I.

INTRODUCTION
Target detection has been studied for many years [1], [2],
but many studies have not considered the multipath ef-
Moving Target Detection Based fect. However, the multipath effect is inevitable in realis-
tic environments for radar systems and severely degrades
on Time Reversal in a detection performance, particularly in indoor and urban
environments and in low-angle detection scenarios [3]–[10].
Multipath Environment Therefore, target-detection problems in multipath environ-
ments have become a popular research topic and have
attracted widespread attention. Two approaches are usu-
ally pursued to deal with multipath effects. On the one
hand, the multipath in radar systems that are regarded as
ZHAOMING ZHANG , Graduate Student Member, IEEE clutter [5]–[7] or jamming [8], [9] can be eliminated, and
BAIXIAO CHEN furthermore, the robustness of detectors to multipath effects
MINGLEI YANG , Member, IEEE can be enhanced, overcoming the influence of multipath
Xidian University Xian, China effects on the detection results [10]–[13].
On the other hand, multipath effects can be exploited
to improve target detection, localization, and tracking per-
formance [3], [14]–[19]. The reflected multipath can be
Time reversal can resolve multipath problems and increase the used to create virtual sensors [15], which can effectively
detection probability, but the detection performance decreases when reduce hardware resources and expand the virtual radar
the target is moving. Because of the change in the channel response
due to target motion, the focusing effect of a time-reversal-transmitted
aperture. Moreover, for indoor scenarios, multipath effects
signal is reduced. In this article, a novel method based on time reversal, are directly and efficiently utilized to increase the signal-to-
which detects a moving target embedded in a multipath environment clutter ratio at the true target locations [3], which is impor-
with channel variation, is proposed. We establish conventional and tant for detection. Some methods of target detection have
time-reversal signal models of a moving target in an indoor diffuse been proposed in [20]–[24] in the case of dense multipath
multipath environment and derive likelihood ratio tests under these
two conditions. The channel parameters and the changes due to target
environments. Regarding the mismatched problem of the
motion are obtained by echo estimation and are exploited as useful traditional adaptive matched filter (AMF) [20] caused by
information by the detectors. The channel changes are taken into multipath effects, the tunable AMF (T-AMF) [21] and mul-
account in the time-reversal detector to improve its robustness. The tipath AMF (MP-AMF) utilize the multipath energy [22]
time-reversal detector proposed in this article possesses the constant- to improve the detection probability. In addition to the
false-alarm-rate property with respect to the unknown background
noise. Although Monte Carlo simulation analysis has illustrated the
energy [23], the spatial diversity of the multipath effect can
superior performance of both conventional and time-reversal detec- be exploited [24] by radar systems to enhance the detection
tors and proved that both methods can improve the detection prob- performance.
ability by exploiting the multipath effect, the time-reversal method In addition to these multipath exploitation methods, the
outperforms the conventional method due to the spatial and temporal time-reversal (TR) technique exploits the characteristics of
focusing effect on the channel.
the multipath effect to realize spatial and temporal focus-
ing by channel matching and to enhance the echo energy
without prior environmental knowledge so that the impact
of the multipath effect is favorable rather than adverse.
In this article, we apply the TR technique to radar target
Manuscript received September 15, 2020; revised January 21, 2021 and detection to detect a moving target in an indoor scenario
March 19, 2021; released for publication March 19, 2021. Date of publi- with channel variation. Although the multipath exploitation
cation April 26, 2021; date of current version October 11, 2021. methods discussed above are very effective, TR can com-
DOI. No. 10.1109/TAES.2021.3074131 bine a variety of advantages, such as the advantage of using
Refereeing of this contribution was handled by K. V. Mishra.
multipath to build virtual sensors to expand the virtual radar
aperture in [15], and the advantage of focusing multipath
This work was supported in part by the National Natural Science Foun- energy on the target location to improve the signal-to-noise
dation of China under Grant 6197010590, in part by the Fund for Foreign
Scholars in University Research and Teaching Programs (the 111 Project)
(SNR) ratio in [3]. In addition, the Doppler effect is not
under Grant B18039, and in part by the National Natural Science Founda- the focus of [15], but a moving target is considered in
tion of Shaanxi Province under Grant 2017CGZH-RGXQ-02. our article, which leads to the fact that Doppler cannot
Authors’ address: The authors are with the National Laboratory
be ignored. The multipath exploitation method in [3] is
of Radar Signal Processing, Xidian University, Xian 710071, suitable for the synthetic aperture radar (SAR) system,
China, E-mail: ([email protected]; [email protected]; whereas this article considers a single-antenna radar system.
[email protected]) (Corresponding author: Baixiao Chen.) Consequently, we utilize TR to solve the problem of target
detection in a multipath environment. The TR technique
0018-9251 © 2021 IEEE. reverses the received signal containing the multipath signal

IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 57, NO. 5 OCTOBER 2021 3221

Authorized licensed use limited to: SHIV NADAR UNIVERSITY. Downloaded on October 09,2024 at 07:21:18 UTC from IEEE Xplore. Restrictions apply.
in the time domain (conjugates in the frequency domain), The remainder of the article is organized as follows. In
to normalize the energy and to retransmit the processed Section II, we describe the detection problem and establish
signal to the same scattering environment. TR was proposed the conventional and TR signal models. In Section III, we
by Fink and applied to acoustics because this technique formalize the binary hypothesis test problems and, then,
can correct signal distortion caused by an inhomogeneous derive the detectors and introduce the parameter estimation
medium without prior information [25], [26]. (PE) method. Section IV presents the performance of the
TR has been proven to be effective in an electromag- detectors. Section V describes the numerical simulations.
netic field [27] when the scattering channel is invariant Finally, Section VI concludes this article.
and reciprocal. Furthermore, due to multipath exploitation, This article makes two main contributions. First, we
TR has been demonstrated to enable improved target de- propose a TR detector that can avoid the adverse multipath
tection compared to conventional methods. However, most effects and instead exploit them to improve the detection
previous studies have assumed that the target is static; that performance. Second, we add the channel disturbance from
is, the channel remains fixed [28], [29]; nevertheless, the the target movement to the detector, solving the TR mis-
effect of TR decreases when the channel changes, which is match problem caused by the channel changes.
a drawback of TR. Jin and collaborators have derived a TR Notations: In this article, we use lowercase boldface
generalized likelihood ratio test (TR–GLRT) detector for a letters to represent vectors and uppercase boldface letters to
single-antenna system [30]. With a stationary channel, the represent matrices. CN (μ, σ 2 ) denotes a complex Gaussian
TR–GLRT detector in [30] shows better performance than distribution with mean μ and variance σ 2 . The symbol
the conventional energy detector (ED). Furthermore, a TR ∗ indicates convolution. XM2 (γ ) represents the noncentral
detector with a single antenna has been studied under more Chi-square distribution with M degrees of freedom and
practical conditions where the channel response changes noncentral parameter γ . The symbols (· )∗ , (· )T , and (· )H
due to clutter [31]. To exploit the spatial diversity char- denote a conjugate, transpose, and Hermitian transpose,
acteristics of antennas, multiple input and multiple output respectively. IN is the N-dimensional identity matrix. A−1
(MIMO) [32] and multistatic [33] radar systems based is the inverse of matrix A. The symbol |A| denotes the
on TR were proposed for dense multipath environments. determinant of the matrix. ||A|| indicates the Frobenius
Although O’ Donoughue and Moura[31] and Jin et al. [32] norm. A  B represents the Hadamard product of the vec-
provide TR detectors under conditions of channel changes, tor or matrix. diag{a} represents a diagonal matrix with
they do not specifically analyze the impact of target motion diagonal vector a. E {·} and Var{·} represent the expected
on the detection results. In the most realistic detection prob- value and variance, respectively. The symbol  represents
lems, the target always moves, leading to channel variation an approximate statistical distribution.
that causes channel mismatch during the TR process and
reduces the focusing effect of TR. In [34] and [35], TR
detectors suitable for single and array antennas, respec- II. SIGNAL MODEL
tively, are developed to detect a moving target with the We consider a scenario in which a point target moves
Doppler frequency. Nevertheless, the channel is assumed in an indoor environment. In this scenario, many multipath
to be unchanged in the TR process in these works, which scattering signals are generated by a large number of ground
is essentially equivalent to a stationary target. However, objects. Furthermore, for a complex indoor scattering envi-
the assumption of a fixed channel is too idealistic to be ronment, the target may be blocked due to many reflectors,
satisfied in many scenarios. For nonstationary but correlated so we assume there is no direct path. We consider the
channels, a more realistic and meaningful study, Setlur and scattering environment that is under surveillance by radar
Devroye [36] demonstrated that TR may still have better is known in advance. Therefore, the number of propagation
mutual information gain performance than conventional paths between the target and the radar can be regarded
systems. as prior knowledge [37]. The reflection environment is
In this article, we develop a TR likelihood ratio test (TR– illustrated in Fig. 1, where the two-way arrow lines for
LRT) to identify a moving target with Doppler frequency each color represent a propagation path. In this article, we
detection. The proposed TR–LRT detector is suitable for consider the signals received by radar consist of two parts:
a multipath environment with channel variation due to One part is reflected only by other objects (unexpected
the target motion, corresponding relatively well to actual stationary targets) or the reflection surface, i.e., background
conditions. In addition, a conventional likelihood ratio test environment, and the other is related to the target. The
(CO–LRT) method exploiting multipath in the detection of signals reflected by other objects or the reflection surface are
a target is derived to benchmark the TR–LRT detector per- defined as clutter (dotted lines in Fig. 1). Their propagation
formance. The parameters of channels are estimated prior to paths do not pass by the target; so, they do not interact
detection, and subsequently, these parameters can be input with the target. Considering that the generation of clutter
into the detectors as useful information. The performance of is independent of the target, the clutter can be eliminated
the two detectors is studied by the Monte Carlo simulation. by background subtraction [28], [30], [38], [39] from the
We demonstrate that the TR detector performs significantly overall received signals. Then, the remaining data are the
better than the conventional detector in multipath environ- target-reflected signals (multipath) and noise. Any signals
ments. related to the target (solid lines in Fig. 1), such as the target

