72 0.46 RF (Hz) 0.18 0.18 0.72 0.46
SNR (dB) −4.54 −7.22 −11.86 −15.26 SNR (dB) −2.54 −5.22 −9.86 −11.26

Proposed Error (m) 0.15 0.15 0.17 0.24 Proposed Error (m) 0.06 0.06 0.11 0.2
RF (Hz) 0.31 0.24 0.27 0.22 RF (Hz) 0.32 0.26 0.29 0.26
SNR (dB) 7.58 4.32 3.67 3.25 SNR (dB) 6.58 3.52 3.27 1.35

MHOC Error (m) 0.35 2.43 1.56 7.25

RF (Hz) 0.14 0.45 0.52 0.44
SNR (dB) −2.67 −5.85 −8.58 −11.35 5.5. Detection performance with different azimuth angles

AM Error (m) 0.24 5.46 4.67 3.98

In this section, the influence on the detection performance is
RF (Hz) 0.37 0.12 0.74 0.63
SNR (dB) 5.35 0.84 −3.69 −6.59 validated due to the different azimuth angles between the actuator
and the radar antenna. The actuator is placed on a desk behind the
wall at a distance of 6 m from the antenna. The detection perfor-
mance with an azimuth angle of 0◦ was analyzed in the previous
short distances. As shown in Figs. 15(b), (c), and (d), significant section. The SD with an azimuth angle of 30◦ between the actuator
clutter remains for distances of 6 m or more. As a result, the pro- and the radar antenna is shown in Fig. 25(a), while the results with
posed method outperforms the CFAR method especially in long an azimuth angle of 60◦ is shown in Fig. 25(c). The corresponding
distance and through-wall conditions. Compared with the three spectrums by employing the STFT of the SD with an azimuth an-
reference methods, the developed method can provide the most gle of 30◦ are shown in Fig. 25(b), while the spectrums with an
accurate range and RF estimations. azimuth angle of 60◦ are given in Fig. 25(d).
As mentioned above, the respiration signal is the most reliable
5.3. Detection performance with different distances indoors basis for human subject detection. As shown in Fig. 25, results
indicate that the largest signal amplitude can be acquired with
In this section, the data sets obtained from human subject at an azimuth angle of 0◦ . However, using the new method, it can
the China National Fire Equipment Quality Supervision Center are be seen that the actuator with weak amplitude can be identified
used to evaluate the performance. accurately even when the azimuth angle is 60◦ between the an-
The SD using the acquired data sets at different distances is tenna and the actuator. The beam angle of the antenna in the radar
given in Fig. 16. The estimated ranges obtained using the STFT system is 60◦ . As a result, the developed method can provide ex-
technique are shown in Fig. 17, and the corresponding RF esti- cellent detection performance.
mations are given in Fig. 18 and Table 3. The errors in the range
estimates with increasing distance are 0.06 m, 0.06 m, 0.11 m, and 5.6. Detection performances with long range
0.2 m, and the corresponding RF estimates are 0.32 Hz, 0.26 Hz,
0.29 Hz, and 0.26 Hz, respectively. Table 3 also gives the esti- The influences of human subject on long range including 16 m
mations for the CFAR method. Results show that the developed indoors and 18 m outdoors in through-wall conditions are studied.
method provides superior performance. The calculated SD using the data sets acquired from human subject
at 18 m indoors is given in Fig. 26(a), while the corresponding
spectrum by employing the STFT is shown in Fig. 26(b). Fig. 26(c)
5.4. Detection performance with actuator
shows the SD using the data sets acquired at 16 m outdoors, while
the estimated range using the STFT is given in Fig. 26(d). The errors
In this section, the data from the actuator experiments are used in the range estimates are 0.22 m, and 0.32 m, respectively. It can
to evaluate the proposed method. be seen that the reference methods including the CFAR, AM, and
The SD values obtained using the acquired data sets indoors MHOC method can only identify human subject in short range as
at different distances are given in Fig. 19, and the corresponding mentioned above. Consequently, there are reasons to believe that
range estimations obtained using the STFT technique are shown the developed method outperform the methods mentioned above
in Fig. 20. The MF estimations are given in Fig. 21 and Table 4. even in long range.
As shown in Fig. 20, the range between the actuator and radar
antenna can be acquired accurately. The MF estimations using 5.7. Clutter suppression
the proposed method are 0.3208 Hz, 0.3223 Hz, 0.3271 Hz, and
0.3283 Hz with the deviations 3.75%, 3.30%, 1.86%, and 1.50%, In this section, clutter suppression is evaluated using the out-
which are less than the estimations using the MHOC method as door data from the first experiment with a distance of 6 m be-
shown in Fig. 22, the AM method as shown in Fig. 23, and the tween the subject and antenna. The FA method is used to cancel
FFT method as shown in Table 4. As shown in Figs. 22–23, the red out the harmonics and respiration and heartbeat signal products.
stars denote the estimations using the reference methods, while The results with FA method performed different numbers of times
the green rectangles denote the true range and frequency values. are given in Fig. 27. This shows that the four FA method can im-
It can be seen that there are notable errors between the estima- prove the SNR better than with one or two times. Table 5 gives
tions and real values. Furthermore, the data set acquired outdoors the corresponding improvement in SNR, and indicates that the
at a distance of 11 m from the antenna is used to further validate SNR improvement is minimal when the FA method is performed
the performance of the new method. The estimated range is shown more than four times. This table also provides the SNR with the
in Fig. 24(a) with the acquired SD given in Fig. 24(b). The MF for FFT method with the corresponding frequency domain result is in
the FFT method is shown in Fig. 24(c), while for the new method Fig. 27(a). As a result, the FA method is performed four times to
is given in Fig. 24(d). Based on all these detection results, it is ob- eliminate harmonics in this paper. This shows that these meth-
vious that the developed method provides superior performance. ods are not effective in suppressing the clutter compared to the
86 X. Liang et al. / Digital Signal Processing 74 (2018) 72–93

