Download as pdf or txt
Download as pdf or txt
You are on page 1of 8

Signal Processing 217 (2024) 109353

Contents lists available at ScienceDirect

Signal Processing
journal homepage: www.elsevier.com/locate/sigpro

Short communication

Adaptive detection with training data in partially homogeneous


environments for colocated MIMO radar
Can Huang a, Yong-Liang Wang a, *, Weijian Liu b, *, Jun Liu c, Qinglei Du b
a
Electronic Information School, Wuhan University, Wuhan, China
b
Wuhan Electronic Information Institute, Wuhan, China
c
Department of Electronic Engineering and Information Science, University of Science and Technology of China, Hefei, China

A R T I C L E I N F O A B S T R A C T

Keywords: The issue of adaptive detection is considered for the colocated multi-input multi-output (MIMO) radar in this
Multi-input multi-output radar paper. Meanwhile, the background is partially homogeneous environments (PHE), where the power mismatch is
Partially homogeneous environments present between the training data and test data. The training data in PHE are utilized to derive effective detectors
Gradient test
on the basis of generalized likelihood ratio test, Rao, Wald, Gradient, and Durbin tests. Since the Durbin test-
Durbin test
based detector coincides with the Rao test-based detector, the two-step design approach for the Durbin test is
used for deriving the new detector. The outcomes of simulation experiments illustrate that the proposed detectors
achieve superior effectiveness to existing approaches. Furthermore, the results also show that when the signal
mismatch exists, the Wald and Durbin tests maintain robust characteristics. Meanwhile, the Rao test ensures
selective property.

1. Introduction environments. A two-step GLRT is devised in [12] to achieve colocated


MIMO radar detection in compound Gaussian clutter scenarios. Addi­
Multi-input multi-output (MIMO) radar has attracted growing con­ tionally, some performance evaluations for colocated MIMO radar
cerns in contemporary radar techniques [1,2]. By using multiple chan­ detection are provided in [13].
nels for transmitting and receiving, the MIMO radar provides Most of the above-mentioned detectors are derived without utilizing
uncorrelated signals and achieves better performance in target detec­ training data. However, in adaptive detection, the data adjacent to the
tion. In recent decades, there have been various studies on colocated cell under test (CUT) can be used as training sample data for improving
MIMO radar, such as image processing [3], location estimation [4], the detection performance [14–19]. Many works have been investigated
waveform analysis [5] and adaptive detection [6,7]. Specifically, a by utilizing the test data and training data collectively to design adaptive
generalized likelihood ratio test (GLRT) is derived in [6] on the basis of detectors. For instance, adaptive detectors on the basis of various design
the MIMO platform. For the identical issue, the Rao and Wald test-based criteria are derived in [20] by utilizing training data, which shows su­
detectors are devised in [7]. By using the Bayesian approach, several perior performance to those without training data. However, the
detectors are proposed in [8] resorting to various detection design aforementioned approaches are only considered in homogeneous envi­
criteria with a priori knowledge being incorporated. In [9] tunable de­ ronments (HE). In realistic situations, because of instrumental factors,
tectors in MIMO radar, including the Rao and Wald tests as special cases, the environments are not always homogeneous. To address the detection
are proposed. The directivity characteristics of the tunable detectors can problem in different nonhomogeneous environments, numerous de­
be flexibly adjusted by changing the tuning parameter. Furthermore, by tectors have been proposed. For instance, in fully-heterogeneous envi­
exploring the persymmetric structure of the covariance matrix, a ronments where covariance matrices can change between test and
GLRT-based detector is derived in [10] to deal with target detection in training data, a comparative analysis is provided in [21] for evaluating
MIMO radar. Besides, a persymmetric Rao test-based detector is devel­ the detectors by utilizing IPIX radar datasets in terms of the constant
oped in [11] by exploiting the persymmetric structure of the disturbance false alarm rate (CFAR) behaviour and probability of detection (PD).
covariance matrix for MIMO radar in Gaussian disturbance Moreover, to deal with the detection issue in heterogeneous

* Corresponding authors.
E-mail addresses: huangcan575@163.com (C. Huang), ylwangkjld@163.com (Y.-L. Wang), liuvjian@163.com (W. Liu), junliu@ustc.edu.cn (J. Liu), dql822@163.
com (Q. Du).

https://doi.org/10.1016/j.sigpro.2023.109353
Received 23 August 2023; Received in revised form 21 November 2023; Accepted 4 December 2023
Available online 7 December 2023
0165-1684/© 2023 Elsevier B.V. All rights reserved.
C. Huang et al. Signal Processing 217 (2024) 109353

