You are on page 1of 5

This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAES.2018.2847958, IEEE
Transactions on Aerospace and Electronic Systems

Fast Angle Estimation for MIMO Radar With


Nonorthogonal Waveforms
Bin Liao, Senior Member, IEEE

Abstract—In this letter, a fast angle estimation method for to robust angle estimation for MIMO radar in the presence of
multiple-input multiple-output (MIMO) radar with nonorthogo- array imperfections [12]–[15].
nal waveforms is devised. This method first estimates the noise-
free covariance matrix by exploiting its low-rank property as
It should be emphasized that, as mentioned earlier, the above
well as the sparse structure of the noise covariance matrix. A approaches for angle estimation are based on the assumption of
subspace-based procedure is then developed to determine the orthogonal waveforms. However, because of various consider-
directions-of-arrival (DOAs) based on the matrix composed of ations in MIMO radar waveform design, perfect orthogonality
the principal eigenvectors of the noise-free covariance matrix cannot be guaranteed [16]–[19]. The angle estimation perfor-
estimate. Compared with the state-of-the-art prewhitening algo-
rithm, the proposed method does not need the knowledge of
mance of existing methods will be degraded if the waveform
the correlation matrix of transmitted waveforms. Moreover, it is nonorthogonality is ignored. Hence, a prewhitening approach
computationally attractive since the DOAs can be estimated in [20] was proposed recently. However, this approach depends
closed-form. In addition, the proposed method offers promising on the exact knowledge of the waveform correlation matrix
DOA estimation performance as verified by simulation results. which may be unavailable or imprecisely known, and requires
Index Terms—MIMO radar, direction-of-arrival, nonorthogo- a spectrum grid search for DOA estimation. This motivates us
nal waveforms, eigendecomposition. to devise a computationally more attractive method which is
free of such knowledge. Briefly speaking, the proposed method
I. I NTRODUCTION formulates a low-rank matrix completion problem for noise-
free covariance matrix estimation, which is exploited for DOA
NLIKE the traditional phased-array radar that transmits
U a single waveform, multiple-input multiple-output (MI-
MO) radar transmits multiple probing waveforms. Owing to
estimation by utilizing a shift-invariant property. Its excellent
performance is demonstrated by simulation results.
Notations: (·)T , (·)H and (·)∗ are the operators of transpose,
the waveform diversity, MIMO radar is capable of providing Hermitian transpose, conjugate, respectively. ⊗ and ⊙ denote
better target detection capability, superior parameter estimation the Kroneker product and Hadamard product, respectively. IN
performance, improved parameter identifiability and higher and JN are the N × N identity matrix and all-ones matrix, re-
spatial resolution [1]–[3]. As one of the most important radar spectively. E[·] denotes the mathematical expectation. diag{·}
applications, estimating the angles of targets with MIMO radar √
generates a diagonal matrix, ȷ = −1 is the imaginary unit,
has attracted much research interest in recent years. ∅ denotes the empty set and ∥·∥2 is the ℓ2 norm. trace(·) and
By implicitly or explicitly assuming orthogonal waveforms, vec(·) denote the matrix trace and vectorization, respectively.
numerous methods for angle estimation in MIMO radar have C and H represent the complex and Hermitian matrix sets,
been proposed [4]–[15]. The maximum likelihood (ML) es- respectively. Xij = [X]ij defines the (i, j)th entry of X.
timator for direction-of-arrival (DOA) estimation was devel-
oped in [4], the Capon method was applied to estimate the
directions in [5], and a combined Capon and approximate ML II. MIMO R ADAR S IGNAL M ODEL
(CAML) method was introduced in [6]. Like their conventional
counterparts, these methods have high complexities due to Consider a monostatic MIMO radar system with M trans-
exhaustive parameter search and spectrum grid search. To mit antennas and N receive antennas. The transmitter emits
this end, dimension reduction and polynomial rooting were M waveforms {sm (t)}M m=1 via omnidirectional transmission,
incorporated into the Capon estimator [7] and multiple signal where t is the fast time index (time index within a radar pulse).
classification (MUSIC) algorithm [8]. Moreover, the concept Assume that K targets with DOAs {θ1 , · · · , θK } are present,
of estimation of signal parameters via rotational invariance the radar return of the kth target is thus given by
techniques (ESPRIT) was utilized for angle estimation in [9]
followed by an improved version without DOA and direction- αk (t, τ ) = rk (τ )aTt (θk )s(t) (1)
of-departure (DOD) paring [10] and an iterative ESPRIT-like
algorithm [11]. Note that many efforts have also been devoted where τ is the slow time index (pulse index), rk (τ ) denotes the
reflection coefficient of the kth target and it is constant within
This work was supported by the National Natural Science Foundation of
China under grants 61771316 and 61471365, Foundation of Department of
a pulse but time-varying over the slow time, at (θk ) ∈ CM ×1
Education of Guangdong under grant 2016KTSCX125, Foundation of Shen- is the transmit steering vector corresponding to the kth target,
zhen under grant JCYJ20170302150044331, and Foundation of Guangdong and s(t) = [s1 (t), · · · , sM (t)]T is the waveform vector.
Science and Technology Department under grant 2017A030313341.
The author is with the College of Information Engineering, Shenzhen Denote by ar (θk ) ∈ CN ×1 the receive steering vector cor-
University, Shenzhen 518060, China (e-mail: binliao@szu.eud.cn). responding to the kth target, the observation x(t, τ ) ∈ CN ×1