Fig. 17. The range estimations using STFT in indoor environment at (a) 4 m; (b) 7 m; (c) 10 m; and (d) 12 m.

Fig. 18. RF estimations using the new method at (a) 4 m; (b) 7 m; (c) 10 m; and (d) 12 m.
 X. Liang et al. / Digital Signal Processing 74 (2018) 72–93 87

Fig. 19. The SD acquired indoors with the actuator at (a) 4 m; (b) 7 m; (c) 10 m; and (d) 12 m.

Fig. 20. The estimated ranges of the actuator using the new method at (a) 4 m; (b) 7 m; (c) 10 m; and (d) 12 m.
88 X. Liang et al. / Digital Signal Processing 74 (2018) 72–93

Fig. 21. The MF estimations using the new proposed method at (a) 4 m; (b) 7 m; (c) 10 m; and 12 m.

Fig. 22. The detection results using MHOC at (a) 4 m; (b) 7 m; (c) 10 m; and (d) 12 m.
 X. Liang et al. / Digital Signal Processing 74 (2018) 72–93 89

Fig. 23. The detection results using AM at (a) 7 m; (b) 10 m; and (c) 12 m.

Table 4
Comparison of MF estimation techniques.

Method 4m 7m 10 m 12 m
MF (Hz) Deviation (%) MF (Hz) Deviation (%) MF (Hz) Deviation (%) MF (Hz) Deviation (%)
MHOC 0.3491 4.74 0.1164 65.10 0.1162 65.14 0.0870 73.89
AM 0.1870 43.89 0.0870 73.89 0.0870 73.89 0.0870 73.89
Proposed 0.3208 3.75 0.3223 3.30 0.3271 1.86 0.3283 1.50
FFT 0.1160 65.20 0.1267 62.00 0.1378 58.66 0.1126 66.22

Fig. 24. The detection results with the data at 11 m (a) the range estimation; (b) the SD; (c) the MF estimation for FFT method; and (d) the MF estimation with the new
method.
90 X. Liang et al. / Digital Signal Processing 74 (2018) 72–93

Fig. 25. (a) The SD with 30◦ azimuth angle; (b) the spectrum using STFT with 30◦ azimuth angle; (c) the SD with 60◦ azimuth angle; (d) the spectrum using STFT with 60◦
azimuth angle; and (e) the MF estimation.