environments caused by power variations of different fluctuating terrain is provided in Section 2. Details of the detector design are supplied in
types and strong clutter outliers, a clustering-CFAR detector is proposed Section 3. The experimental findings are presented in Section 4. The
in [22]. Furthermore, the heterogeneous scenarios caused by interfer­ conclusion is drawn in Section 5.
ence power variations between consecutive samples are investigated in
[23], and the maximum likelihood approach is proposed to achieve 2. Problem formulation
effective detection. In [24] a GLRT-like detector is derived for coherent
target detection in heterogeneous environments. Furthermore, resorting It is hypothesized that the uniform linear array (ULA) transmitter
to the Rao and Wald tests, the detectors based on the hoc design pro­ and receiver with M and N elements, respectively, are the components of
cedure in [25] are adopted in heterogeneous clutter scenarios for the colocated MIMO radar system. Meanwhile, the waveforms contain K
moving target detection. Another type of nonhomogeneity is partially sampling points. For the binary hypothesis testing, under hypothesis H0 ,
homogeneous environments (PHE), in which the covariance matrices of the N × K -dimensional test data matrix X only contains the noise matrix
the test and training data are identical but the power levels are different N = [n1 , n2 , ⋯, nK ], where nk , k = 1, 2, …, K, is the noise modeled as a
[26]. Meanwhile, the PHE assumptions precisely reflect the surround­ zero-mean complex circular Gaussian vector with an unknown covari­
ings for ground radar. It can also be employed in the detection issue with ance matrix Tt . To estimate Tt , we assume that there exist training data
interference from electronic countermeasures [27–29]. The problems in zl , l = 1, 2, …, L, containing the noise el which has the identical
PHE have received much consideration lately and numerous detectors structure to nk but with a covariance matrix T. In HE, Tt = T. In
have been developed. For instance, in [30] several detectors resorting to contrast, in PHE there exists the power mismatch [26], denoted as σ2 , i.
different design criteria are derived for detecting double subspace tar­
e., Tt = σ2 T. Z = [z1 , z2 , ⋯, zL ] and E = [e1 , e2 , ⋯, eL ] are the N × L
gets in PHE. Moreover, an adaptive coherent estimator is developed in
-dimensional training data and noise matrices, respectively. Under hy­
[31,32] to reduce the requirement for the training sample in PHE.
pothesis H1 , the target signal is denoted as [6,7]
Apart from the design criteria mentioned above, another two com­
mon design criteria are Gradient and Durbin tests which have been G = αs ar aHt S, (1)
applied in adaptive detection in recent years. For instance, to address the
problem of clutter with symmetrical ordered power spectral density, where αs is the unknown amplitude, ar and at are the known receiving
Gradient and Durbin test-based detectors are developed in [33]. More­ and transmitting steering vectors, respectively, and S is the known
over, the Gradient and Durbin test-based detectors are provided in [34] waveform matrix. The dimensions of ar , at and S are N × 1, M × 1 and
to deal with the problem of structural nonhomogeneity, and these two M × K, respectively. Hence, the detection issue is stated as
detectors are shown to be robust to signal mismatch. On the basis of the ⎧

Gradient test, two detectors are derived in [35] for distributed targets ⎪



{
X = N,

under the Gaussian noise background. Due to the excessive number of ⎪


H 0 :
Z = E.
unknowns, it is essential to note that the optimal detectors do not exist { (2)
for the MIMO radar detection. Hence, we utilize the GLRT, Rao, Wald, ⎪

⎪ X = αs ar aHt S + N,

⎪ H 1 :
Gradient and Durbin tests to derive detectors with training data in PHE. ⎪

⎪ Z = E.

The experimental outcomes indicate that the proposed approaches
attain superior performance to the methods without training data.
It is worth pointing out that the model for the detection problem in 3. Derivations of detectors
this paper is a special case of [30] from the viewpoint of mathematics.
However, the detection model in this paper provides a specific appli­ 3.1. GLRT-based detector
cation scenario based on the colocated MIMO radar platform. Further­
more, there are some significant differences between this paper and that The joint probabilities of density function (PDFs) of X and Z under
in [30]. First, apart from the GLRT, Rao and Wald tests, new detectors hypotheses H1 and H0 are
based on the Gradient and Durbin tests are derived, which ensure rela­ { }
tively superior robust characteristics when the signal mismatch exists. ( ) 1 [( )H ( )]
exp − tr T− 1 W − 2 tr X − αs ar aHt S T− 1 X − αs ar aHt S
Second, the derivation procedures of the detectors in this paper are f1 (X, Z) =
σ
different from those in [30]. Third, in the aspects of performance πN(L+K) σ 2NK |T|L+K
assessment, we analyze the detection performance under different (3)
power mismatch parameters and the effectiveness in the cases of inter­
fering targets and clutter edges [36–38], which have not been investi­ and
gated in [30]. These evaluations provide more comprehensive {
( ) 1 ( )
}
improvements for the issues of colocated MIMO radar adaptive detec­ exp − tr T− 1 W − 2 tr XH T− 1 X
(4)
σ
tion. The innovations of the paper are two aspects. Firstly, five detectors f0 (X, Z) = ,
π N(L+K) σ2NK |T|L+K
are proposed to deal with the issue of MIMO radar adaptive detection in
PHE when training data are available, and the proposed approaches
respectively, where W = ZZH is the sample covariance matrix. Nulling
achieve better performance than other competitors. Secondly, the
the differential of the natural logarithm of (3) with respect to (w.r.t.) T
directivity characteristics of the proposed approaches to signal
yields the maximum likelihood estimation (MLE) of T as
mismatch are analyzed. It is found that the Wald test-based and Durbin [ ]
test-based detectors possess robust characteristics, while the Rao ̂1 = 1 1( )( )H
T W + 2 X − αs ar aHt S X − αs ar aHt S . (5)
test-based detector ensures selective characteristics. L+K σ
The remaining part is arranged as follows. The problem formulation
Taking (5) into (3) yields