0018-9251 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAES.2018.2847958, IEEE
Transactions on Aerospace and Electronic Systems

at the receiver can thus be expressed as Let us define E1 , [IN −1 , 0(N −1)×1 ] ∈ C(N −1)×N and E2 ,

K [0(N −1)×1 , IN −1 ] ∈ C(N −1)×N , it can be deduced that
x(t, τ ) = rk (τ )ar (θk )aTt (θk )s(t) + w(t, τ ) (2)
E2 ar (θ) = eȷπ sin θ E1 ar (θ). (8)
k=1

where w(t, τ ) is the zero-mean additive white Gaussian noise Accordingly, assuming P1 = IM ⊗ E1 and P2 = IM ⊗ E2 ,
(AWGN) vector with E[w(t1 , τ )wH (t2 , τ )] = σn2 IN δ(t1 −t2 ). recalling a(θ) = at (θ) ⊗ ar (θ) and using the known identity
Assuming the pulsewidth is Tp and matching the observation (W ⊗ X)(Y ⊗ Z) = (WY) ⊗ (XZ), one gets

with sm (t) yields the output ym (τ ) = Tp x(t, τ )s∗m (t)dt ∈
Pκ a(θ) = (IM ⊗ Eκ )(at (θ) ⊗ ar (θ)) = at (θ) ⊗ Eκ ar (θ) (9)
CN ×1 , m = 1, · · · , M , as
for ∀κ ∈ {1, 2}. Therefore, it can be found from (8) and (9)

K
ym (τ ) = rk (τ )ar (θk )at,m (θk ) + nm (τ ) (3) that P2 a(θ) = eȷπ sin θ P1 a(θ). By applying this relation to all
k=1
steering vectors in A = [a(θ1 ), · · · , a(θK )], we have

where at,m (θk ) = aTt (θk ) Tp s(t)s∗m (t)dt and nm (τ ) = P2 A = P1 AΨ (10)