Fig. 26. (a) The SD indoors at 18 m; (b) the spectrum using STFT indoors at 18 m; (c) the SD outdoors at 16 m; (d) the spectrum using STFT outdoors at 16 m.
 X. Liang et al. / Digital Signal Processing 74 (2018) 72–93 91

Fig. 27. Clutter suppression using (a) FFT; (b) one FA; (c) two FA; (d) four FA; and (e) six FA.

Table 5 References
Improvement of SNR using different methods.

Method FFT One FA Two FA Four FA Six FA [1] S.D. Liang, Sense-through-wall human detection based on UWB radar sensors,
Signal Process. 126 (2016) 117–124.
SNR (dB) −24.37 −11.44 −8.89 −3.92 −3.88 [2] A.H. Muqaibel, A.T. Abdalla, M.T. Alkhodary, S.A. Alawsh, Through-the-wall
radar imaging exploiting Pythagorean apertures with sparse reconstruction,
Digit. Signal Process. 61 (2017) 86–96.
[3] M.I. Stanciu, S. Azou, E. Rădoi, A. Şerbănescu, A statistical analysis of multipath
proposed method. This indicates that the method is better able to interference for impulse radio UWB systems, J. Franklin Inst. 352 (12) (2014)
remove clutter and improve the SNR. 5952–5967.
[4] J. Wang, X. Wang, L. Chen, J. Huangfu, Noncontact distance and amplitude-
independent vibration measurement based on an extended DACM algorithm,
6. Conclusion IEEE Trans. Instrum. Meas. 63 (1) (2014) 145–153.
[5] A. Singh, X. Gao, E. Yavari, M. Zakrzewski, Data-based quadrature imbalance
In this paper, a new algorithm for vital sign detection was de- compensation for a CW Doppler radar system, IEEE Trans. Microw. Theory Tech.
61 (4) (2013) 1718–1724.
veloped by employing the Impulse UWB radar. The range of human [6] G. Wang, C. Gu, T. Inoue, C. Li, A hybrid FMCW-interferometry radar for indoor
subject can be acquired based on the short-time Fourier transform precise positioning and versatile life activity monitoring, IEEE Trans. Microw.
(STFT) technique, and the respiration frequency (RF) was estimated Theory Tech. 62 (11) (2014) 2812–2822.
using the ensemble empirical mode decomposition (EEMD) based [7] M. Mercuri, P.J. Soh, L. Boccia, D. Schreurs, Optimized SFCW radar sensor aim-
ing at fall detection in a real room environment, in: Proc. IEEE Biomed. Wire-
frequency accumulation (FA) method. The performance of the de-
less Technol. Netw. Sens. Syst., Austin, TX, USA, 2013, pp. 4–6.
veloped method was validated compared with several well-known [8] M. Mercuri, J.S. Ping, G. Pandey, P. Karsmakers, Analysis of an indoor biomedical
algorithms in different conditions. Results indicated that the new radar-based system for health monitoring, IEEE Trans. Microw. Theory Tech.
method can suppress the clutter effectively and can be imple- 61 (5) (2013) 2061–2068.
[9] A. Muqaibel, A. Safaai-Jazi, Characterization of wall dispersive and attenuative
mented easily.
effects on UWB radar signals, J. Franklin Inst. 345 (6) (2008) 640–658.
[10] Y. Wang, Q. Liu, A.E. Fathy, Simultaneous localization and respiratory detection
Acknowledgments of multiple people using low cost UWB biometric pulse Doppler radar sensor,
in: IEEE MTT-S Int. Microw. Symp. Dig., Montreal, QC, Canada, 2012, pp. 1–3.
[11] Y. Wang, Q. Liu, A.E. Fathy, CW and pulse-Doppler radar processing based on
This work was funded by the Nature Science Foundation of FPGA for human sensing applications, IEEE Trans. Geosci. Remote Sens. 51 (5)
China (41527901, 61701462, 61501424), the Major Program of Chi- (2013) 3097–3107.
na’s Second Generation Satellite Navigation System (GF**********03), [12] Y.S. Koo, L. Ren, Y. Wang, A.E. Fathy, UWB micro Doppler radar for human
gait analysis, tracking more than one person, and vital sign detection of mov-
the Fundamental Research Funds for the Central Universities ing persons, in: IEEE MTT-S Int. Microw. Symp. Dig., Seattle, WA, USA, 2013,
(201713018), the National High Technology Research and Devel- pp. 1–4.
opment Program of China (2012AA061403), the National Science [13] Y. Nijsure, W.P. Tay, E. Gunawan, F. Wen, An impulse radio ultrawideband sys-
and Technology Pillar Program during the Twelfth Five-year Plan tem for contactless noninvasive respiratory monitoring, IEEE Trans. Biomed.
Eng. 60 (6) (2013) 1509–1517.
Period (2014BAK12B00), Qingdao National Laboratory for Marine
[14] J. Li, L. Liu, Z. Zeng, Advanced signal processing for vital sign extraction with
Science and Technology (2017ASKJ01), and Qingdao Science and applications in UWB radar detection of trapped victims in complex environ-
Technology Plan (17-1-1-7-jch). ments, IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 7 (3) (2014) 783–791.
92 X. Liang et al. / Digital Signal Processing 74 (2018) 72–93