1( )H ( )− (L+K)
̂ 1 ) = λσ −
f1 (X, Z; T 2NK
|W|− (L+K)
|IK + X − αs ar aHt S W− 1 X − αs ar aHt S | , (6)
σ 2

2
C. Huang et al. Signal Processing 217 (2024) 109353

where λ = [(L + K)/(eπ)]N(L+K) and IK is the K × K -dimensional identity


∂lnf1 (X, Z) T [ − 1 ̂ ] ∂lnf1 (X, Z)
matrix. Nulling the differential of the natural logarithm of (6) w.r.t. αs tRao = | F ( Θ 0 ) Θr ,Θr |Θ=Θ̂0 , (14)
∂Θr ∂Θ∗r
yields the MLE of αs as Θ=Θ̂0

T
aHr W− 1 XSH at where Θ = [ΘTr , ΘTs ] is the parameter vector, Θr = αs , Θs =
α
̂s = . (7)
aHt SSH at aHr W− 1 ar [σ2 , vecT (T)]T , Θ
̂ 0 is the MLE of Θ under H0 and F(Θ) is the Fisher in­
Taking (7) into (6) yields formation matrix (FIM) for complex-valued signals, shown as [39]

̂ 1, ̂
f1 (X, Z; T α s ) = λσ− 2NK |W|− (L+K)
( )H ( )− (L+K)
1 aH W− 1 XSH at aH W− 1 XSH at
⋅|IK + 2 X − H r H H − 1 ar aHt S W− 1 X − H r H H − 1 ar aHt S |
σ at SS at ar W ar at SS at ar W ar
(∼ )H ( ∼ ) (8)
1 ̃ SH a ̃ SH a |− (L+K)
= λσ − 2NK |W|− (L+K) |IK + 2 X − P̃ar XP X − P̃ar XP
σ t t

1 (∼H ⊥ ̃ ̃H ̃ ⊥ |
)− (L+K)
= λσ − 2NK
|W|− (L+K)
⋅|IK + X P̃a X + P ⊥
S H X P XP H
a ̃ S a .
σ2 r t ar t

SH a aH S ∼ 1/2 1/2
where P⊥
SH at
= IK − PSH at , PSH at = aH SSt Ht at , ar = W− ar , X
̃ = W− X, P̃⊥ [ ]
t r
a
H ∼H ∼
∂lnf1 (X, Z) ∂lnf1 (X, Z)
= IN − P̃ar and P̃ar = ̃
ar ̃
ar /(ar ar ). F(Θ) = E
∂Θ∗

∂ΘT
2
The MLE of σ under H1 , denoted as ̂ σ 21 , is the sole solution to [26] [ ] (15)
FΘr ,Θr (Θ) FΘr ,Θs (Θ)
= .

r1
ξk1 NK FΘs ,Θr (Θ) FΘs ,Θs (Θ)
= , (9)
ξ + σ 2 L +K
It can be derived from (3) that
k1 =1 1k

∼H ∼ ∼H ∼ ∂lnf1 (X, Z) 1 H ( )H
where r1 = rank(X P⊥
∼ X + P H X P∼ XP H ), and ξk is the k1 -th non-
⊥ ⊥
1
= 2 at S X − αs ar aHt S T− 1 ar , (16)
a S at ar S at
r ∂αs σ
∼H ∼ ∼H ∼
zero eigenvalue of X P∼⊥ X + P⊥
SH at
X P∼ar XP⊥
SH at
.
a r ∂lnf1 (X, Z) 1 H − 1 ( )
Nulling the derivative of (4) w.r.t. T yields the MLE of T with fixed σ 2 = 2 ar T X − αs ar aHt S SH at . (17)
∂α∗s σ
as
( ) Setting αs = 0 in (16) and (17), we have
1 1
̂0 =
T W + 2 XXH . (10) [
∂lnf1 (X, Z) ∂lnf1 (X, Z)
]
K+L σ FΘr ,Θr (Θ) = E ⋅

∂αs ∂αs
Plugging (10) into (4) yields
1 [ ]
= 4 aHr T− 1 E XSH at aHt SXH T− 1 ar (18)
1 ∼H ∼
̂ 0 ) = λσ −
f0 (X, Z; T 2NK
|W|− (L+K)
|IK + X X |− (L+K)
. (11) σ
σ 2
1
= aHt SSH at ⋅aHr T− 1 ar ,
The MLE of σ under H0 , denoted as σ
2
̂ 20 , is the sole solution to [26] σ2
r0
∑ λk0 NK where E[XSH at aH H 2 H H
t SX ] = σ at SS at ⋅T is utilized. From [20], we know
= , (12) that FΘr ,Θs (Θ) is a null matrix. Thus, it can be derived that
λ + σ2 L + K
k0 =1 k0
( )− 1
[ −1 ] [ ]− 1 1 H H
where r0 is the smaller value between N and K, λk0 is the k0 -th eigenvalue F (Θ) Θr ,Θr = FΘr ,Θr (Θ) = a
2 t
SS at ⋅aH − 1
r T ar . (19)
σ
∼H ∼
of X X.
Plugging (16), (17) and (19) into (14), and setting αs = 0 yields the
Performing the (L +K) th root of the ratio between (8) and (11),
Rao test with fixed T and σ2 as
along with ̂ σ 21 and ̂σ 20 , yields the GLRT with training data in PHE (GLRT-
( )− 1
TD-PHE) as 1 1 1
tRaoT,σ2 = 2 aHt SXH T− 1 ar ⋅ 2 aHt SSH at ⋅aHr T− 1 ar ⋅ 2 aHr T− 1 XSH at
⃒ ⃒ σ σ σ
(20)
NK ⃒
( 2 )K+L 1 ∼H ∼⃒
̂σ 0 ⃒⃒IK + 2 X X⃒⃒ 2
σ0
̂ |aHr T− 1 XSH at |
tGLRT− TD− PHE = ⃒ ( )⃒. (13) = 2 H H H − 1
.
NK ⃒
( 2 )K+L 1 ∼H ⊥ ∼ ∼H ∼ ⃒ σ at SS at ⋅ar T ar
σ1 ⃒ ⊥ ⊥ ⃒
̂ ⃒IK + ̂ 2 X P∼ar X + PSH at X P∼a r XPSH at ⃒
σ1
σ 20 into (20) yields the Rao test with training data in
Taking (10) and ̂
PHE (Rao-TD-PHE) as
3.2. Rao test-based detector