3222 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 57, NO. 5 OCTOBER 2021

Authorized licensed use limited to: SHIV NADAR UNIVERSITY. Downloaded on October 09,2024 at 07:21:18 UTC from IEEE Xplore. Restrictions apply.
 L propagation paths and can be further expressed as

L
h (t ) = hl (t ). (4)
l =1

Here, hl (t ) is the channel impulse response of the lth

propagation path
hl (t ) = αl (t )δ (t − τl ), l = 1, . . . , L. (5)
τl is the time delay of the lth path, and δ (t ) denotes the
Dirac delta function. αl (t ) is the attenuation coefficient of
the lth path. The pulse duration is much shorter than the
coherence time of the radar-target channel. Hence, αl (t )
can be represented as αl [40], [41]. We consider a Rayleigh
fading channel. The propagation paths are independent of
each other, implying that αl is independent and identically
distributed, and αl satisfies a complex Gaussian distribution
with zero mean and variance σα2 , i.e.,
αl ∼ CN (0, σα2 ). (6)
We consider only the overlapping echoes and ignore
Fig. 1. Schematic representation of the scattering scenario. multipath signals with delay times longer than signal dura-
tion T . Therefore, the multipath time delay τl < T .
The entire TR process includes two steps: The first step
scattering signal propagating to the reflecting surface, or
is called forward transmission and is consistent with the
the signal scattered to other objects and received by the
conventional approach; the second step is defined as back-
radar, are defined as multipath signals. We consider a single
ward transmission and is the reverse processing performed
omnidirectional antenna radar system. The initial probing
on the basis of conventional echo; and the retransmitted
signal is s (t ) with energy Es and signal duration T
signal in the backward transmission is focused on the target
 T by matching the channel after the TR processing to achieve
Es = |s (t )|2 dt. (1) an echo with higher energy.
0

The target is assumed to be moving at a constant speed vs A. Clutter Response

relative to the radar with angle θ. There are assumed to be This phase estimates the clutter response hc (t ). The
L multiple paths related to the target in the scattering envi- surveillance environment is assumed to be fixed, and no
ronment. The direction of arrival of the lth path is defined target is present. The received signal is the clutter and noise.
as φl . In the indoor environment, rich diffuse multipaths are The transmitting and recording of M independent snapshots
received by the radar antenna at random angles. Therefore, of data are given by
we can consider φl to be a uniform distribution variable from
0 to 2π. The Doppler frequency shift is derived as [34] yc,m (t ) = s (t ) ∗ hc (t ) + nc (t )
(7)
= c (t ) + nc (t ), m = 1, . . . , M
2|vs |(cos θ cos φl + sin θ sin φl )
fdl = where nc (t ) is additive, zero mean, complex white Gaussian
c/ fc (2)
noise. For sufficient snapshots, we have
= βl fc , l = 1, . . . , L
c (t ) ≈ ĉ (t )
where βl is defined as the Doppler frequency factor of the
1 
M
lth propagation path, c is the speed of electromagnetic wave = yc,m (t ) (8)
propagation, and fc is the radar carrier frequency. M m=1
The channel response that is defined as H (t ) consists of
= s (t ) ∗ hc (t ).
clutter response hc (t ) and target multipath response h (t )1
It is straightforward to obtain hc (t ) by deconvolution of
H (t ) = hc (t ) + h (t ) (3) ĉ (t ). Accordingly, we assume that the clutter response hc (t )
is known in the following analysis.
where target multipath response h (t ) is composed of the
superposition of the impulse responses corresponding to the B. Conventional Signal Model–Forward Transmission
In the forward transmission, when there is a moving tar-
1 As mentioned previously, the clutter would be subtracted, so its expression get with the Doppler frequency factor βl in the environment,
is not denoted here. the received data contain channel information, and with the

ZHANG ET AL.: MOVING TARGET DETECTION BASED ON TIME REVERSAL IN A MULTIPATH ENVIRONMENT 3223

Authorized licensed use limited to: SHIV NADAR UNIVERSITY. Downloaded on October 09,2024 at 07:21:18 UTC from IEEE Xplore. Restrictions apply.
transmitted signal s (t ), the echo d (t ) is given by h (t ). There are L paths related to the target in the scattering
environment, and each path varies due to the target move-

L
ment, which may lead to inconsistencies in the variation
d (t ) = hl (t ) ∗ s (t )e jωc βl t + hc (t ) ∗ s (t ) + v(t ) (9)
of different paths and different correlation coefficients for
l =1
each path in the backward channel. Therefore, the correla-
where v(t ) is the additive complex white Gaussian noise tion coefficient between the forward and backward channel
with zero mean and variance σv2 . Using the background responses of the lth path can be defined as ρl . Subsequently,
subtraction technique, the signals related to the target x (t ) the backward target transmission channel can be further
can be obtained defined as the superposition of the forward channel and the
x (t ) = d (t ) − s (t ) ∗ hc (t ) disturbance that is expressed as

L

L
= hl (t ) ∗ s (t )e jωc βl t + v(t ) h (t ) = ρl hl (t ) + (t )
l =1
(10) l =1

L (12)
jωc βl (t−τl ) 
L
= αl s (t − τl )e + v(t ) = ρl αl δ (t − τl ) + (t ).
l =1
l =1
= r (t ) + v(t )
Transform forward and backward channel responses to
where r (t ) is defined as the target-response signal the frequency domain. Then, the variance  of the forward

L channel response can be obtained as Var( Ll=1 αl e− jωτl ) =
r (t ) = αl s (t − τl )e jωc βl (t−τl ) , (11) Lσα2 , and the variance of the backward channel response is
 
l =1 Var( Ll=1 ρl αl e− jωτl ) + σ 2 = Ll=1 ρl2 σα2 + σ 2 . Here, we
which is the superposition of echo signals corresponding to constrain h (t ) and h (t ) to follow the same distribu-
different delay propagation paths and reflects the channel tion [31], [32]. Thus, we have
scattering conditions.

L

C. TR Signal Model–Backward Transmission ρl2 σα2 + σ 2 = Lσα2 . (13)

l =1
In the backward transmission step, the data received 
from the forward transmission are reversed in the time To simplify the analysis, Ll=1 ρl2 σα2 in (13) can be replaced
domain, the energy is normalized, and the signals are re- by Lρ 2 σα2 , where ρ ranges from 0 to 1, which is given by
transmitted to the same environment. 
When the environment is unchanged and the target is sta- σ2
tionary, the forward transmission channel and the backward ρ = 1− . (14)
Lσα2
transmission channel are completely consistent, and an op-
timum channel matching effect can be obtained. However, It need be pointed out that the correlation coefficient ρ is
because the target is moving, the channel response changes obtained by a statistical analysis. Consequently, h (t ) can be
and leads to the mismatch of the reciprocal channel. We note rewritten as
that for an indoor scenario, although the target is moving,
its speed is often slow, and thus, the variation in the channel h (t ) = ρh (t ) + (t ). (15)
response over a short time will not be very large [42], [43].
Hence, the backward target multipath channel response h (t ) When ρ is equal to 0, the forward and backward channels
in the TR process is correlated with the forward transmission are completely decorrelated; this case is not the focus of our
channel h (t ). Clutter is independent of the target. Therefore, research and is not discussed in this article.
the clutter response is considered to be constant and is still As described in (15), in the backward transmission
hc (t ). For the multipath channel response, the variation for stage, the propagation channel is composed of two parts:
each path due to target movement may be inconsistent. One part is considered to be consistent with the forward
Nevertheless, the channel variation can be regarded as an channel, and the other is the disturbance term that includes
entirety that is defined as channel disturbance (t ), which the variations in the channel and the Doppler frequency
is a random variable. We should note that the channel is factor caused by the target motion. Hence, although the
random and time varying. Since the variation can be taken parameters of the multipath and the Doppler frequency in
part into the disturbance term (t ), αl (t ) can be written as αl . the backward channel may change, we can still assume
As introduced in [44], dense multipath results in a complex that the number of propagation paths in the TR process is
Gaussian process model for the impulse response; hence, we consistent with that of the conventional stage. The Doppler
can assume that the channel disturbance follows a complex frequency factor β l of the TR process can be considered the
Gaussian distribution with zero mean and variance σ 2 , i.e., same as βl . All of the variations are unified in the disturbance
(t ) ∼ CN (0, σ 2 ). In addition, similar to the work in [31] item (t ). At this stage, the transmitted signal is kx ∗ (−t ),
and [32], the disturbance is assumed to be independent of where k is the energy normalized factor ensuring that the

3224 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 57, NO. 5 OCTOBER 2021

Authorized licensed use limited to: SHIV NADAR UNIVERSITY. Downloaded on October 09,2024 at 07:21:18 UTC from IEEE Xplore. Restrictions apply.
total energy of each transmitted signal is consistent where w(t ) is the additive complex circular white Gaus-
 sian noise with zero mean and variance σw2 . The detailed
Es derivation of y (t ) is introduced in Appendix A. As the
k = T . (16)
0 |x (t )| dt
2 target moves faster, the channel disturbance becomes larger,
resulting in an enhanced channel disturbance response and
The received signal is given by poor focusing effect.
 L 
 