[15] Z. Li, W. Li, H. Lv, A novel method for respiration-like clutter cancellation in [45] C. Li, J. Lin, Random body movement cancellation in Doppler radar vital sign
life detection by dual-frequency IR-UWB radar, IEEE Trans. Microw. Theory detection, IEEE Trans. Microw. Theory Tech. 56 (12) (2008) 3143–3152.
Tech. 61 (5) (2013) 2086–2092. [46] B.-K. Park, O. Boric-Lubecke, V.M. Lubecke, Arctangent demodulation with DC
[16] W. Hu, Z. Zhao, Y. Wang, H. Zhang, F. Lin, Noncontact accurate measurement offset compensation in quadrature Doppler radar receiver systems, IEEE Trans.
of cardiopulmonary activity using a compact quadrature Doppler radar sensor, Microw. Theory Tech. 55 (5) (2007) 1073–1079.
IEEE Trans. Biomed. Eng. 61 (3) (2014) 725–735. [47] L. Ren, H. Wang, K. Naishadham, O. Kilic, A.E. Fathy, Phase-based methods for
[17] C. Gu, C. Li, Assessment of human respiration patterns via noncontact sensing heart rate detection using UWB impulse Doppler radar, IEEE Trans. Microw.
using Doppler multi-radar system, Sensors 15 (3) (2015) 6383–6398. Theory 64 (10) (2016) 3319–3331.
[18] A. Lazaro, D. Girbau, R. Villarino, Techniques for clutter suppression in the [48] R.G. Lyons, Understanding Digital Signal Processing, Publishing House of
presence of body movements during the detection of respiratory activity Electronics Industry, Beijing, 2012.
through UWB radars, Sensors 14 (2) (2014) 32595–32618. [49] S. Haykin, B. Van Veen, Signals and Systems, John Wiley & Sons, New York,
[19] L. Chuantao, C. Fuming, J. Jingxi, L. Hao, L. Sheng, L. Guohua, W. Jianqi, 1999.
A method for remotely sensing vital signs of human subjects outdoors, Sensors [50] R.G.T. Mello, L.F. Oliveira, J. Nadal, Digital Butterworth filter for subtract-
15 (7) (2015) 14830–14844. ing noise from low magnitude surface electromyogram, Comput. Methods
[20] L. Ren, Noncontact multiple heartbeats detection and subject localization us- Programs Biomed. 87 (1) (2007) 28–35.
ing UWB impulse Doppler radar, IEEE Microw. Wirel. Compon. Lett. 25 (10) [51] C. Li, F. Chen, F. Qi, M. Liu, Z. Li, F. Liang, X. Jing, G. Lu, J. Wang, Search-
(2015) 690–692. ing for survivors through random human-body movement outdoors by
[21] M.C. Huang, J.J. Liu, W. Xu, C. Gu, C. Li, A self-calibrating radar sensor sys- continuous-wave radar array, PLoS ONE 11 (4) (2016) 1–13.
tem for measuring vital signs, IEEE Trans. Biomed. Circuits Syst. 10 (2) (2016) [52] X. Liang, H. Zhang, T.A. Gulliver, G. Fang, S. Ye, An improved algorithm for
352–363. through-wall target detection using ultra-wideband impulse radar, IEEE Access
[22] F. JalaliBidgoli, S. Moghadami, S. Ardalan, A compact portable microwave 5 (2017) 22101–22118.
life-detection device for finding survivors, IEEE Embed. Syst. Lett. 8 (1) (2016) [53] X. Liang, H. Zhang, T. Lu, T.A. Gulliver, Energy detector based TOA estimation
10–13. for MMW systems using machine learning, Telecommun. Syst. 64 (2) (2017)
[23] G. Gennarelli, G. Ludeno, F. Soldovieri, Real-time through-wall situation aware- 417–427.
ness using a microwave Doppler radar sensor, Remote Sens. 8 (8) (2016) 621. [54] R. Buccheri, B. Sacco, Time Analysis in Astronomy: Tools for Periodicity
[24] C. Le, T. Dogaru, L. Nguyen, M.A. Ressler, Ultra-wideband radar imaging of Searches, vol. 24, Springer, US, 1985, pp. 15–27.