Tp
w(t, τ )s∗m (t)dt. Define at (θk ) , [at,1 (θk ), · · · , at,M (θk )]T where Ψ = diag{eȷπ sin θ1 , · · · , eȷπ sin θK }.
and n(τ ) , [nT1 (τ ), · · · , nTM (τ )]T , then stack all matched Next, let us define a matrix U ∈ CM N ×K (details of the
outputs as y(τ ) , [y1T (τ ), · · · , yM
T
(τ )]T ∈ CM N ×1 , one gets estimation of U will be presented later) which spans the same
∑ column space as A ∈ CM N ×K , i.e.,
K
( )
y(τ ) = at (θk ) ⊗ ar (θk ) rk (τ )+n(τ ) = Ar(τ )+n(τ ) (4)
∃ T ∈ CK×K
nonsingular : A = UT (11)
k=1

where r(τ ) = [r1 (τ ), · · · , rK (τ )]T , A = [a(θ1 ), · · · , a(θK )], where CK×K


nonsingular stands for the set of nonsingular matrices.
and a(θ) , at (θ) ⊗ ar (θ). ∫
Substituting (11) into (10) yields P2 U = P1 UΞ, where
Recall that at,m (θk ) = aTt (θk ) Tp s(t)s∗m (t)dt, we have
Ξ = TΨT−1 = (UPH
1 P1 U)
−1
UPH
1 P2 U. (12)
at (θk ) , [at,1 (θk ), · · · , at,M (θk )]T = RTs at (θ), where Rs ∈
CM ×M is the correlation matrix of transmitted waveforms as This means that Ξ and Ψ are similar matrices having identical
  eigenvalues. Hence, the DOAs {θk }K
1 β12 · · · β1M k=1 can be determined as
∫  β21 · · · β2M 
 1  θk = arcsin(π −1 arg(ξk )) (13)
Rs = s(t)sH (t)dt =  . . .. ..  (5)
 .
. .
. . . 
Tp
where {ξk }K k=1 are the eigenvalues of Ξ. It is interesting to
βM 1 βM 2 · · · 1 note that the above process does not need the knowledge of

where βij = Tp si (t)s∗j (t)dt. Likewise, the noise term n(t) the correlation matrix of transmitted waveforms.
∫ Now, let us proceed to the estimation of U (and hence Ξ),
can be expressed as n(τ ) = Tp s∗ (t) ⊗ w(t, τ )dt and
which is required to estimate the DOAs as (13). According to
Rn = E[n(τ )nH (τ )] = R∗s ⊗ σn2 IN . In particular, when the
(4), it is known that the covariance matrix of y(τ ) is
waveforms are orthogonal, we have βij = 0 for i ̸= j, and
thus Rs = IM , a(θ) = at (θ)⊗ar (θ) and Rn = σn2 IM N . This Ry = Rnf + Rn , Rnf , ARr AH (14)
enables us to straightforwardly utilize traditional methods for
angle estimation. where Rnf defines the noise-free covariance matrix with Rr =
However, in the case of nonorthogonal waveforms, the E[r(τ )rH (τ )]. It is readily known that if Rn is a scaled identi-
steering vector a(θ) is related to the waveform correlations and ty matrix, then U corresponds to the K principal eigenvectors
the resulting noise n(τ ) is colored. Thus, Zheng [20] recently of Ry . Unfortunately, the waveforms are nonorthogonal (i.e.,
proposed to whiten the output y(τ ) as Rs ̸= I) and Rn = R∗s ⊗ σn2 IN corresponds to a colored
noise model. This is the reason why the recent approach [20]
−1
z(τ ) = R̄n 2 y(τ ) (6) requires the knowledge of Rs to whiten y(τ ). In what follows,
we shall show that the noise-free covariance matrix Rnf can be
where R̄n = R∗s ⊗ IN . With such a prewhitening process,
determined by exploiting its low-rank property and the sparse
traditional approaches can be directly employed. However, as
structure of Rn . By doing so, U can be estimated via the
mentioned earlier, this approach depends on the knowledge b nf .
eigendecomposition of R
of the correlation matrix Rs , and hence, is not feasible for
We first notice that the number of entries of Rn is M 2 N 2 ,
scenarios where Rs is unknown or imperfectly known. In what
and the number of nonzero entries, denoted as #nonzero , is
follows, a new method free of this knowledge is devised.
#nonzero ≤ M 2 N (15)
III. FAST A NGLE E STIMATION A LGORITHM
where the equality holds only if Rs does not contain nonzero
For illustration purpose, assume that the receiver deploys a
entries. For instance, if N = 10, at least 90% entries of Rn are
uniform linear array (ULA) with half-wavelength interelement
equal to zero. This indicates that Rn is a sparse matrix with
spacing, then we have
only a small portion of entries being nonzero, and therefore,
ar (θ) = [1, eȷπ sin θ , · · · , eȷπ(N −1) sin θ ]T . (7) Ry and Rnf are almost identical except for those entries with