L
e(t ) = ρ hl (t ) + (t ) ∗ kx ∗ (−t )e jωc β l t III. DETECTORS
l =1 l =1
(17)
After digital sampling in the time domain, we use the
+ hc (t ) ∗ kx ∗ (−t ) + w(t ). discrete form of the signals for analysis. We pose a binary
Similar to (10), hc (t ) ∗ kx ∗ (−t ) is subtracted out from the hypothesis test to describe the detection problem. Under
received signal e(t ), and then, the TR target echo is given the null hypothesis H0 , there are no targets in the scattering
by environment. Under the alternative hypothesis H1 , the echo
data contain the target signals. The TR–LRT and CO–LRT
y (t ) = e(t ) − hc (t ) ∗ kx ∗ (−t ) are derived in a Neyman–Pearson sense as follows. Both
(18)
≈ z1 (t ) + z2 (t ) + w(t ). tests can use multipath to detect a target in a dense scattering
environment.
z1 (t ) is the focused signal response that is the echo of the
signal that achieves space–time focus at the target position A. TR Likelihood Ratio Test
and is obtained by the backward transmission signal. The
focusing effect is directly related to the correlation between According to (18), the TR echo y (t ) consists of the
the forward and backward channels. The signal is fully fo- focused signal z1 (t ), disturbance signal z2 (t ) due to target
cused on the target when the channel correlation coefficient movement, and fusion noise w(t ). The time domain discrete
ρ is 1, that is, when the channel has not changed. z1 (t ) can sample y (n ) of the TR echo is utilized to detect a target
be expressed as y (n ) = z1 (n ) + z2 (n ) + w(n ), n = 0, . . . , N − 1 (24)

L
z1 (t ) = kρ |αl |2 ˜s∗ (−t )e− jωc βl τl where N is the total number of samples in the bandwidth B
l =1
(19) with the sampling frequency fs . z1 (n ) and z2 (n ) represent the
discrete forms of the focused signal z1 (t ) and the disturbance
= kρDL s∗ (−t )
term z2 (t ), respectively
where DL [34] highlights the multipath focusing effect

L

L z1 (n ) = kρ |αl |2 ˜s∗ (−n )e− jωc βl nτl / fs (25)
2 − jωc βl τl
DL = |αl | e . (20) l =1
l =1 
L 
L

1/ fc is a small value; hence, the phase term ωc βl τl = z2 (n ) = kαl (n ) ∗ s (−n − nτl )

(26)
2π fc βl τl changes by 2π rad with small target motions l =1 l  =1
jωc βl (−n−nτl )/ fs jωc βl  n/ fs
βl τl [34], indicating that DL is a random process. For the ×e e
actual complex scattering environment, a large number of
propagation paths always exist that are received from differ- where nτl denotes the start of the sample corresponding to
ent directions by the radar antenna. Consequently, according the time delay τl of the lth path and s (n ) is the discrete
to the central limit theorem [45], DL is regarded as a complex sample of the transmitted signal.
Gaussian random variable LEMMA 1: w(n ) denotes the discrete sample of noise w(t )
DL  CN (0, Lσα4 ). (21)  L 
 
L

z2 (t ) is the channel disturbance response that is caused by w(n ) = k ρhl (n ) + (n ) ∗ v(n )e jωc β l n + w(n )
l =1 l =1
the target motion and is given by
 L  ∼ CN 0, k Lσα σv +
22 2
σw2
L 
z2 (t ) = k (t ) ∗ ∗ ∗
αl s (−t − τl )e jωc βl (−t−τl )
e jωc βl  t . ∼ CN 0, σw2 (27)
l  =1 l =1
(22) where

w(t ) is the fusion noise composed of w(t ) and the noise σw2 = k 2 Lσα2 σv2 + σw2 . (28)
caused by v(t ), which is expressed as
LEMMA 2: The focused term z1 (n ) is a Gaussian random
 L 
 
L process based on the central limit theorem for a large
w(t ) = k ρhl (t ) + (t ) ∗ v(t )e jωc β l t + w(t ) number of paths
l =1 l =1
(23) z1 (n )  CN (0, k 2 ρ 2 Lσα4 |s (−n )|2 ). (29)

ZHANG ET AL.: MOVING TARGET DETECTION BASED ON TIME REVERSAL IN A MULTIPATH ENVIRONMENT 3225

Authorized licensed use limited to: SHIV NADAR UNIVERSITY. Downloaded on October 09,2024 at 07:21:18 UTC from IEEE Xplore. Restrictions apply.
LEMMA 3: The disturbance term z2 (n ) is distributed as Here, the matrix STR is composed of the transmitted signal
  data, and the lth column is the transmitted signal with the

L
z2 (n )  CN 0, Lk 2 σα2 σ 2 |s∗ (−n − nτl )|2 (30) time delay of the lth propagation path
l =1 STR = [s0 , s1 , . . . , sL ]N×(L+1) (39)
since the channel disturbance (n ) follows a complex Gaus-
sian distribution (n ) ∼ CN (0, σ 2 ). sl = [s (−0 − nτl ), . . . , s (−(N − 1) − nτl )]T . (40)
We note that the disturbance term also contains target In particular, nτl is zero, corresponding to l = 0.  is a
and environmental information. The disturbance term influ- diagonal matrix
ences the focusing effect of the TR transmitted signal, and it
 = diag {[λ0 , λ1 , . . . , λL ]} (41)
gives rise to channel mismatch. When the disturbance term
is known, it can still provide environmental information and where
be useful for detection. Therefore, in the method described k 2 ρ 2 Lσα4 , l=0
ahead, the disturbance term is not eliminated but instead λl = (42)
Lk 2 σα2 σ 2 , l = 1, . . . , L.
applied to the detector. And z1 (n ), z2 (n ), and w(n ) are
independent of each other. The covariance matrix CTR is given by

PROPOSITION 1: For each time domain sample, the echo CTR = S∗TRSTTR + σw2 IN . (43)
statistics are given by Under the null hypothesis H0 , only noise is included in the
y (n )  CN (0, (n )) (31) echo data, which are distributed as
where the variance (n ) is expressed as y ∼ CN (0, σw2 IN ). (44)
(n ) = k 2 ρ 2 Lσα4 |s (−n )|2 Based on the matrix inversion lemma [46], the inversion
matrix C− 1
TR can be obtained

L
+L k 2 σα2 σ 2 |s∗ (−n − nτl )|2 + σw2 . (32) 1 1
C−
TR =
1
I − 2 S∗TR (σw2  −1 + STTR S∗TR )−1 STTR . (45)
l =1 σw2
σw
The probability density functions (pdfs) of the TR echo data
The vector form of the TR echo is written as are described as
y = [y (0), . . . , y (N − 1)]T . (33) 1
p(y|H1 ) = N exp −yH C− 1
TR y (46)
We define the signal response z (n ) as the sum of z1 (n ) and π |CTR |
z2 (n ), and the vector form of z (n ) is given by 1 yH y
p(y|H0 ) = exp − . (47)
z = [z (0), . . . , z (N − 1)]T . (34) (πσw2 )N σw2
The LRT can be expressed as
The fusion noise vector and additive noise vector, respec-
tively, are written as p(y|H1 )
(y) = . (48)
w = [w(0), . . . , w(N − 1)]T (35) p(y|H0 )
In the derivation of the TR–LRT detector, we assume that
and
the Doppler frequency factor, time delay, amplitude of the
w = [w(0), . . . , w(N − 1)]T . (36) propagation paths, and variance of the channel disturbance
are known in advance and that they can be regarded as
The TR detection problem with channel variation can
constant terms. In a realistic scenario, these parameters
be described as a binary hypothesis test
are unknown and need to be estimated. The estimation
H1 : y = z + w processes are introduced in the following section.
(37)
H0 : y = w . THEOREM 1: The TR–LRT detector is derived by taking the
When the target exists, i.e., under the alternative hypothesis logarithm of (48) and, then, discarding the constant terms
H1 , the echo data in the vector form are distributed as σw2 − σw2
  TR–LRT (y) = yH I + S∗TR − 1 T
TR STR y
L
σw2 σw2
4 ∗ T ∗ T
y ∼ CN 0, k ρ Lσα s0 s0 + Lk σα σ
2 2 2 2 2
s l s l + σw I N
2
= yH  TR y (49)
l =1
 

L where
∼ CN 0, λl sl∗ slT + σw2 IN σw2 − σw2
l =0  TR = I + S∗TR − 1 T
TR STR (50)
σw2 σw2
∼ CN 0, S∗TR STTR + σw2 IN
and
∼ CN (0, CTR ) .
(38)  TR = σw2 σw2  −1 + STTR S∗TR . (51)

3226 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 57, NO. 5 OCTOBER 2021

Authorized licensed use limited to: SHIV NADAR UNIVERSITY. Downloaded on October 09,2024 at 07:21:18 UTC from IEEE Xplore. Restrictions apply.
REMARK 1: Although the correlation coefficients are uni- The pdfs of the conventional echo data under hypotheses
formly replaced by ρ, it does not affect the detection result. H1 and H0 are
Because according to the echo variance (n ) and the matrix 1
 in detectorTR–LRT , they all contain Lρ 2 that can all be p(x|H1 ) = exp −xH C− 1
CO x (63)
π N |CCO |
replaced by Ll=1 ρl2 , and the detector does not involve ρ
except for  . 1 xH x
p(x|H0 ) = exp − . (64)
(πσv2 )N σv2
B. Conventional Likelihood Ratio Test
The LRT is defined as
The discrete version of the conventional echo can be
p(x|H1 )
expressed as CO–LRT = . (65)
p(x|H0 )