The Rao test is denoted as [39]

3
C. Huang et al. Signal Processing 217 (2024) 109353

( )− 1 2
1 { ̂ − 1 XSH at }
|aHr W + 2 XXH XSH at | 1 H H ̂− 1 aH T
σ0
̂ tGradient− a SX T 0 ar ⋅ r 1
= Re
(21)
TD− PHE 2 t
tRao− TD− PHE = ( )− 1 . σ0
̂ ̂ − 1 ar
aHt SSH at aHr T
1 1
2 H H
σ 0 at SS at ⋅ar W + 2 XXH
̂ H
ar ⎧ ( )− 1 ⎫
σ0
̂ ⎪
⎪ H 1 1
̂ XSH at ⎪

⎪ (30)
⎪ H
⎨at SX W + ̂ 2
XXH ar ⋅aHr T 1 ⎪

σ0
= Re − 1
.

⎪ σ 20 aHt SSH at ⋅aHr T
̂ ̂ ar ⎪

3.3. Wald test-based detector ⎪
⎩ 1 ⎪

The Wald test is denoted as [39]


3.5. Durbin test-based detector
{[ ] }− 1
tWald = ( Θ ̂ 1)
̂ r1 − Θr0 )H F− 1 ( Θ ̂ r1 − Θr0 ),
(Θ (22)
Θr ,Θr
The Durbin test is denoted as [41]
{ [ −1 ] }
where Θ ̂ r1 is the MLE of Θr under H1 , Θr0 is the value of Θr under H0 . ̂ r01 − Θr0 )H [F( Θ
tDurbin = ( Θ ̂ 0 )] ̂ ̂ ̂
Θr Θr F ( Θ 0 ) Θr Θr [F( Θ 0 )]Θr Θr ( Θ r01 − Θr0 ),
Then, we attain (31)
{[ ] }− 1 1
F− 1 (Θ) Θr ,Θr = 2 aHt SSH at ⋅aHr T− 1 ar . (23) where Θ̂ r01 = argmax f1 (X,Z; Θr , Θ ̂ s0 is the MLE of Θs under H0 . It
̂ s0 ), Θ
σ Θr
can be derived that
The MLE of αs with fixed T is [20]
− 1
̂ XSH at
aHr T
aH T− 1 XSH a
̂ T = H r H H − 1t .
α (24)
̂ r01 =
Θ 0
. (32)
at SS at ar T ar ̂ − 1 ar
aH SSH at ⋅aH T
t r 0

Plugging (23) and (24) into (22) yields the Wald test with fixed T and Plugging (18), (19) and (32) into (31) yields the Durbin test-based
σ 2 as detector with fixed σ2 as

aHt SXH T− 1 ar 1 H H aH T− 1 XSH at ̂ − 1 ar


aHt SXH T 1 H ̂− 1 H
tWaldT,σ2 = H H H − 1
⋅ 2 at SS at ⋅aHr T− 1 ar ⋅ H r H H − 1 tDurbinσ2 = 0 ̂ − 1 ar ⋅ ar T 0 XS at
⋅ 2 aHt SSH at ⋅aHr T 0
at SS at ar T ar σ at SS at ar T ar H H Ĥ −
at SS at ⋅ar T 0 ar
1
σ H H H ̂− 1
at SS at ⋅ar T 0 ar
2
(25) (33)
|aHr T− 1 XSH at | ̂ XSH at | − 1 2
= 2 H H H − 1
. |aHr T 0
σ at SS at ar T ar = .
̂ − 1 ar
σ a SSH at ⋅aH T
2 H
t r 0
Taking (7) and σ 21 into (5) yields the MLE of T under H1 as

[ ( )( )H ]
̂1 = 1 1 aH W− 1 XSH at aH W− 1 XSH at
T W + 2 X − H r H H − 1 ar aHt S X − H r H H − 1 ar aHt S
L+K σ1
̂ at SS at ar W ar at SS at ar W ar
{ [ ( − 1) ( )]}
1 1 H
ar ar W W− 1 ar aHr
= W + 2 XP⊥ H X
S at
H
+ I N − XP H X
S at
H
I N − (26)
L+K σ1
̂ aHr W− 1 ar aHr W− 1 ar
[ ( ) ]
1 1 1 ∼ ∼H ∼ ∼H 1
= W2 IN + 2 XP⊥ SH at X + P̃