0018-9251 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAES.2018.2847958, IEEE
Transactions on Aerospace and Electronic Systems

indices (which can be concluded from the special structure of Algorithm 1 Proposed fast angle estimation algorithm.
Rn ) in the following set: 1: Estimate Ry from P pulses as Rb y = 1 ∑P y(τ )yH (τ )
{ } P
b nf by solving (19) using CVX
τ =1
Ω̄ = (i, j) i − j = 0, ±N, · · · , ±(M − 1)N (16) 2: Obtain R
3: Eigendecompose R b nf to obtain U
b
where i, j ∈ {1, · · · , M N }. Therefore, let Ω be the comple- b b b
4: Estimate Ξ as Ξ = (UP1 P1 U)−1 UP
H b H P2 U b
1
mentary set of Ω̄, then Ω ∩ Ω̄ = ∅ and Ω ∪ Ω̄ = {(i, j)|i = b
5: Eigendecompose Ξ to obtain the eigenvalues ξˆ1 , · · · , ξˆK
1, · · · , M N ; j = 1, · · · , M N }, and the entries of Ry and 6: Estimate the DOAs as θ̂k = arcsin(π −1 arg(ξˆk ))
Rnf with indices in Ω are identical. Furthermore, the rank of
Rnf , N , is generally much less than its dimension, M N . This
allows us to recover Rnf from a portion of its entries with the
low-rank matrix completion technique [21]. a symmetric structure. To be more practical, we consider the
According to the observation above, we define a sampling imperfect knowledge of the true Rs and assume that
operator PΩ : CM ×M → CM ×M as βbij = βbji = (1 + ϵ)βij , ϵ ∼ unif(−ρ, ρ) (20)
{
Xij , (i, j) ∈ Ω b s will thus be applied to the algorithms which require
[PΩ (X)]ij = (17) and R
0, (i, j) ∈ Ω̄
the knowledge of waveform correlation matrix. ϵ is uniformly
then we can determine Rnf by minimizing its rank subject to distributed within [−ρ, ρ]. In particular, ρ = 0 or ϵ = 0 means
a constraint PΩ (Rnf ) = PΩ (Ry ). A careful examination of that Rs is exactly known. In all simulations, P = 1000 puleses
(16) and (19) shows that the constraint can be expressed as are collected to calculate the covariance matrix estimate R b y.
ω ⊙ vec(Rnf − Ry ) = 0 (18) The waveform correlation coefficients βij are generated from
a uniform distribution unif(0, 0.15).
where ω , vec(JM ⊗ (JN − IN )). It is known that rank min- In the first example, we compare the DOA estimates of the
imization problem is NP-hard and not solvable in polynomial proposed method with those obtained by spatial spectra of the
time deterministically. A well-recognized effective solver is prewhitening method [20] and MUSIC (ignoring the waveform
relaxing the problem by nuclear norm, which corresponds to nonorthogonality) in various scenarios. More precisely, assume
the sum of singular values. Since Rnf is Hermitian, minimiz- the signal-to-noise ratio (SNR) is 10 dB, four cases, i.e., ρ = 0,
ing its trace is thus the objective. Consequently, the following 0.1, 0.3, 0.5, are considered and 10 independent trials are run
linear programming problem is formulated to estimate Rnf : for each case. The spatial spectra of the prewhitening method
minimize trace(M) (red line) and MUSIC (blue line), and the DOA estimates of
M the proposed method (marked with ×) are shown in Fig.1. It is
subject to ω ⊙ vec(M − Ry ) = 0 (19) seen that when the correlation matrix is known exactly (i.e.,
M ∈ HM N ×M N ρ = 0), the prewhitening method offers good performance.
However, its performance is dramatically degraded when the
which can be solved using, e.g., CVX [22]. Assume R b nf is the