building interior: measurements and predictions, IEEE Trans. Geosci. Remote [55] B. Allen, Short term spectral analysis, synthesis, and modification by dis-
Sens. 47 (5) (2009) 1409–1420. crete Fourier transform, IEEE Trans. Audio Speech Lang. Process. 25 (3) (1977)
[25] Q. Huang, L. Qu, G. Fang, UWB through-tall imaging based on compressive 235–238.
sensing, IEEE Trans. Geosci. Remote Sens. 48 (3) (2010) 1408–1415. [56] J. Antoni, The spectral kurtosis: a useful tool for characterising non-stationary
[26] V.T. Vu, T.K. Sjogren, M.I. Pettersson, A. Gustavsson, Detection of moving tar- signals, Mech. Syst. Signal Process. 20 (2) (2006) 282–307.
gets by focusing in UWB SAR theory and experimental results, IEEE Trans. [57] L. Durak, O. Arikan, Short-time Fourier transform: two fundamental proper-
Geosci. Remote Sens. 48 (10) (2010) 3799–3815. ties and an optimal implementation, IEEE Trans. Signal Process. 51 (5) (2003)
[27] X. Zhuge, A. Yarovoy, A sparse aperture MIMO-SAR based UWB imaging sys-
1231–1242.
tem for concealed weapon detection, IEEE Trans. Geosci. Remote Sens. 49 (1)
[58] K. Wojcicki, M. Milacic, A. Stark, J. Lyons, Exploiting conjugate symmetry of
(2011) 509–518.
the short-time Fourier spectrum for speech enhancement, IEEE Signal Process.
[28] M. Ascione, A. Buonanno, M. D’Urso, L. Angrisani, A new measurement method
Lett. 15 (2008) 461–464.
based on music algorithm for through-the-wall detection of life signs, IEEE
[59] A.H. Liao, C.C. Shen, P.C. Li, Potential contrast improvement in ultrasound pulse
Trans. Instrum. Meas. 62 (1) (2013) 13–26.
inversion imaging using EMD and EEMD, IEEE Trans. Ultrason. Ferroelectr.
[29] L. Liu, Z. Liu, B. Barrowes, Through-wall bio-radiolocation with UWB impulse
Freq. Control 57 (2) (2010) 317–326.
radar—observation, simulation and signal extraction, IEEE J. Sel. Top. Appl.
[60] Z. Liu, Y. Liu, H. Shan, B. Cai, Q. Huang, A fault diagnosis methodology for
Earth Obs. Remote Sens. 4 (4) (2011) 791–798.
gear pump based on EEMD and Bayesian network, PLoS ONE 10 (5) (2015)
[30] L. Liu, Z. Liu, H. Xie, B. Barrowes, A.C. Bagtzoglou, Numerical simulation of
e0125703.
UWB impulse radar vital sign detection at an earthquake disaster site, Ad Hoc
[61] Z. Liu, Y. Cui, W. Li, A classification method for complex power quality distur-
Netw. 13 (1) (2014) 34–41.
bances using EEMD and rank wavelet SVM, IEEE Trans. Smart Grid 6 (4) (2015)
[31] M. Baldi, G. Cerri, F. Chiaraluce, L. Eusebi, P. Russo, Non-invasive UWB sensing
1678–1685.
of astronauts’ breathing activity, Sensors 15 (1) (2015) 565–591.
[62] Z. Wang, D. Wu, J. Chen, A. Ghoneim, M.A. Hossain, A triaxial accelerometer-
[32] Z. Li, W. Li, H. Lv, Y. Zhang, A novel method for respiration-like clutter cancel-
based human activity recognition via EEMD-based features and game-theory-
lation in life detection by dual-frequency IR-UWB radar, IEEE Trans. Microwave
based feature selection, IEEE Sens. J. 16 (9) (2016) 3198–3207.
Theory Tech. 61 (5) (2013) 2086–2092.
[63] Y. Yang, H. Chen, T. Jiang, Nonlinear response prediction of cracked rotor based
[33] J. Rovňáková, D. Kocur, Weak signal enhancement in radar signal processing,
on EMD, J. Franklin Inst. 352 (8) (2015) 3378–3393.
in: 2010 20th International Conference Radioelektronika, 2010, pp. 1–4.