XPSH at X P̃⊥ W2 .
L+K σ1
̂ ar ar

σ 21 into (25) yields the Wald test with training data


Inserting (26) and ̂ Taking (10) and ̂
2
σ 0 into (33), and eliminating the constant yields the
in PHE (Wald-TD-PHE) as Durbin test with training data in PHE (Durbin-TD-PHE) as
− 1 2 ( )− 1 2
̂ XSH at |
|aHr T 1
tWald− TD− PHE = 1
. (27) |aHr W + 2 XXH XSH at |
̂ − 1 ar σ0
̂
σ 2 aH SSH at aH T
̂ 1 t r 1 tDurbin− TD− PHE = ( )− 1 , (34)
2 H H 1
σ 0 at SS at ⋅ar W + 2 XXH
̂ H
ar
3.4. Gradient test-based detector σ0
̂

The Gradient test is denoted as [40] which coincides with the Rao-TD-PHE.
{ } Another method is the two-step approach [26]. Precisely, the de­
∂lnf1 (X, Z; Θ) tector is derived by claiming that the covariance matrix is known.
tGradient = Re | ( ̂ r1 − Θr0 ) ,
Θ (28)
∂ΘTr Θ=Θ̂0
Subsequently, it is replaced with a proper estimation.
Nulling the derivative of (4) w.r.t. σ 2 yields the MLE of σ2 under H0
It can be derived from (16) that with fixed T as
∂lnf1 (X, Z; Θ) 1 ( )/
̂ − 1 ar .
|Θ=Θ̂0 = 2 aHt SXH T 0 (29) σ 2T = tr XH T− 1 X NK.
̂ (35)
∂ΘTr σ0
̂
Inserting (35) into (33), eliminating the constant and substituting T
σ 20 into (28) yields the Gradient
Plugging (10), (24), (26), (29) and ̂ by W yields the two-step Durbin test with training data in PHE (2S-
test-based detector in PHE (Gradient-TD-PHE) as Durbin-TD-PHE) as

4
C. Huang et al. Signal Processing 217 (2024) 109353

Fig. 1. PD against SNR in PHE.

Fig. 2. PD against cos2 ϕ in PHE.

|aH W− 1 XSH at |
2
Durbin-TD-PHE ensures the CFAR characteristics with respect to σ2 .
t2S− = ( H − 1r ) H H (36)
σ 21 and ̂
σ 20 can be expressed as ̂
σ 21 = σ 2 λ′k0 and ̂
σ 20 = σ2 ξ′k1 ,
.
Durbin− TD− PHE
tr X W X ⋅at SS at ⋅aHr W− 1 ar Furthermore, ̂
respectively, where λ′k0 and ξ′k1 are the eigenvalues of the terms
⌢H ⌢ − 1 ⌢ ⌢H ⌢ ⌢H ⌢
3.6. CFAR property X0 W X0 and X 0 P⊥
⌢ X0 + P H X P⌢ X0 P H , respectively. If ς0 and ς1
a

S at 0 a r0

S at
r0

are the solutions to the equations


To evaluate the CFAR property of the proposed detectors with

r0
λ′k0 NK
respect to σ2 , the term XH W− 1 X can be expressed as ′
= (38)
k0 =1 λ k0 + ς
L+K
(37)
⌢H ⌢ − 1 ⌢
XH W− 1 X = σ 2 X0 W X0 ,
and
⌢ ⌢ ⌢ ⌢
1/2 − 1/2 − 1/2
where X0 = σ and W = T
T− X
WT . Each column of X0 is ∑
r1
ξ′k1 NK
= , (39)
distributed as a zero-mean complex circular Gaussian random vector, ′
k1 =1 ξ k1 + ς
L+K

with a covariance matrix IN . Meanwhile, W is ruled by the complex
central Wishart distribution with an associated covariance matrix IN . respectively, then, ς0 σ 2 and ς1 σ 2 are the solutions to (12) and (9),
− 1 − 1 ⌢H ⌢ − 1 ⌢ respectively. Plugging ς0 σ 2 and ς1 σ2 into the detection statistics of the
Similarly, the terms aH H
r W X and ar W ar can be recast as σ 2 a r0 W X0 GLRT-based, Rao test-based, Wald test-based and Gradient test-based
⌢H ⌢ − 1 ⌢
detectors cancels out the dependence on σ 2 . Hence, all the proposed
1/2
and σ2 a r0 W a r0 , respectively, where a r0 = . Then, inserting
⌢ T− ar
σ
these terms into (36) cancels out the dependence on σ 2 . Hence, the 2S- detectors ensure the CFAR characteristics with respect to σ 2 .

5
C. Huang et al. Signal Processing 217 (2024) 109353

Fig. 3. PD against the non-homogeneous environment parameter σ2 .