knowledge of Rs is not precise enough. On the other hand,
b
solution to (19), then U includes the K principal eigenvectors the proposed method performs well in all cases since it is not
of Rb nf , and hence, Ξ b = (UPb H P1 U)b −1 UP
b H P2 U.b
1 1 affected by the knowledge of Rs . In addition, as expected, it
b
In practice, Ry in (19) is replaced by Ry obtained from a is noted that once the nonorthogonality is ignored, traditional
finite set of pulses, and the equality constraint does not hold methods such as MUSIC cannot perform properly.
exactly. Thus, a quadratic constraint ∥ω⊙vec(M−Ry )∥2 ≤ ε In the second example, we first measure the performance of
may be utilized, where ε depends on the accuracy of R b y . If proposed method in terms of root mean square error (RMSE)
the number of pulses is large, ε is assigned a small value. of DOA estimation (as defined in [20]), which is calculated
In particular, if the number of pulses is sufficiently large, we from 1000 independent trials. Since the radar output y(τ ) in
can simply assume ε = 0, i.e., using the equality constraint, (4) obeys a colored noise mode, the Cramér-Rao bound (CRB)
without visibly affecting the performance. derived in [23] for unknown noise is adopted for performance
Finally, it is seen that the main computations in the proposed evaluation. The curves of RMSE versus SNR are drawn in
method include solving (19) and performing the eigendecom- Fig.2(a). It is seen that if the waveform correlation is exactly
position of R b nf . Thus, the overall complexity is O(M 3 N 3 ),
known (i.e., ρ = 0), the prewhitening method is able to attain
which is lower that that of prewhitening method [20]. The the CRB at high SNRs, but it is sensitive to the precision of
proposed fast angle estimation algorithm for MIMO radar with the knowledge of waveform correlation. For instance, when
nonorthogonal waveforms is summarized in Algorithm 1. ρ = 0.3, the prewhitening method results in significantly
degraded performance. It appears that the proposed method
IV. S IMULATION R ESULTS does not attain the CRB and there is always a performance
We assume that both the transmitter and receiver are ULAs gap. This is due to the fact that the proposed method uses the
of M = N = 8 elements with half-wavelength inter-element ESPRIT concept which does not make full use of the effective
spacing. K = 3 targets locate at −5◦ , −1◦ and 2◦ . Following array aperture. Nevertheless, compared with the prewhitening
the Swerling II target model, {rk }K k=1 are assumed to be method, it has lower computational complexity as well as
zero-mean Gaussian distribution with variance σr2 . The wave- higher resolution, and therefore, provides lower RMSE when
forms are nonorthogonal with correlation matrix Rs having the SNR is relatively low, e.g., less than 10 dB.

0018-9251 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAES.2018.2847958, IEEE
Transactions on Aerospace and Electronic Systems

1
10
Spatial Spectrum (dB)

Spatial Spectrum (dB)


Proposed method
0 0 Prewhitening method ρ = 0
Prewhitening method ρ = 0.3
−10 −10 Prewhitening method ρ = 0.5

RMSE (Degree)
0
10 CRB

−20 ρ =0 −20 ρ =0.1

−30 −30 10
−1
−8 −5 −2 1 4 6 −8 −5 −2 1 4 6
Azimuth (Deg) Azimuth (Deg)
(a) (b)

Spatial Spectrum (dB)


Spatial Spectrum (dB)