L
x (n ) = αl s (n − nτl )e jωc βl (n−nτl )/ fs + v(n ) Similar to the TR–LRT detector, the time delay, the Doppler
l =1
(52) frequency shift, and the amplitude of each propagation
path are assumed to be known. The estimation of these
= r (n ) + v(n ), n = 0, . . . , N − 1
parameters is introduced in the following.
where r (n ) and v(n ) are discrete samples of the target
response signal r (t ) and noise v(t ), respectively. THEOREM 2: Taking the logarithm of (65) and discarding
the constant terms, the conventional detector is derived as
LEMMA 4: The statistics for the target response signal r (n ) −1
based on the central limit theorem with a large number of 1 σv2
CO–LRT = 2 xH SCO IL + SH
CO SCO SH
CO x
propagation paths are given by σv σα2 (66)
 L  =x H
SCO − 1 H
 CO SCO x
r (n )  CN 0, σα |s (n − nτl )| .
2 2
(53)
where
l =1
 CO = σv2 σv2 /σα2 IL + SH
CO SCO . (67)
The vector forms of the data can be written as
C. Parameter Estimation
x = [x (0), . . . , x (N − 1)]T (54)
In a more realistic scenario, the time delay of each prop-
r = [r (0), . . . , r (N − 1)]T (55) agation path, the reflection coefficient, and the disturbance
v = [v(0), . . . , v(N − 1)]T . (56) related to the motion of the target are always unknown.
However, these parameters are essential in target detection.
Conventional detection is a binary hypothesis test problem We estimate these parameters to calculate the test statistics.
that is described as 1) Time Delay and Doppler Estimation: There are
H1 : x = r + v many time-delay estimation algorithms, and the traditional
(57) AMF method is always used. The conventional echo after
H0 : x = v .
multiple snapshot accumulations is convolved with s∗ (−t )
The additive noise v is independent of the signal r. There- to realize a matched filter. Then, high-amplitude peak points
fore, under alternative hypothesis H1 , we obtain are found in front of L, which correspond to the propagation
x ∼ CN (0, CCO ) (58) path time delay. However, the estimation accuracy of this
method is inadequate. A novel Doppler-delay estimation
where method based on the Kumaresan–Tufts (KT) algorithm
CCO = σα2 SCO SH suitable for a multipath environment has been proposed by
CO + σv IN
2
(59)
Friedlander [47] to estimate the time delay τˆl and Doppler
and frequency β̂l .
SCO = [s1 , . . . , sL ]N×L . (60) 2) Amplitude Estimation: After obtaining the estima-
tion of the time delay and Doppler frequency, we can
Each column vector sl in SCO is given by determine the attenuation coefficients of the propagation
sl = [s (0 − nτl )e jωc βl (0−nτ1 )/ fs , . . . , paths that are represented in the following vector form:
(61)
s (N − 1 − nτl )e jωc βl (N−1−nτl )/ fs ]T . a = [α1 , . . . , αL ]T . (68)
We note that SCO and STR have similar forms, except The frequency shift term e jωc βl (n−nτl )/ fs in (52) can be written
for the dimensions. Hence, we can adopt a similar approach in a matrix form for N samples and L paths
to address the TR detection problem. Based on the matrix ⎡ ⎤
inversion lemma [46], the inversion matrix C− 1 e jωc β̂1 (0−nτ1 )/ fs ··· e jωc β̂L (0−nτL )/ fs
CO is derived ⎢ ⎥
.. ..
as =⎢ ⎣ .
..
. .
⎥.
⎦
−1
1 1 σv2 e jωc β̂1 (N−1−nτ1 )/ fs · · · e jωc β̂L (N−1−nτL )/ fs
C− 1
= 2 IN − 2 SCO IL + SH
CO SCO CO . (62)
SH
CO
σv σv σα2 (69)

ZHANG ET AL.: MOVING TARGET DETECTION BASED ON TIME REVERSAL IN A MULTIPATH ENVIRONMENT 3227

Authorized licensed use limited to: SHIV NADAR UNIVERSITY. Downloaded on October 09,2024 at 07:21:18 UTC from IEEE Xplore. Restrictions apply.
The conventional echo data of (54) can be recast as
Algorithm 1: Computation of the TR Test Statistic
x = (SCO   ) a + v. (70) TR–LRT .
Require: The conventional echo x (n );
The maximum likelihood requires a solution
  The TR echo y (n );
â = arg min ||x − (SCO   ) a||2 . (71) The transmitted signal s (n );
a
Ensure:The TR test statistic TR–LRT ;
PROPOSITION 2: The estimation of a is 1: Compute SCO by (60) and  by (69);
 −1 2: Estimate the amplitude αl by (72);
â = (SCO   )H (SCO   ) (SCO   )H x. (72)
3: Estimate the disturbance covariance σ 2 by (73),
3) Disturbance Estimation: The channel disturbance and compute the correlation coefficient ρ by
(n ) can be solved by deconvolution; however, we analyze (76);
the TR–LRT detector and note that (n ) is not directly 4: Compute  TR by (50);
utilized to calculate the test statistics, but its variance σ 2 5: Compute TR–LRT = yH  TR y;
is used. Thus, we need to estimate only the channel distur- 6: return TR–LRT .
bance variance σˆ2 .
PROPOSITION 3: The TR echo consists of z1 (n ), z2 (n ), and
w(n ). Based on (14) and (31), the channel disturbance Algorithm 2: Computation of the Conventional Test
variance is Statistic CO–LRT .
Require: The conventional echo x (n );
Var y (n ) − Var (z1 (n )) − σw2 The transmitted signal s (n );
σˆ2 = (73)
Var (z2 (n )) L − Var (z1 (n )) /(Lσα̂2 ) Ensure:The conventional test statistic CO–LRT ;
where 1: Compute SCO by (60) and  by (69);
2: Estimate the amplitude αl by (72);

L
3: Compute  CO by (67);
z 1 (n ) = k |α̂l |2 ˜s∗ (−n )e− jωc β̂l nτ̂l / fs (74)
4: Compute CO–LRT = xH SCO − 1 H
CO SCO x;
l =1
5: return CO–LRT .

L
z 2 (n ) = k α̂l∗ e− jωc nτ̂l / fs s∗ (−n − nτ̂l ). (75)
l =1
that matrix  TR is also a Hermitian matrix. In addition,  TR
With the estimated σα̂2 and σ ˆ2 , the estimation of the is a nonsingular matrix, implying that  − 1
TR is a Hermitian
correlation coefficient is straightforward matrix. Furthermore, we know that  TR is also a Hermitian
 matrix. Therefore,  TR can be written as the product of a
σ2
ρ̂ = 1 − ˆ 2 . (76) matrix and its corresponding Hermitian transpose. Conse-
Lσα̂ quently, the result of singular value decomposition (SVD)
REMARK 2: Although the detector uses the variance of the of  TR is obtained
amplitude term σα2 instead of αl , it still needs to use an
 TR = UTR TR UTR
H
accurate amplitude term for each path when calculating the (77)
variance of the disturbance term σ 2 . = PTR
H
PTR .

D. Algorithm Flowchart Then, we rewrite the test statistic in (49)

Algorithms 1 and 2 show the steps involved in evaluating
TR–LRT (y) = yH PTR
H
PTR y
the TR and conventional test statistics, respectively. (78)
= ||PTR y||2 .
IV. PERFORMANCE ANALYSIS
Under the alternative hypothesis H1
In this section, we study the performance of the TR
and conventional detectors. The pdfs of the test statistics of PTR y
two detectors are developed under the null and alternative ⎡ ⎤⎡ ⎤
hypotheses. On the one hand, we analyze the detection p(TR)
0 ,0 ··· p(TR)
0,N−1 z (0) + w(0)
⎢ . .. ⎥⎢ .. ⎥
probability, which is the most important index for the =⎢
⎣
.. ..
. .
⎥⎢
⎦⎣ .
⎥
⎦
evaluation of the detector performance. On the other hand,
the normalized J-divergence [48] can be used to exhibit p(TR)
N−1,0 · · · p (TR)
N−1,N−1 z ( N − 1) + w( N − 1)
⎡ ⎤
mathematical insight and tractable methods. N−1 (TR)
n=0 p0,n (z (n ) + w(n ))
⎢ .. ⎥
A. TR–LRT Detector Performance =⎢
⎣ .
⎥
⎦
N−1 (TR)
The matrix STTR S∗TR is a Hermitian matrix, and diagonal p
n=0 N−1,n ( z ( n ) + w( n ))
matrix  −1 is a positive definite matrix. Hence, it is obvious (79)

3228 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 57, NO. 5 OCTOBER 2021

Authorized licensed use limited to: SHIV NADAR UNIVERSITY. Downloaded on October 09,2024 at 07:21:18 UTC from IEEE Xplore. Restrictions apply.
and the test statistic can be written as can obtain the following:

PCO x
1
N−
⎡ ⎤⎡ ⎤
TR–LRT (y) = |dTR,m | 2
(80) p(CO) · · · p(CO) r (0) + v(0)
1,0 1,N−1
m=0 ⎢ . .. ⎥ ⎢ .. ⎥
=⎢⎣
.. ..
. .
⎥⎢
⎦⎣ .
⎥
⎦
where |dTR,m |2 is a noncentral Chi-square distributed vari- p(CO)
L,0 · · · p (CO)
L,N−1 r ( N − 1) + v( N − 1)
able with two degrees of freedom ⎡ ⎤
N−1 (CO)
n=0 p1,n (r (n ) + v(n ))
⎢ .. ⎥
 N− 2 =⎢⎣
⎥.
⎦
  (TR)
1 .
 N−1 (CO)
|dTR,m |2 =  pm,n (z (n ) + w(n ))  p ( r ( n ) + v( n ))
n=0 L,n
n=0 (86)
N−1 N−1 (TR)
n=0 |p(TR)
m,n | σw 2 |
2 2
pm,n z (n )|2 Consequently, CO–LRT can be written as
∼ X2 N−n=0
1 (TR) 2 2
n=0 |pm,n | σw /2
2