Fig. 4. PD against SNR in the presence of interfering targets. Fig. 5. PFA against distance of the clutter edge position from the CUT for N =
24 and the nominal value of PFA = 10− 3 .
4. Performance evaluation
Moreover, we let N = 12, M = 8, K = 2N, and σ2 = 2. 106 and 107 trials
Monte Carlo experiments are utilized to verify the effectiveness of are utilized to determine the PD and detection threshold, respectively.
proposed approaches. For comparison, the detectors without training The signal-to-noise ratio (SNR) in PHE is defined as
data in [6,7] are also present. The detection statistics of detectors in [6,
1
7] are expressed as SNR = ⋅|αs |2 ⋅aHr T− 1 ar ⋅aHt SSH at . (43)
σ2
|XXH | Besides, the signal mismatch is considered, where the actual steering
tGLRT = ( )
(40)
⃒ ⃒ H
vector is distinct from the presumed one [42]. A quantity to measure the
H
⃒ ⊥ ⃒ aHt SX P⊥ ar XS at
⃒XPSH at XH ⃒ 1 +
aHt SSH at amount of signal mismatch, denoted as cos2 ϕ, is
2
( )− 1 |aHr T− 1 ar,0 |
|aH XXH XSH at |
2
cos2 ϕ = , (44)
(41)
H − 1 H − 1
tRao = ( r H )− 1 , ar T ar ⋅ar,0 T ar,0
aHr XX ar ⋅aHt SSH at
where ar,0 is the actual received signal steering vector. The lower degree
( )− 1
cos2 ϕ is, the more serious signal mismatch becomes.
2
|aHr
XP⊥ SH a t X
H
XSH at |
tWald = ( )− 1 , (42) The PDs as functions of the SNRs in PHE are displayed in Fig. 1. It is
aHr XP⊥
SH at X
H
ar ⋅aHt SSH at shown that the proposed detectors with training data provide higher PDs
than those without training data in [6,7]. Furthermore, the GLRT-based
− 1/2 detector nearly achieves identical performance to the Gradient
where X = (XP⊥ XH ) H H
ar = IN − Par , Par = ar ar /(ar ar ) and ar =
X, P⊥
SH at test-based detector, which provides slightly superior effectiveness to
H − 1/2
(XP⊥
SH at
X ) ar . The probability of false alarm (PFA) is set to be 10− 5 . other competitors. Similar trends are present in Fig. 1(b) that with the
increase of training data, the performance of detectors with training data

6
C. Huang et al. Signal Processing 217 (2024) 109353

with the performance in the absence of interfering targets in Fig. 1, all


detectors suffer from performance loss due to the increasing uncertainty
when the interfering targets exist. Furthermore, it is seen that the pro­
posed detectors still provide superior performance to the approaches
without training data owing to the better estimation precision of the
covariance matrix.
The curves of PFAs against the distance d are displayed in Fig. 5,
where the parameter d denotes the distance between the clutter edge
and CUT. The clutter edge can be modeled as an abrupt clutter power
transition [36]. It is simulated that the clutter samples prior to the
clutter edge maintain the power P1 , while the samples after the clutter
edge ensure the power P2 which is lower than P1 . The clutter-to-noise
ratio (CNR), defined as P1 /P2 [36], is set to be 20 dB. d < 0 indicates
that the CUT is in the clutter with lower power. It is observed that the dip
in PFA at d = − 1 is present when the CUT is in the lower power clutter,
which leads to the degradation of PFA. In contrast, the significant in­
crease of PFA appears at d = 0 when the CUT is in the heavy power
clutter. Furthermore, when d > 0, the proposed detectors with training
Fig. 6. PD versus K for the cases of L = 2N and L = 4N, SNR = 13dB. data achieve better PFA regulations than other competitors.
To analyze the effect of the waveform matrix on selecting the
training data, the PDs versus different dimensions K of the waveform
matrix for the cases of L = 2N and L = 4N are depicted in Fig. 6. The
results reveal that the proposed detectors achieve higher PDs with the
increase of dimensions K. Meanwhile, with more training data, the
detection performance of detectors will enhance significantly. Hence, to
achieve effective signal detection for a waveform matrix with lower
dimensions K, more training data are required.
The counter-tests for the case of HE are depicted in Fig. 7. Precisely,
five proposed detectors in our work are utilized to achieve effective
detection in HE and compared with the existing methods in [20] which
were originally designed for HE on the basis of the MIMO platform. It is
discovered that apart from the 2S-Durbin-TD-PHE, the proposed de­
tectors in PHE provide close PDs to existing competitors in HE, which
verifies the effectiveness of the proposed approaches.