[64] L. Liu, Z. Liu, B. Barrowes, Through-wall bio-radiolocation with UWB impulse
[34] E. Conte, A. Filippi, S. Tomasin, ML period estimation with application to vital
radar—observation, simulation and signal extraction, IEEE J. Sel. Top. Appl.
sign monitoring, IEEE Signal Process. Lett. 17 (11) (2010) 905–908.
Earth Obs. Remote Sens. 4 (4) (2011) 791–798.
[35] A. Nezirovíc, A. Yarovoy, L. Ligthart, Signal processing for improved detection
of trapped victims using UWB radar, IEEE Trans. Geosci. Remote Sens. 48 (4) [65] L. Marple, Computing the discrete-time “analytic” signal via FFT, IEEE Trans.
(2010) 2005–2014. Signal Process. 47 (9) (1999) 2600–2603.
[36] J. Sachs, M. Helbig, R. Herrmann, M. Kmec, K. Schilling, E. Zaikov, Remote vi- [66] Y. Yang, A.E. Fathy, Development and implementation of a realtime see-
tal sign detection for rescue, security, and medical care by ultra-wideband through-wall radar system based on FPGA, IEEE Trans. Geosci. Remote Sens.
pseudo-noise radar, Ad Hoc Netw. 13 (2014) 42–53. 47 (5) (2009) 1270–1280.
[37] C. Kim, J. Lee, ToA-based multi-target localization and respiration detection [67] L. Liu, G. Fang, A novel UWB sampling receiver and its applications for impulse
using UWB radars, EURASIP J. Wirel. Commun. Netw. 2014 (1) (2014) 1–15. GPR systems, IEEE Geosci. Remote Sens. Lett. 7 (4) (2010) 690–693.
[38] Z. Zhang, Human-target detection and surrounding structure estimation un-
der a simulated rubble via UWB radar, IEEE Trans. Geosci. Remote Sens. 10 (2)
(2013) 328–331. Xiaolin Liang was born in Shandong, China, in
[39] Y. Xu, S. Wu, C. Chen, J. Chen, A novel method for automatic detection of 1988. He is currently working towards the Ph.D. de-
trapped victims by ultrawideband radar, IEEE Trans. Geosci. Remote Sens. gree in the Department of Electronic Engineering,
50 (8) (2012) 3132–3142. Ocean University of China, Qing Dao, China. His re-
[40] Y. Xie, G. Fang, Equi-amplitude tracing algorithm based on base-band pulse
search interests include ultra-wideband radio systems,
signal in vital sign detecting, Electron. Inf. Technol. 31 (5) (2009) 1132–1135.
60 GHz wireless systems, signal processing, ultra-
[41] A. Nezirovic, A.G. Yarovoy, L.P. Ligthart, Signal processing for improved detec-
tion of trapped victims using UWB radar, IEEE Trans. Geosci. Remote Sens. wideband radar, and vital sign detection.
48 (4) (2010) 2005–2014.
[42] S. Wu, K. Tan, Z. Xia, J. Chen, Improved human respiration detection method Hao Zhang was born in Jiangsu, China, in 1975. He
via ultra-wideband radar in through-wall or other similar conditions, IET Radar
received the B.S. degree in telecom engineering and
Sonar Navig. 10 (3) (2016) 468–476.
industrial management from Shanghai Jiaotong Uni-
[43] Y. Xu, S. Dai, S. Wu, J. Chen, Vital sign detection method based on multiple
higher order cumulant for ultra-wideband radar, IEEE Trans. Geosci. Remote
versity, China, in 1994, the MBA degree from New
Sens. 50 (4) (2012) 1254–1265. York Institute of Technology, USA, in 2001, and the
[44] X. Hu, T. Jin, Short-range vital signs sensing based on EEMD and CWT using Ph.D. degree in electrical engineering from the Univer-
IR-UWB radar, Sensors 16 (12) (2016) 2025. sity of Victoria, Canada, in 2004. From 1994 to 1997,
 X. Liang et al. / Digital Signal Processing 74 (2018) 72–93 93