L
m = 0, . . . , N − 1. CO (x) = |dCO,l |2 (87)
(81) l =1

where |dCO,l |2 is drawn from the noncentral Chi-square

COROLLARY 1: Under H1 , the test statistic is distributed as distribution

TR–LRT,1 1
N−

 1 (TR) 2 2 N−1 (TR) |dCO,l |2 = p(CO)

l,n (r (n ) + v(n ))
1 N−
N−
n=0 |pm,n | σw 2 | pm,n z (n )|2 (82) n=0
∼ X2 N−n=0 . N−1  1 (CO)
1 TR |p(CO)
l,n | σv
2 2
| N−
n=0 pl,n r (n )|
2
n=0 |pm,n | σw /2
2 ( ) 2 2
n=0
m=0 ∼ X22 N−1 (CO) 2 2
n=0 |pl,n | σv /2
2
Under the null hypothesis H0 , the target signal is zero; thus, l = 1, . . . , L.
we have (88)

1  N−
N− 1

2

COROLLARY 2: CO–LRT is the sum of L independent non-
TR–LRT,0 =  TR )
p(m,n w(n ) central Chi-square random variables. Subsequently, we can
m=0 n=0 obtain
N−1 (83)
1
N−
n=0 |p(m,n | σv 2
TR ) 2 2
CO–LRT,1
∼ X2 (0).
2  L N−1  1 (CO)
m=0
(CO) 2 2
n=0 |pl,n | σv | N−
n=0 pl,n r (n )|
2
(89)
∼ X2 N−1 (CO)
2
.
n=0 |pl,n | σv /2
2 2 2
With the test statistics, the detection probability Pd,TR l =1

and the threshold ηTR are derived in Appendix B. Under the null hypothesis H0 , CO–LRT,0 is distributed as

 L N−1 (CO) 2 2
B. Conventional Detector Performance n=0 |pl,n | σv
CO–LRT,0 ∼ X22 (0). (90)
2
The matrix  CO is a Hermitian matrix and is invertible, l =1
which means that its inverse is also a Hermitian matrix. The derivation of the detection probability Pd,CO and the
Hence, we can perform SVD of  − CO
1
threshold ηCO are described in Appendix C.

−
CO = UCO CO UCO
1 H
C. Normalized J-Divergence
(84)
= CO .
 CO H Although the detection probability Pd and the false
probability Pfa can directly reflect the performance of the
The test statistics can be recast as detectors, it is difficult to calculate these terms by deriving
theoretical formulas. To study the performance of the detec-
CO–LRT (x) = xH SCO CO H H tors, we utilize the normalized J-divergence [31], [32], [48]
CO SCO x
= || H
CO SCO x||
H 2
(85) |E {|H1 } − E {|H0 }|2
ψ= . (91)
= ||PCO x||2 . 1
(Var{|H1 } + Var{|H0 } )
2
Under the alternative hypothesis H1 , the echo data con- COROLLARY 3: According to the statistical properties of
sist of the target signal and additive noise. Therefore, we the noncentral Chi-square distribution, we obtain E {} and

ZHANG ET AL.: MOVING TARGET DETECTION BASED ON TIME REVERSAL IN A MULTIPATH ENVIRONMENT 3229

Authorized licensed use limited to: SHIV NADAR UNIVERSITY. Downloaded on October 09,2024 at 07:21:18 UTC from IEEE Xplore. Restrictions apply.
Var{} under H1 and H0 for the TR–LRT detector to (107) and (110), it is apparent that the actual noise power
does not appear in the pdf under H0 . Therefore, the proposed
E {TR–LRT |H1 }
 1 (TR) 2 2   1 (TR)  TR–LRT detector possesses the CFAR property.
1 N−
N−
n=0 |pm,n | σw | N−
n=0 pm,n z (n )|
2
= 2 + N−1 (TR)
n=0 |pm,n | σw /2
2 2 2
m=0
(92) V. NUMERICAL SIMULATIONS
1 N−
N− 1 In this section, we use Monte Carlo simulations based
E {TR–LRT |H0 } = |p(TR)
m,n | σw ,
2 2
on 200/Pfa independent trials to assess the performance
m=0 n=0 of the proposed TR–LRT and CO–LRT detectors along
(93) with that of the ED [30], TR–GLRT [30] and TR (no
Var{TR–LRT |H1 } disturbance) detectors [34]. The ED detector is essentially
 N−1    1 (TR)  a conventional generalized likelihood ratio detector. The
1
N− (TR) 2 2 2
n=0 |pm,n | σw | N−
n=0 pm,n z (n )|
2
TR–GLRT detector assumes that the target is stationary and
= 4 + 4 N−1 (TR)
that the forward and backward channels are consistent. The
n=0 |pm,n | σw /2
2 2 2
m=0
(94) TR (no disturbance) method does not consider the channel
variation due to target motion; this method considers only
and the Doppler frequency. In addition, some other non-TR
 N−1 2 techniques that exploit multipath are used to compare the
1
N−
n=0 |p(TR)
m,n | σw
2 2
Var{TR–LRT |H0 } = 4 . (95) proposed detectors. Kelly’s detector derived in [49] is a
m=0
2 GLRT detector. The algorithm T-AMF is introduced in [21],
and MP-AMF is proposed in [22]. A snapshot process of
COROLLARY 4: For the conventional detector, the statistical
TR includes two transmissions, i.e., forward transmission
characteristics of CO–LRT can be described as
and backward transmission, whereas each snapshot of the
E {CO–LRT |H1 } conventional detection process includes only one transmis-
L N−1
  1 (CO)  sion. Therefore, in order to ensure fairness, the number of
 (CO) 2 2
n=0 |pl,n | σv | N−
n=0 pl,n r (n )|
2
= 2 + N−1 (CO) snapshots of the conventional detection is twice that of the
n=0 |pl,n | σv /2
2 2 2
TR detection. The simulation steps are as follows.
l =1
(96) 1) Generate the initially transmitted signal s (t ) and chan-
nel response h (t ) containing multiple paths with random
 1
L N−
E {CO–LRT |H0 } = |p(CO) time delays and amplitudes.
l,n | σv
2 2

l =1 n=0
2) Generate complex white Gaussian noise v(t ) and
(97) obtain the conventional echo data x (t ) with the specified
SNR. The SNR is defined as
Var{CO–LRT |H1 }
 N−1 (CO)    1 (CO) 
 2 2 2
L
n=0 |pl,n | σv | N−
n=0 pl,n r (n )|
2
= 4 + 4 N−1 (CO) Es σα2
2 n=0 |pl,n | σv /2
2 2 SNR = . (100)
l =1 σv2
(98)
and
 N−1 2 3) Reverse the conventional signal x (t ) in the time

L
n=0 |p(CO)
l,n | σv
2 2
domain and normalize the energy. The processed data are
Var{CO–LRT |H0 } = 4 . (99)
2 the TR transmitted signal.
l =1
4) Obtain the TR echo by the convolution of the TR
With the expected values and variances in the statistics transmitted signal kx ∗ (−t ) with the backward channel re-
of TR–LRT and CO–LRT, we can calculate the normalized sponse h (t ) and add noise w(t ) with the same power as that
J-diversity of the two detectors. The specific numerical of v(t ).
simulations are presented in Section V. 5) Estimate the Doppler frequency factors by Wigner–
Hough transform [50], multipath delays by the modified
D. CFAR Performance
time–frequency correlation function based on the Wigner–
We analyze the CFAR property of the TR–LRT detector. Ville distribution [50], amplitudes of the paths, and the
Under hypothesis H0 , the received signal contains only variance of the channel disturbance.
noise and there is no multipath in the environment, i.e., 6) Calculate the test statistics of the conventional and
L = 0, which implies that STR is composed of one vector s0 . TR echoes and, then, compare them with the corresponding
Subsequently,  TR reduces to σw2 STTR S∗TR , and  TR simpli- detection threshold.
fies to 1/σw2 S∗TR (STTR S∗TR )−1 STTR . Hence, the elements p(TR)
m,n 7) Calculate the detection probability Pd .
of matrix PTR all have a common coefficient 1/σw , and the The transmitted signal is a linear frequency modulation
remain terms are independent on σw2 . Eventually, according (LFM) signal with bandwidth B = 200 MHz and timewidth

3230 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 57, NO. 5 OCTOBER 2021

Authorized licensed use limited to: SHIV NADAR UNIVERSITY. Downloaded on October 09,2024 at 07:21:18 UTC from IEEE Xplore. Restrictions apply.
T = 0.1 μs 2 . The sampling rate fs is 1 GHz. The carrier
frequency is 10 GHz. The velocity is set to be 1.5 m/s. Note
that we consider only the overlapping echoes and a dense
multipath scattering environment; hence, the delay of each
propagation path is generated randomly within the time
of a pulse width. The direction of arrival φl is randomly
distributed between 0 and 2π, which leads to a Doppler
frequency that is also a random number in the range of
−100 Hz to 100 Hz. The noise variance σv2 = σw2 = 1. We
study the following aspects of the detector performance:

1) the relationship between SNR and the corresponding

Pd under the different false-alarm probabilities Pfa ;
2) the effect of the different numbers of propagation
paths on the detection probability;
3) the influence of the channel disturbance on the de-
tection performance;
4) the influence of the PE.