5. Conclusion

The issue of adaptive detection in PHE was investigated on the basis


of the colocated MIMO radar platform. Five detectors were proposed
Fig. 7. PD against SNR for the case of HE. based on different design criteria. The outcomes illustrated that the
Durbin-TD-PHE coincides with the Rao test-based detectors. In contrast,
will improve. the two-step Durbin tests and detectors on the basis of other one-step
Fig. 2 plots the PDs versus various cos2 ϕ in PHE. As the signal design criteria in PHE are different. Numerical experiments indicate
mismatch becomes more severe, the PDs of the Rao-TD-PHE decrease that the proposed methods in PHE achieve superior performance to
significantly, indicating that the Rao test-based detector attains more detectors without training data. Besides, the Wald test-based and Durbin
selective characteristics. In contrast, the Wald-TD-PHE and 2S-Durbin- test-based detectors are robust to signal mismatch, while the Rao test-
TD-PHE provide higher PDs than other detectors, verifying their based detector maintains selective characteristics. For future work, the
strongly robust characteristics. Furthermore, the directivity character­ specific approaches which deal with the detection issue in the presence
istics of the proposed Gradient-TD-PHE and GLRT-TD-PHE are balanced, of interfering targets and clutter edges deserve further investigations.
even in the presence of adequate training data.
Fig. 3 depicts the curves of PDs as functions of the non-homogeneous CRediT authorship contribution statement
parameters σ2 in PHE. The experimental findings indicate that when the
power mismatch σ2 increases, the PDs of all detectors with training data Can Huang: Investigation, Methodology, Writing – original draft.
will decline. It is because the estimation precision of the covariance Yong-Liang Wang: Conceptualization, Supervision. Weijian Liu: Re­
matrix degrades when the degree of power mismatch becomes severe. sources, Writing – review & editing. Jun Liu: Conceptualization. Qin­
Besides, the 2S-Durbin-TD-PHE shows lower PDs than other proposed glei Du: Conceptualization.
detectors, which, in turn, achieves superior performance to existing
competitors. Furthermore, combining the outcomes of Fig. 3(a) and (b), Declaration of Competing Interest
it is shown that with a fixed parameter σ2 , more training data L are
capable of improving the detection performance of all proposed The authors declare that they have no competing financial or non-
detectors. financial interests.
Fig. 4 plots the curves of the PD versus SNR for the case of interfering
targets. It is assumed that there are two interfering targets which share Acknowledgments
the same mathematical model as the expected target in the test data. The
interfering target to noise ratio (INR) is equal to the SNR. Compared This work was supported in part by National Natural Science Foun­
dation of China under Contract 62071482, the Youth Innovation