−2
0 0 10
−4 −2 0 2 4 6 8 10 12 14 16 18 20
SNR (dB)
−10 −10 (a)
1
−20 ρ =0.3 −20 ρ =0.5
0.8
−30 −30
−8 −5 −2 1 4 6 −8 −5 −2 1 4 6

Probability
0.6
Azimuth (Deg) Azimuth (Deg)
(c) (d)
0.4
Fig. 1. Spatial spectrum of MUSIC algorithm (blue line) and prewhitening Proposed method
method (red line), and DOA estimates of the proposed method (marked with 0.2 Prewhitening method ρ = 0
×). (a) ρ = 0, (b) ρ = 0.1, (c) ρ = 0.3, and (d) ρ = 0.5. Prewhitening method ρ = 0.3
Prewhitening method ρ = 0.5
0
−4 −2 0 2 4 6 8 10 12 14 16 18 20
SNR (dB)
Following the above settings, the resolution probability is (b)
investigated (here, sources are deemed to be resolved if the
bias of each DOA estimate is less than 0.3◦ ). 1000 independent Fig. 2. Illustration of statistical performance. (a) RMSE versus SNR, and
trials are run for this purpose. Fig.2(b) shows the resulting (b) Resolution probability versus SNR.
resolution probability versus SNR. It is obvious that the
proposed method offers much better than the prewhitening
method even if the waveform correlation matrix is exactly [3] J. Li and P. Stoica, “MIMO radar with colocated antennas,” IEEE Signal
Process. Mag., vol. 24, no. 5, pp. 106–114, Sep. 2007.
known. Furthermore, it is seen that the proposed method has [4] I. Bekkerman and J. Tabrikian, “Target detection and localization using
lower SNR threshold to properly determine the DOAs. This MIMO radars and sonars,” IEEE Trans. Signal Process., vol. 54, no. 10,
coincides with the results in Fig.2(a). pp. 3873–3883, Oct. 2006.
[5] H. Yan, J. Li, and G. Liao, “Multitarget identification and localization
using bistatic MIMO radar systems,” EURASIP J. Adv. Signal Process.,
V. C ONCLUSIONS vol. 2008, no. 1, p. 283483, Dec. 2007.
[6] L. Xu, J. Li, and P. Stoica, “Target detection and parameter estimation
Orthogonal waveforms are desirable in MIMO radar system- for MIMO radar systems,” IEEE Trans. Aerosp. Electron. Syst., vol. 44,
s, however, the orthogonality may not always hold in practice. no. 3, pp. 927–939, Jul. 2008.
[7] R. Xie, Z. Liu, and Z. Zhang, “DOA estimation for monostatic MIMO
If this is ignored without properly treatment, the performance radar using polynomial rooting,” Signal Process., vol. 90, no. 12, pp.
of parameter estimation may be considerably deteriorated. In 3284 – 3288, 2010.
this work, we have devised a fast angle estimation method for [8] Y. Zhang, G. Zhang, and X. Wang, “Computationally efficient DOA
estimation for monostatic MIMO radar based on covariance matrix
MIMO radar with nonorthogonal waveforms. By exploiting reconstruction,” Electron. Lett., vol. 53, no. 2, pp. 111–113, 2017.
the features such as shift-invariant virtual steering vector, low- [9] C. Duofang, C. Baixiao, and Q. Guodong, “Angle estimation using
rank structure of the noise-free covariance matrix, and sparsity ESPRIT in MIMO radar,” Electron. Lett., vol. 44, no. 12, pp. 770–771,
June 2008.
of the noise covariance matrix, it has been shown that the [10] C. Jinli, G. Hong, and S. Weimin, “Angle estimation using ESPRIT
proposed method estimates the DOAs in a computationally without pairing in MIMO radar,” Electron. Lett., vol. 44, no. 24, pp.
efficient manner. More importantly, this method does not need 1422–1423, Nov. 2008.
[11] F. K. W. Chan, H. C. So, L. Huang, and L. T. Huang, “Parameter
the waveform correlation knowledge. Simulation results illus- estimation and identifiability in bistatic multiple-input multiple-output
trated that the proposed method offers promising performance radar,” IEEE Trans. Aerosp. Electron. Syst., vol. 51, no. 3, pp. 2047–
as compared with the most recent prewhitening method. 2056, Jul. 2015.
[12] X. Wang, W. Wang, J. Liu, Q. Liu, and B. Wang, “Tensor-based real-
valued subspace approach for angle estimation in bistatic MIMO radar
R EFERENCES with unknown mutual coupling,” Signal Process., vol. 116, pp. 152 –
158, 2015.
[1] E. Fishler, A. Haimovich, R. S. Blum, L. J. Cimini, D. Chizhik, and [13] C. Zhang, H. Huang, and B. Liao, “Direction finding in MIMO radar
R. A. Valenzuela, “Spatial diversity in radars-models and detection with unknown mutual coupling,” IEEE Access, vol. 5, pp. 4439–4447,
performance,” IEEE Trans. Signal Process., vol. 54, no. 3, pp. 823–838, 2017.
Mar. 2006. [14] B. Liao and S. C. Chan, “Direction finding in MIMO radar with un-
[2] A. D. Maio and M. Lops, “Design principles of MIMO radar detectors,” known transmitter and/or receiver gains and phases,” Multidimensional
IEEE Trans. Aerosp. Electron. Syst., vol. 43, no. 3, pp. 886–898, Jul. Syst. Signal Process., vol. 28, no. 2, pp. 691–707, Apr. 2017.
2007. [15] H. Jiang, J. K. Zhang, and K. M. Wong, “Joint DOD and DOA estimation