Fig. 2 shows the detection probability versus the

SNR for the different false-alarm probabilities Pfa =
[10−4 , 10−3 , 10−2 ]. The TR–GLRT clearly outperforms
ED, with approximately 1-dB SNR gain, indicating that
TR can, indeed, exploit the multipath effect. Moreover, the
TR detectors have significant performance gains compared
with other non-TR detectors due to the focusing effect,
such as Kellys GLRT, T–AMF and MP–AMF in this sim-
ulation situation. However, TR–GLRT performs slightly
worse than CO–LRT with a channel correlation coefficient
of ρ = 0.35. This occurs because channel variation gives
rise to mismatch during the TR process, whereas CO–LRT
exploits multipath effects more effectively by estimating
the multipath parameters. Nevertheless, the curves suggest
that TR still offers advantages in multipath exploitation.
The TR detectors that considered target motion perform
significantly better than CO–LRT. Overall, the proposed
TR–LRT detector in this article is observed to be the most
effective because this detector takes channel variation into
account. The TR (no disturbance) method performs slightly
worse than TR–LRT since TR (no disturbance) does not
account for the influence of the channel changes, even
though it does consider the Doppler frequency. Remarkably,
when the SNR is low, for instance, for an SNR of less than
-7 dB in Fig. 2(c), the detection performance of TR–LRT is
slightly worse than that of the TR (no disturbance) detector.
This result is reasonable because TR–LRT achieves poor
channel PE accuracy for a low SNR, which indicates that Fig. 2. Detection probability versus SNR. (a) Pfa = 10−4 . (b)
the proposed TR–LRT detector is related to the PE perfor- Pfa = 10−3 . (c) Pfa = 10−2 . The number of paths L = 25. The variance
mance. of the multipath amplitude attenuation coefficient σα2 = 2. The
correlation coefficient ρ is set to 0.35.
Fig. 3 illustrates the influence of the number of prop-
agation paths on the detection probability. Fig. 3(a) shows which clearly indicates the effect of the number of prop-
the curves of the detection probability versus the SNR for agation paths on the detection performance. The simulation
different numbers of paths. Fig. 3(b) depicts the curves results show that the detection probability increases with
of the detection probability versus the number of paths, an increasing number of paths. This phenomenon indicates
that the multipath effect can be effectively exploited by the
2 Note
that the transmitted signal of the proposed algorithm is not limited two detectors we proposed. Nevertheless, these detectors
to LFM. Different transmitted signals correspond to different delay and have different multipath exploitation efficiencies. The TR
Doppler estimation methods, e.g., see [47] and [51]. technique can clearly make better use of multipaths than the

ZHANG ET AL.: MOVING TARGET DETECTION BASED ON TIME REVERSAL IN A MULTIPATH ENVIRONMENT 3231

Authorized licensed use limited to: SHIV NADAR UNIVERSITY. Downloaded on October 09,2024 at 07:21:18 UTC from IEEE Xplore. Restrictions apply.
 Fig. 4. Detection probability versus the correlation coefficient. The
SNR is 0 dB. The number of paths L = 20. The false-alarm probability
Pfa = 10−2 . The variance of the amplitude coefficient σα2 = 2.

Fig. 3. Detection probability for different numbers of propagation

paths. (a) Detection probability curves versus the SNR. (b) Detection Fig. 5. Influence of PE on detection performance. The number of paths
probability curves versus the number of paths. The digital suffixes in the L = 25. The false-alarm probability Pfa = 10−2 . The variance of the
legend denote the number of propagation paths. The false-alarm amplitude coefficient σα2 = 2 with the channel correlation coefficient
probability is set as Pfa = 10−2 . The variance of the amplitude coefficient ρ = 0.35.
σα2 = 2. The correlation coefficient ρ = 0.35.
the channel variation and its corresponding parameters to
conventional method since the focusing effect is utilized to the detection process in the TR–LRT that allows better sens-
improve energy. The TR–LRT detector derived in our article ing of environmental changes and improves the robustness
offers significantly better detection performance than the of the detector.
TR (no disturbance) detector proposed in [34] because the Fig. 5 depicts the detection performance of the two
changes in the scattering environment are taken into account detectors we proposed in the cases of channel PE and
in the TR–LRT detector and the multipath parameters are es- nonparametric estimation (NPE), respectively. In the case
timated to provide target information to improve detection, of NPE, the channel parameters are known prior and are
whereas the TR detector in [34] does not consider channel substituted into the detector. The parameters obtained by es-
variation and, thus, cannot adapt to a diffuse multipath timation cause loss of detection performance. Specifically,
environment. the performance loss of the TR detector is approximately
Fig. 4 shows the detection probability versus correlation 2 dB, which also demonstrates that the proposed detector is
coefficient ρ with SNR=0 dB, i.e., the influence of channel robust.
disturbance on the detection performance. The detection Additionally, we use normalized J-divergence to assess
performance of the TR (no disturbance) detector in [34] the detection performance of we proposed detectors. Fig. 6
is almost consistent with that of the TR–LRT detector depicts the normalized J-divergence versus the SNR for the
proposed in this article when the value of ρ approaches TR and conventional detectors. As expected, the detection
1. However, for small ρ, the detection performance of performance of TR–LRT and CO–LRT improves with in-
the TR (no disturbance) detector with a constant channel creasing SNR. However, as shown in Fig. 7, TR–LRT detec-
assumption in [34] is significantly worse. In contrast, the tion performance deteriorates when the channel correlation
TR–LRT detector still has a high-detection probability at is weak. The focusing effect becomes poor and the detection
ρ = 0.1, indicating the superior robustness of the TR–LRT performance is reduced due to channel mismatch, consistent
detector. This superior robustness is due to the addition of with the phenomenon shown in Fig. 4.

3232 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 57, NO. 5 OCTOBER 2021

Authorized licensed use limited to: SHIV NADAR UNIVERSITY. Downloaded on October 09,2024 at 07:21:18 UTC from IEEE Xplore. Restrictions apply.
 of the retransmitted signal kx ∗ (−t ) with the Doppler effect,
and the backward propagation channel h (t ) consists of the
target channel response and disturbance response, which is
derived as
y (t )
= e(t ) − hc (t ) ∗ kx ∗ (−t )

L
=ρ hl (t ) ∗ kx ∗ (−t )e jωc β l t
l =1

L
+ (t ) ∗ kx ∗ (−t )e jωc β l t + w(t )
l =1
Fig. 6. Normalized J-divergence versus the SNR. Variance of the L 
amplitude coefficient σα2 = 2, and Pfa is 10−2 . The variance in the channel 
L 
correlation coefficient ρ = 0.8. The number of propagation paths L is 20.
=ρ hl (t ) ∗ s∗ (−t )e− jωc βl  t ∗ hl∗ (−t ) + v(t ) ejωc β l t
l =1 l  =1
L 

L 
+ (t ) ∗ s∗ (−t )e− jωc βl  t ∗ hl∗ (−t ) + v(t ) ejωc β l t
l =1 l  =1
+ w(t )
 L 

L 
= kρ hl (t ) ∗ s∗ (−t )e− jωc βl  t ∗ hl∗ (−t ) e jωc β l t
l =1 l  =1
  
ya (t )
 L 

L 
+k (t ) ∗ s∗ (−t )e− jωc βl  t ∗ hl∗ (−t ) e jωc β l t
l =1 l  =1
  
yb (t )
Fig. 7. Normalized J-divergence versus the correlation coefficient with
SNR = 0 dB. The number of paths is 20, and Pfa is 10−2 . Variance in the 
L 
L

amplitude coefficient σα2 = 2. +k ρhl (t ) ∗ v(t )e jωc β l t + k (t ) ∗ v(t )e jωc β l t + w(t )

l =1 l =1
  
w(t )
VI. CONCLUSION
(101)
This article studies conventional and TR–LRTs for a
where
moving target in a diffuse multipath indoor environment
with channel variation. Both detectors can utilize the mul- ya (t )
tipath effect to improve the detection probability. However,  L 

L 
∗ − jωc βl t
the CO–LRT detector uses only multipath information to = kρ hl (t ) ∗ s (−t )e ∗ hl∗ (−t ) e jωc β l t
improve the detection performance, whereas the TR–LRT l =1 l  =1
detector not only exploits multipath information but also

L 
L
uses the TR technique to realize spatial and temporal fo- = kρ αl αl∗ s∗(−t + τl −τl  )e jωc βl  (−t +τl −τl ) e jωc βl (t−τl )
cusing to enhance the echo energy. We present analytical l =1 l  =1
expressions for the detection probability of the detectors and

L
employ normalized J-divergence to determine the perfor- = kρ |αl |2 ˜s∗ (−t )e− jωc βl τl
mance of the detectors. Monte Carlo simulations are carried l =1
out to analyze the detection probability. The TR detector   
ya1 (t )
is shown to perform significantly better than conventional
detectors, particularly for a severe multipath environment. 
L

Furthermore, the TR–LRT detector proposed in this article +k αl αl∗ s∗ (−t + τl − τl )e jωc βl  (−t +τl −τl ) e jωc βl (t−τl) .
is robust with respect to heavy disturbances due to target l=l 
  
motion and possesses the CFAR property. ya2 (t )
(102)
APPENDIX A
For a dense multipath scattering environment, the amplitude
DERIVATION OF THE TR ECHO
of ya2 (t ) is very small compared to that of ya1 (t ) [34], [35].
We derive the TR echo in this section. Similar to the con- Therefore, the influence of ya2 (t ) on the echo can be ignored
ventional echo, the TR echo is obtained by the convolution while only ya1 (t ) is retained. yb (t ) is caused by the channel

ZHANG ET AL.: MOVING TARGET DETECTION BASED ON TIME REVERSAL IN A MULTIPATH ENVIRONMENT 3233

Authorized licensed use limited to: SHIV NADAR UNIVERSITY. Downloaded on October 09,2024 at 07:21:18 UTC from IEEE Xplore. Restrictions apply.
disturbance, which is specifically described as as a noncentral Chi-square distribution with two degrees of
freedom xm,1 ∼ X22 (γm,1 ), for which the pdf is given by
yb (t )
 L  1 √

L  fxm,1 (xm,1 , γm,1) = e−(xm,1 +2)/2 I0 ( γm,1 xm,1 ), xm,1 > 0
∗ − jωc βl  t
=k (t ) ∗ s (−t )e ∗ hl∗ (−t ) e jωc β l t 2
(111)
l =1 l  =1
 L  where I0 (z ) is a modified Bessel function of the first kind