7
C. Huang et al. Signal Processing 217 (2024) 109353

Promotion Association CAS (CX2100060053), and the Natural Science [22] S. Lu, W. Yi, W. Liu, G. Cui, L. Kong, X. Yang, Data-dependent clustering-CFAR
detector in heterogeneous environment, IEEE Trans. Aerosp. Electron. Syst. 54 (1)
Foundation of Anhui Province under Grant 2208085J17.
(2018) 476–485.
[23] J. Liu, D. Massaro, D. Orlando, A. Farina, Radar adaptive detection architectures
References for heterogeneous environments, IEEE Trans. Signal Process. 68 (2020)
4307–4319.
[1] W. Liu, J. Liu, C. Hao, Y. Gao, Y.L. Wang, Multichannel adaptive signal detection: [24] A. Coluccia, D. Orlando, G. Ricci, A GLRT-like CFAR detector for heterogeneous
basic theory and literature review, Sci. China Inf. Sci. 65 (2) (2022), 121301. environments, Signal Process. 194 (2022), 108401.
[2] Y. Xiong, W. Xie, Y. Wang, Space time adaptive processing for airborne MIMO [25] N. Li, G. Cui, H. Yang, L. Kong, Q.H. Liu, S. Iommelli, Adaptive detection of moving
radar based on space time sampling matrix, Signal Process. 211 (2023), 109119. target with MIMO radar in heterogeneous environments based on Rao and Wald
[3] L. Ding, W. Chen, W. Zhang, Em-based sparse imaging for colocated MIMO radar tests, Signal Process. 114 (2015) 198–208.
under phase synchronization mismatch, in: Proceedings of the IEEE International [26] E. Conte, A.D. Maio, G. Ricci, GLRT-based adaptive detection algorithms for range-
Conference on Acoustics, Speech and Signal Processing, 2013, pp. 4106–4109. spread targets, IEEE Trans. Signal Process. 49 (7) (2001) 1336–1348.
[4] S. Chao, CRB of target location and velocity estimation in MIMO radar, in: [27] M. Sun, W. Liu, J. Liu, C. Hao, Rao and Wald tests for target detection in coherent
Proceedings of the IET International Radar Conference, 2013, pp. 1–6. interference, IEEE Trans. Aerosp. Electron. Syst. 58 (3) (2022) 1906–1921.
[5] H. Zang, S. Zhou, X. Lv, Y. Cao, L. Xu, H. Liu, Joint optimization of waveforms and [28] F. Bandiera, O. Besson, D. Orlando, G. Ricci, L.L. Scharf, GLRT-based direction
transmit array for colocated MIMO radar, in: Proceedings of the IEEE Radar detectors in homogeneous noise and subspace interference, IEEE Trans. Signal
Conference, 2015, pp. 374–378. Process. 55 (6) (2007) 2386–2394.
[6] L. Xu, J. Li, P. Stoica, Target detection and parameter estimation for MIMO radar [29] W. Liu, J. Liu, L. Wang, K. Duan, Z. Chen, Y. Wang, Adaptive array detection in
systems, IEEE Trans. Aerosp. Electron. Syst. 44 (3) (2008) 927–939. noise and completely unknown jamming, Digit. Signal Process 46 (2015) 41–48.
[7] W. Liu, Y. Wang, J. Liu, W. Xie, H. Chen, W. Gu, Adaptive detection without [30] W. Liu, W. Xie, J. Liu, Y. Wang, Adaptive double subspace signal detection in
training data in colocated MIMO radar, IEEE Trans. Aerosp. Electron. Syst. 51 (3) Gaussian background-part II: partially homogeneous environments, IEEE Trans.
(2015) 2469–2479. Signal Process. 62 (9) (2014) 2358–2369.
[8] J. Liu, J. Han, Z.J. Zhang, J. Li, Bayesian detection for MIMO radar in Gaussian [31] Y. Gao, G. Liao, S. Zhu, X. Zhang, D. Yang, Persymmetric adaptive detectors in
clutter, IEEE Trans. Signal Process. 66 (24) (2018) 6549–6559. homogeneous and partially homogeneous environments, IEEE Trans. Signal
[9] J. Liu, S. Zhou, W. Liu, J. Zheng, H. Liu, J. Li, Tunable adaptive detection in Process. 62 (2) (2014) 331–342.
colocated MIMO radar, IEEE Trans. Signal Process. 66 (4) (2018) 1080–1092. [32] W. Liu, J. Liu, T. Liu, H. Chen, Y.L. Wang, Detector design and performance
[10] J. Liu, W. Liu, J. Han, B. Tang, Y. Zhao, H. Yang, Persymmetric GLRT detection in analysis for target detection in subspace interference, IEEE Signal Process Lett. 30
MIMO radar, IEEE Trans. Veh. Technol. 67 (12) (2018) 11913–11923. (2023) 618–622.
[11] J. Liu, J. Han, W. Liu, S. Xu, Z.J. Zhang, Persymmetric Rao test for MIMO radar in [33] H. Ye, Y.L. Wang, W. Liu, J. Liu, H. Chen, Adaptive detection based on gradient test
Gaussian disturbance, Signal Process. 165 (2019) 30–36. and Durbin test in spectrally symmetric interference, Signal Process. 203 (2023),
[12] G. Cui, L. Kong, X. Yang, Performance analysis of colocated MIMO radars with 108793.
randomly distributed arrays in compound-Gaussian clutter, Circuits Syst. Signal [34] M. Sun, W. Liu, J. Liu, C. Hao, Multichannel adaptive detection based on gradient
Process. 31 (4) (2011) 1407–1422. test and durbin test in deterministic interference and structure nonhomogeneity,
[13] I. Bekkerman, J. Tabrikian, Target detection and localization using MIMO radars IEEE Signal Process Lett. 29 (2022) 592–596.
and sonars, IEEE Trans. Signal Process. 54 (10) (2006) 3873–3883. [35] P. Tang, R. Dong, W. Liu, J. Liu, Q. Du, Y.L. Wang, Adaptive multichannel detectors
[14] M. Greco, S. Fortunati, F. Gini, Maximum likelihood covariance matrix estimation for distributed target based on gradient test, Signal Process. 191 (2022), 108350.
for complex elliptically symmetric distributions under mismatched conditions, [36] A. Zaimbashi, M.R. Taban, M.M. Nayebi, Y. Norouzi, Weighted order statistic and
Signal Process. 104 (2014) 381–386. fuzzy rules CFAR detector for Weibull clutter, Signal Process. 88 (3) (2008)
[15] A. De Maio, G. Alfano, Polarimetric adaptive detection in non-Gaussian noise, 558–570.
Signal Process. 83 (2) (2003) 297–306. [37] A. Zaimbashi, An adaptive cell averaging-based CFAR detector for interfering
[16] J.A. Guan, X.L. Zhang, Subspace detection for range and Doppler distributed targets and clutter-edge situations, Digit Signal Process 31 (2014) 59–68.
targets with Rao and Wald tests, Signal Process. 91 (1) (2011) 51–60. [38] A. Zaimbashi, J. Li, Tunable adaptive target detection with kernels in colocated
[17] C. Hao, X. Ma, X. Shang, L. Cai, Adaptive detection of distributed targets in MIMO radar, IEEE Trans. Signal Process. 68 (2020) 1500–1514.
partially homogeneous environment with Rao and Wald tests, Signal Process. 92 [39] W. Liu, Y. Wang, W. Xie, Fisher information matrix, Rao Test, and Wald Test for
(4) (2012) 926–930. complex-valued signals and their applications, Signal Process. 94 (2014) 1–5.
[18] W. Liu, J. Liu, L. Huang, D. Zou, Y. Wang, Rao tests for distributed target detection [40] M. Sun, W. Liu, J. Liu, C. Hao, Complex parameter Rao, Wald, gradient, and Durbin
in interference and noise, Signal Process. 117 (2015) 333–342. Tests for multichannel signal detection, IEEE Trans. Signal Process. 70 (2022)
[19] C. Huang, Y.L. Wang, W. Liu, Q. Du, J. Liu, Determination between target and 117–131.
jamming based on multiple alternative hypotheses, IET Signal Proc. 17 (2) (2023) [41] J. Durbin, Testing for serial correlation in least-squares regression when some of
e12172. the regressors are lagged dependent variables, Econometrica 38 (3) (1970)
[20] L. Zeng, Y.L. Wang, W. Liu, J. Liu, Z. Zhang, Adaptive detectors for colocated 410–421.
MIMO radar with training data, IEEE Geosci. Remote Sens. Lett. 19 (2022) 1–5. [42] J. Liu, W. Liu, B. Chen, H. Liu, H. Li, Detection probability of a CFAR matched filter
[21] A. Coluccia, A. Fascista, D. Orlando, G. Ricci, Radar detectors for heterogeneous with signal steering vector errors, IEEE Signal Process Lett. 22 (12) (2015)
environments: a comparison on IPIX data, in: Proceedings of the International 2474–2478.
Conference on Radar Systems, Edinburgh, United kingdom, Institution of
Engineering and Technology, 2022, pp. 324–329.

You might also like