he was the Assistant President of ICO (China) Global Communications Italy. From 2001 to 2003, he was a Special Foreign Research Fellow of the
Company. He is now a Professor of the Department of Electrical Engineer- Japan Society for the Promotion of Science with Tohoku University, Sendai,
ing at Ocean University of China. He is also an Adjunct Professor at the Japan. Since 2004, he has been a Professor with the Institute of Electron-
University of Victoria. He is a Senior Member of IEEE. His research inter- ics, Chinese Academy of Sciences, Beijing, China, and the Director of the
ests include ultra-wideband systems, MIMO wireless systems, and spread Key Laboratory of Electromagnetic Radiation and Sensing Technology. He
spectrum communications. is the author of more than 100 publications. His research interests include
ultrawideband radar, ground-penetrating radar signal processing and iden-
Shengbo Ye received the Ph.D. degree from the tification methods, and computational electromagnetics.
Institute of Electronics, Chinese Academy of Sciences
(CAS), Beijing, China, in 2011. Since 2011, he has been T. Aaron Gulliver received the Ph.D. degree in elec-
with the Key Laboratory of Electromagnetic Radiation trical engineering from the University of Victoria, Vic-
and Sensing Technology, CAS. His research interests toria, British Columbia, Canada, in 1989. From 1989 to
include UWB through-wall radar detection, imaging, 1991, he was employed as a Defence Scientist at the
ground-penetrating radar signal processing life detec- Defence Research Establishment Ottawa, Ottawa, On-
tion, and other related applications. tario, Canada. He has held academic positions at Car-
leton University, Ottawa, and the University of Canter-
Guangyou Fang received the B.S. degree in elec- bury, Christchurch, New Zealand. He joined the Uni-
trical engineering from Hunan University, Changsha, versity of Victoria in 1999 and is a Professor in the
China, and the M.S. and Ph.D. degrees in electrical en- Department of Electrical and Computer Engineering. He is a Senior Mem-
gineering from Xi’an Jiaotong University, Xi’an, China, ber of the IEEE and a Member of the Association of Professional Engineers
in 1984, 1990, and 1996, respectively. From 1990 to of Ontario, Canada. In 2002, he became a Fellow of the Engineering In-
1999, he worked as an Engineer, an Associate Profes- stitute of Canada, and in 2012 a Fellow of the Canadian Academy of
sor, and a Professor with the China Research Institute Engineering. His research interests include information theory and com-
of Radiowave Propagation, Xinxiang, China. From 2000 munication theory, algebraic coding theory, cryptography, construction of
to 2001, he was a Visiting Scholar at the University of optimal codes, iterative coding, MIMO communications, space-time cod-
Trieste, Trieste, Italy, and the International Center for Science and High ing, and ultra-wideband communications.
Technology-United Nations Industrial Development Organization, Trieste,