L  of order 0 [52]. xm,0 is a central Chi-squared distribution
=k (t ) ∗ αl∗ s∗ (−t − τl  )e jωc βl  (−t−τl  ) e jωc βl t
random variable with 2 degrees of freedom, which the pdf
l =1 l  =1
 L  is [52]

L 
=k (t ) ∗ αl∗ s∗ (−t − τl )e jωc βl (−t−τl ) e jωc βl  t . 1 −xm,0 /2
fxm,0 (xm,0 ) = e , xm,0 > 0. (112)
l  =1 l =1 2
(103) FTR,i (i )
Hence, the TR echo can be written as  i  ∞  ∞  
1 x0,i

L = ··· N−1 (TR) fx0,i (TR)
· · · fxN−2,i
y (t ) ≈ kρ |αl |2 ˜s∗ (−t )e− jωc βl τl −∞ 0
  0 m =0 Am,i A 0 ,i
l =1 N−1
 L     N−2 

L  xN−2,i TR,i − m=0 xm,i
+k (t ) ∗ αl∗ s∗ (−t − τl )e jωc βl (−t−τl )
e jωc βl  t
× fxN−1,i
l  =1 l =1 A(TR)
N−2,i
(TR)
AN−1,i
+ w(t ). dx0,i · · ·dxN−2,i di , i = 0, 1. (113)
(104)   
N−1

APPENDIX B Then, we calculate the cumulative density function (cdf)

of TR–LRT,1 and TR–LRT,0 that is written as (113), which is
DERIVATION OF THE DETECTION PROBABILITY AND
shown at the bottom of the page.
THRESHOLD FOR THE TR–LRT DETECTOR
Subsequently, we can obtain the detection threshold ηTR
We can rewrite (82) and (83) in the following unified by the inverse function
form:
ηTR = F−TR–LRT,
1
( Pf a ) (114)
1
N− 0

TR–LRT,i = A(TR) 2 (TR)

m,i X2 (γm,i ), i = 0, 1 (105) and the detection probability is given by
m=0
Pd,TR = FTR–LRT,1 (ηTR ). (115)
where
N−1 APPENDIX C
|p(TR)
m,n | σw
2 2
n=0
A(TR) = , m = 0, . . . , N − 1 (106)
m,1
2 DERIVATION OF THE DETECTION PROBABILITY AND
N−1 (TR) 2 2 THRESHOLD FOR THE CO–LRT DETECTOR
n=0 |pm,n | σw
A(TR)
m,0 = , m = 0, . . . , N − 1 (107)
2 The test statistics CO–LRT of (89) and (90) can be rewrit-
 1 (TR) ten as
| N− pm,n z (n )|2
γm,1 = N−n=0
(TR)
, m = 0, . . . , N − 1 (108)
1 (TR) 2 2 
L
n=0 |pm,n | σw /2 CO–LRT,i = (CO)
ACO,l X22 (γl,i ), i = 0 , 1 (116)
(TR)
γm, 0 = 0, m = 0, . . . , N − 1. (109) l =1

fTR,i where
 ∞  ∞   N−1
1 x0,i n=0 |p(CO)
l,n | σv
2 2
= ··· N−1 (TR) fx0,i (TR)
· · · fxN−2,i ACO,l = , l = 1, . . . , L (117)
A A 2
  
0 0 m =0 m,i 0 ,i N−1
N−1 | n=0 p(CO)
l,n r (n )|
2
   N−2  γl,(CO)
1 = N−1 (CO) , l = 1, . . . , L (118)
TR,i − n=0 |pl,n | σv /2
xN−2,i 2 2
m=0 xm,i
× fxN−1,i
A(TR)
N−2,i A(TR)
N−1,i γl,(CO)
0 = 0, l = 1, . . . , L (119)
dx0,i · · · dxN−2,i , i = 0, 1 (110) FCO,i (i )
    i  ∞  ∞
N−1 1 x1,i
= ··· L fx1,i · · · fxL−1,i
For the sum of two independent random variables ∞ z= −∞ 0 A ACO,1
   0 l =1 CO,l
x1 + x2 , the pdf is described as fz (z ) = −∞ fx1 (z − L−1
x2 ) fx2 (x2 )dx2 . The pdf of y = a · x is written as fy (y ) =   1 
(1/|a| ) fx (y/a ). Hence, under hypotheses H1 and H0 , the xL−1,i CO,i − L−l =1 xl,i
× fxL,i
pdf of TR–LRT is derived as (110), where xm,1 is distributed ACO,L−1 ACO,L

3234 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 57, NO. 5 OCTOBER 2021

Authorized licensed use limited to: SHIV NADAR UNIVERSITY. Downloaded on October 09,2024 at 07:21:18 UTC from IEEE Xplore. Restrictions apply.
 dx1,i · · · dxL−1,i di , i = 0, 1. (120) [14] H. T. Hayvaci, A. De Maio, and D. Erricolo
   Improved detection probability of a radar target in the presence
L−1 of multipath with prior knowledge of the environment
IET Radar, Sonar, Navigat., vol. 7, no. 1, pp. 36–46,
We note that the forms of (105) and (116) are the same 2013.
except for the number of accumulated terms. Therefore, [15] P. Setlur, G. E. Smith, F. Ahmad, and M. G. Amin
similar to (110) and (113), we can obtain the cdf of CO–LRT Target localization with a single sensor via multipath exploita-
that is shown as (120). Then, the detection threshold ηCO tion
IEEE Trans. Aerosp. Electron. Syst., vol. 48, no. 3,
and the detection probability Pd,CO are derived as pp. 1996–2014, Jul. 2012.
[16] P. Setlur, T. Negishi, N. Devroye, and D. Erricolo
ηCO = F−–LRT,
1
0
( Pf a ) (121) Multipath exploitation in non-LOS urban synthetic aperture
radar
Pd,CO = FCO–LRT,1 (ηCO ). (122)
IEEE J. Sel. Topics Signal Process., vol. 8, no. 1, pp. 137–152,
Feb. 2014.
REFERENCES [17] P. Setlur and N. Devroye
Bayesian and Cramer-Rao bounds for single sensor target lo-
[1] S. L. Wilson and B. D. Carlson
calization via multipath exploitation
Radar detection in multipath
in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process., 2013,
IET Radar, Sonar, Navigat., vol. 146, no. 1, pp. 45–50, 1999.
pp. 5845–5849.
[2] Y. Jang, H. Lim, and D. Yoon
[18] X. Zeng, M. Yang, B. Chen, and Y. Jin
Multipath effect on radar detection of nonfluctuating targets
Estimation of direction of arrival by time reversal for low-angle
IEEE Trans. Aerosp. Electron. Syst., vol. 51, no. 1, pp. 792–795,
targets
Jan. 2015.
IEEE Trans. Aerosp. Electron. Syst., vol. 54, no. 6,
[3] P. Setlur, M. Amin and F. Ahmad
pp. 2675–2694, Dec. 2018.
“Multipath model and exploitation in through-the-wall and
[19] X. Zeng, F. Zhang, B. Wang, and K. J. R. Liu
urban radar sensing
Radio frequency based direction sensing using massive MIMO
IEEE Trans. Geosci. Remote Sens., vol. 49, no. 10,
IEEE Access, vol. 8, pp. 26827–26838, 2020.
pp. 4021–4034, Oct. 2011.
[20] F. C. Robey, D. R. Fuhrmann, E. J. Kelly, and R. Nitzberg
[4] Y. Yang, D. J. Feng, X. S. Wang and S. P. Xiao
A CFAR adaptive matched filter detector
Effects of k distributed sea clutter and multipath on radar
IEEE Trans. Aerosp. Electron. Syst., vol. 28, no. 1, pp. 208–216,
detection of low altitude sea surface targets
Jan. 1992.
IET Radar, Sonar, Navigat., vol. 8, no. 7, pp. 757–766, 2014.
[21] A. Aubry, A. De Maio, G. Foglia, and D. Orlando
[5] R. Cardinali, F. Colone, C. Ferretti, and P. Lombardo
Diffuse multipath exploitation for adaptive radar detection
Comparison of clutter and multipath cancellation techniques for
IEEE Trans. Signal Process., vol. 63 no. 5, pp. 1268–1281,
passive radar
Mar. 2015.
in Proc. IEEE Nat. Radar Conf., 2007, pp. 469–474.
[22] U. Kumbul and H. T. Hayvaci
[6] Z. Zhao, X. Wan, Q. Shao, Z. Gong, and F. Cheng
Multipath exploitation for knowledge-aided adaptive target de-
Multipath clutter rejection for digital radio mondiale-based HF
tection
passive bistatic radar with OFDM waveform
IET Radar, Sonar, Navigat., vol. 13, no. 6, pp. 863–870, 2019.
IET Radar, Sonar, Navigat., vol. 6, no. 9, pp. 867–872, 2012.
[23] J. Shi, G. Hu, H. Zhou, and M. Chen
[7] Z. Zhao, X. Zhou, S. Zhu, and S. Hong
Novel detection model for MIMO radar in multipath environ-
Reduced complexity multipath clutter rejection approach for
ment
DRM-Based HF passive bistatic radar
in Proc. IEEE Chin. Guid., Navigat., Control Conf., 2016,
IEEE Access, vol. 5, pp. 20228–20234, 2017.
pp. 1525–1529.
[8] R. L. Fante
[24] S. Sen and A. Nehorai
Cancellation of specular and diffuse jammer multipath using a
Adaptive OFDM radar for target detection in multipath scenar-
hybrid adaptive array
ios
IEEE Trans. Aerosp. Electron. Syst., vol. 27, no. 5, pp. 823–837,
IEEE Trans. Signal Process., vol. 59, no. 1, pp. 78–90,
Sep. 1991.
Jan. 2011.
[9] R. L. Fante and J. A. Torres
[25] M. Fink, C. Prada, F. Wu, and D. Cassereau
Cancellation of diffuse jammer multipath by an airborne adap-
Self focusing in inhomogeneous media with time reversal
tive radar
acoustic mirrors
IEEE Trans. Aerosp. Electron. Syst., vol. 31, no. 2, pp. 805–820,
in Proc. IEEE Ultrason. Symp., 1989, pp. 681–686.
Apr. 1995.
[26] M. Fink
[10] Y. Cao, S. Wang, Y. Wang, and S. Zhou
Time reversal of ultrasonic fields—Part I: Basic principles
Target detection for low angle radar based on multi-frequency
IEEE Trans. Ultrason. Ferroelect. Freq. Control, vol. 39, no. 5,
order-statistics
pp. 555–566, Sep. 1992.
J. Syst. Eng. Electron., vol. 26, no. 2, pp. 267–273, 2015.
[27] G. Lerosey, J. De Rosny, A. Tourin, A. Derode, G. Montaldo, and
[11] L. Fang, X. Wan, G. Fang, F. Cheng, and H. Ke
M. Fink
Passive detection using orthogonal frequency division multiplex
Time reversal of electromagnetic waves
signals of opportunity without multipath clutter cancellation
Phys. Rev. Lett., vol. 92, no. 19, pp. 19–21, 2004.
IET Radar, Sonar, Navigat., vol. 10, no. 3, pp. 516–524, 2016.
[28] Y. Jin and J. M. F. Moura
[12] D. O. Walsh
Time-reversal detection using antenna arrays
Performance analysis of the incoherent and skewness matched
IEEE Trans. Signal Process., vol. 57, no. 4, pp. 1396–1414,
filter detectors in multipath environments
Apr. 2009.
IEEE J. Ocean. Eng., vol. 20, no. 1, pp. 80–84, Jan. 1995.
[29] J. M. F. Moura and Y. Jin
[13] H. Zhou, G. Hu, J. Shi, and Y. Xiao
Time reversal imaging by adaptive interference canceling
Weighted target detector via estimated SNR in a multipath
IEEE Trans. Signal Process., vol. 56, no. 1, pp. 233–247,
environment
Jan. 2008.
J. Eng., vol. 20, pp. 7119–7124, 2019.
ZHANG ET AL.: MOVING TARGET DETECTION BASED ON TIME REVERSAL IN A MULTIPATH ENVIRONMENT 3235