0018-9251 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAES.2018.2847958, IEEE
Transactions on Aerospace and Electronic Systems

for bistatic MIMO radar in unknown correlated noise,” IEEE Trans. Veh.
Technol., vol. 64, no. 11, pp. 5113–5125, Nov. 2015.
[16] Y. Yang and R. S. Blum, “MIMO radar waveform design based on
mutual information and minimum mean-square error estimation,” IEEE
Trans. Aerosp. Electron. Syst., vol. 43, no. 1, pp. 330–343, Jan. 2007.
[17] Z. Cheng, Z. He, B. Liao, and M. Fang, “MIMO radar waveform design
with PAPR and similarity constraints,” IEEE Trans. Signal Process.,
vol. 66, no. 4, pp. 968–981, Feb. 2018.
[18] Y. Li and S. A. Vorobyov, “Fast algorithms for designing unimodular
waveform(s) with good correlation properties,” IEEE Trans. Signal
Process., vol. 66, no. 5, pp. 1197–1212, Mar. 2018.
[19] G. Cui, H. Li, and M. Rangaswamy, “MIMO radar waveform design
with constant modulus and similarity constraints,” IEEE Trans. Signal
Process., vol. 62, no. 2, pp. 343–353, Jan. 2014.
[20] G. Zheng, “DOA estimation in MIMO radar with non-perfectly orthogo-
nal waveforms,” IEEE Commun. Lett., vol. 21, no. 2, pp. 414–417, Feb.
2017.
[21] E. J. Candès and Y. Plan, “Matrix completion with noise,” Proc. IEEE,
vol. 98, no. 6, pp. 925–936, Jun. 2010.
[22] M. Grant and S. Boyd, “CVX: Matlab software for disciplined convex
programming, version 2.1,” http://cvxr.com/cvx, 2014.
[23] A. B. Gershman, P. Stoica, M. Pesavento, and E. G. Larsson, “Stochastic
cramér-rao bound for direction estimation in unknown noise fields,” IEE
Proc., Radar Sonar Nuvig., vol. 149, no. 1, pp. 2–8, Feb. 2002.

0018-9251 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

You might also like