Authorized licensed use limited to: SHIV NADAR UNIVERSITY. Downloaded on October 09,2024 at 07:21:18 UTC from IEEE Xplore. Restrictions apply.
[30] J. M. F. Moura and Y. Jin [46] X. Zhang
Detection by time reversal: Single antenna Matrix Analysis and Applications. Beijing, China: Tsinghua
IEEE Trans. Signal Process., vol. 55, no. 1, pp. 187–201, Univ., 2004.
Jan. 2007. [47] B. Friedlander
[31] O’ N. Donoughue, and J. M. F. Moura An efficient parametric technique for Doppler-delay estimation
Gaussian target detection in multipath clutter with single- IEEE Trans. Signal Process., vol. 60, no. 8, pp. 3953–3963,
antenna time reversal Aug. 2012.
IEEE Trans. Signal Process., vol. 61, no. 15, pp. 3733–3744, [48] E. Fishler, A. Haimovich, R. Blum, D. Chizhik, L. Cimini, and R.
Aug. 2013. Valenzuela
[32] Y. Jin, J. M. F. Moura, and O’ N. Donoughue Spatial diversity in radars-models and detection performance
Time reversal in multiple-input multiple-output radar IEEE Trans. Signal Process., vol. 54, no. 3, pp. 823–838,
IEEE J. Sel. Topics Signal Process., vol. 4, no. 1, pp. 210–225, Mar. 2006.
Feb. 2010. [49] E. J. Kelly
[33] Z. Zhang, B. Chen, and M. Yang. An adaptive detection algorithm
“Time reversal detection in multistatic radar system IEEE Trans. Aerosp. Electron. Syst., vol. AES-22, no. 1,
IEEE Access, vol. 8, pp. 93594–93607, 2020. pp. 115–127, Mar. 1986.
[34] Y. Jin, J. M. F. Moura, O’ N. Donoughue, and J. Harley [50] X. Tian, G. Liao, and Y. Wu
Single antenna time reversal detection of moving target Joint estimation of Doppler and multipath time delay of over-
in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process., 2010, lapping echoes for LFM pulse radar
pp. 3558–3561. in Proc. CIE Int. Conf. Radar Proc., 2001, pp. 829–833.
[35] Z. Zhang, B. Chen, M. Yang, and H. Xu [51] M. D. Hahm, Z. I. Mitrovski, and E. L. Titlebaum
Moving target detection of array antennas based on time reversal Deconvolution in the presence of Doppler with application to
in Proc. IEEE Sensor Array and Multichannel Signal Process. specular multipath parameter estimation
Workshop, 2020, pp. 1–5. IEEE Trans. Signal Process., vol. 45, no. 9, pp. 2203–2219,
[36] P. Setlur and N. Devroye Sep. 1997.
An information theoretic take on time reversal for nonstationary [52] N. L. Johnson and S. Kotz
channels Continuous Univariate Distributions, 2nd ed. New York, NY,
IEEE Signal Process. Lett., vol. 20, no. 4, pp. 327–330, Apr. USA: Wiley, 1970, ch. 28.
2013. Zhaoming Zhang (Graduate Student Member,
[37] S. Sen and A. Nehorai IEEE) received the B.E. degree in electronic
Adaptive OFDM radar for target detection in multipath scenar- and information engineering from Xidian Uni-
ios versity, Xi’an, China, in 2017, where he has
IEEE Trans. Signal Process., vol. 59, no. 1, pp. 78–90, been working toward the Ph.D. degree in signal
and information processing with the National
Jan. 2011.
Laboratory of Radar Signal Processing since
[38] J. M. Bourgeois and G. S. Smith September 2017.
A fully three-dimensional simulation of a ground-penetrating His current research interests include time-
radar: FDTD theory compared with experiment reversal technology, radar target detection, mul-
IEEE Trans. Geosci. Remote Sens., vol. 34, no. 1, pp. 36–44, tistatic radar, and automatic target recognition.
Jan. 1996.
Baixiao Chen was born in Anhui, China, in
[39] F. Foroozan and A. Asif 1966. He received the graduate degree in elec-
Time-reversal ground-penetrating radar: Range estimation with tronic engineering from the Metallurgy College
Cramer-Rao lower bounds of East China, Ganzhou, China, in 1987, and
IEEE Trans. Geosci. Remote Sens., vol. 48, no. 10, the master’s degree in circuit and system and
pp. 3698–3708, Oct. 2010. the Ph.D. degree in signal and information pro-
[40] H. T. Hayvaci, P. Setlur, N. Devroye, and D. Erricolo cessing from Xidian University, Xi’an, China, in
Maximum likelihood time delay estimation and Cramér-Rao 1994 and 1997, respectively.
bounds for multipath exploitation He was with the Metallurgy College of East
China, Ganzhou, China from July 1987 to July
in Proc. IEEE Radar Conf., 2012, pp. 764–768.
1991. Since then, he has been a Faculty Member
[41] H. T. Hayvaci, A. De Maio, and D. Erricolo with the National Laboratory of Radar Signal Processing, Xi’an, where
Diversity in receiving strategies based on time-delay analysis he was a Lecturer from October 1997 to May 1999 and an Associate
in the presence of multipath Professor from June 1996 to May 2003. In 2006, he was selected into
in Proc. IEEE Radar Conf., 2011, pp. 1040–1045. “New Century elitist Support Program” of the ministry of education. He
[42] D. Liu, J. Krolik, and L. Carin is currently a Professor in signal and information processing. His general
Electromagnetic target detection in uncertain media: Time- research interests include radar signal processing, new radar system design,
reversal and minimum-variance algorithms array signal processing, precise guidance, etc.
IEEE Trans. Geosci. Remote Sens., vol. 45, no. 4, pp. 934–944, Minglei Yang (Member, IEEE) received the
Apr. 2007. B.E. degree in electronic engineering and the
[43] D. Liu, S. Vasudevan, J. Krolik, G. Bal, and L. Carin Ph.D. degree in signal and information process-
Electromagnetic time-reversal source localization in changing ing from Xidian University, Xi’an, China, in
2004 and 2009, respectively.
media: Experiment and analysis
Since 2009, he has been with the National
IEEE Trans. Antennas Propag., vol. 55, no. 2, pp. 344–354, Laboratory of Radar Signal Processing, Xidian
Feb. 2007. University, and is currently an Associate Pro-
[44] H. L. Van Trees fessor. From 2014 to 2015, he was a Visiting
Detection, Estimation, and Modulation Theory, Part III: Radar- Scholar of the Elisha Yegal Bar-Ness Center for
Sonar Signal Processing and Gaussian Signals in Noise, New Wireless Communications and Signal Process-
York, NY, USA: Wiley, 2001. ing Research (CWCSPR), New Jersey Institute of Technology. He have
[45] P. Billingsley authored/coauthored more than 60 peer-reviewed journals and conference
Convergence of Probability Measures. New York, NY, USA: papers and more than 40 inventions. His research interests include array
signal processing, MIMO signal processing, and polarization information
Wiley, 1968.
processing.

3236 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 57, NO. 5 OCTOBER 2021

Authorized licensed use limited to: SHIV NADAR UNIVERSITY. Downloaded on October 09,2024 at 07:21:18 UTC from IEEE Xplore. Restrictions apply.