Received 1 December 2022, accepted 18 December 2022, date of publication 20 December 2022,

date of current version 27 December 2022.

Digital Object Identifier 10.1109/ACCESS.2022.3230988

Sparse Parameter Estimation and Imaging in

mmWave MIMO Radar Systems With Multiple
Stationary and Mobile Targets
MEESAM JAFRI , (Student Member, IEEE), SURAJ SRIVASTAVA , (Member, IEEE),
SANA ANWER , AND ADITYA K. JAGANNATHAM , (Senior Member, IEEE)
Department of Electrical Engineering, Indian Institute of Technology Kanpur, Kanpur 208016, India
Corresponding author: Meesam Jafri ([email protected])
The work of Aditya K. Jagannatham was supported in part by the Qualcomm Innovation Fellowship, and in part by the Arun Kumar Chair
Professorship.

ABSTRACT This work conceives novel target detection and parameter estimation schemes in millimeter-
wave (mmWave) multiple-input multiple-output (MIMO) radar (mMR) systems for both stationary and
mobile targets/radar platform. Initially, the orthogonal matching pursuit (OMP)-based mmR (OmMR)
algorithm is proposed for stationary targets to estimate their radar cross-section (RCS) coefficients, angle,
range locations together with the number of targets. Next, mMR systems with mobile targets and platform are
considered, followed by development of the simultaneous OMP (SOMP)-based mMR (SmMR) algorithm
for RCS, angle/range estimation together with their Doppler velocities. The proposed algorithms lead to a
significant improvement in performance since they exploit the inherent sparsity of the mMR scattering scene
in contrast to the conventional schemes. Two-dimensional (2D) mMR imaging procedures are also presented
for both scenarios in the angle, range, and Doppler dimensions. Analytical expressions are derived for the
Cramér-Rao bounds (CRBs) for the mean-squared error (MSE) of joint estimation of the RCS coefficients
and Doppler velocities. Simulation results demonstrate that proposed schemes perform well even in low
signal-to-noise ratio (SNR) scenarios with a few snapshots of the scattering environment and yield improved
performance in comparison to existing sparse as well as non-sparse schemes.

INDEX TERMS Millimeter wave (mmWave), MIMO radar, RCS coefficients, Doppler velocity, radar
imaging, parameter estimation, sparsity, simultaneous orthogonal matching pursuit.

I. INTRODUCTION is expected to play a pivotal role in future communication

Automotive Radar is a critical and promising technology for systems due to its ultra-high data rates that can enable various
autonomous vehicles and driving assistance systems. Auto- applications such as Vehicle-to-Everything (V2X) [3], body
motive radar technology can help avoid traffic congestion area networks, UAV communication, to name a few [4].
and road accidents, enable automatic cruise control (ACC), This is well suited also for radar applications, since the
automatic emergency brakes (AEB) and forward collision mmWave band provides a large bandwidth, which leads to a
warning (FCW) systems, in addition to greatly improving significantly higher angle and range resolution in radar, thus
the fuel efficiency [1]. However, it is important to note aiding in the accurate detection of various targets followed by
that the reliability of these functions in autonomous driving the estimation of their location parameters [1]. The temporal
systems depends critically on the accuracy of the angle, range and angular sparsity of the mmWave channel [5] can be
and velocity estimates obtained by the radar in a multi- efficiently exploited to further improve the accuracy of radar
target scenario [2]. Millimeter-wave (mmWave) technology detection and imaging [6]. The small wavelength of signals
at mmWave frequencies enables the packing of large antenna
The associate editor coordinating the review of this manuscript and arrays on wireless devices of limited sizes. According to
approving it for publication was Irfan Ahmed . the well-known Rayleigh criterion [7], a large antenna array

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
132836 VOLUME 10, 2022
M. Jafri et al.: Sparse Parameter Estimation and Imaging in mmWave MIMO Radar Systems

can significantly strengthen the angular resolution of the filter, for the one-dimensional direction finding problem
radar. Therefore, Multiple-input multiple-output (MIMO) of MIMO radar. The results therein demonstrate that the
technology, which uses large arrays with multiple transmit proposed scheme performs well in a system with a large
and receive antennas for sensing the scattering environment number of TAs and RAs, and has the ability to resolve all
and detection of targets, can lead to a significant improvement the targets as long as they are fewer in number than the total
in radar performance [8]. This is due to the fact that number of antennas.
MIMO technology supports the simultaneous transmission of Other researchers have presented MIMO radar techniques,
multiple probing signals followed by the reception of multiple such as Capon [19], APES [20] for the detection of multiple
signals reflected by the targets. Therefore, deploying multiple targets using data dependent algorithms followed by the esti-
antennas in the radar system enables the transmission of mation of the target location and associated RCS parameters.
multiple probing signals and also the subsequent reception However, it must be noted that the Capon scheme yields
of multiple reflected copies of the probing signals, which good estimates of target locations, while its RCS coefficient
yields increased degrees of freedom obtained via exploiting estimates are biased downward. APES on the other hand
spatial diversity. This, in turn, leads to enhanced estimation provides accurate estimates of the RCS coefficients at the
accuracy and better parameter identifiability in comparison to expense of a poorer resolution and hence leads to poorer
a standard phased-array radar, which transmits phase shifted estimation accuracy of the target locations. In [21], the
versions of a single waveform [9], [10], [11], [12]. A brief authors proposed a CAPES algorithm, which combines the
review of the existing works in this rapidly evolving area of best features of the Capon and APES algorithms, to refine the
mMR systems is presented next. estimates of the RCS coefficients and target locations. Zhang
et. al in [22] proposed a reduced dimensional Capon (RD-
A. REVIEW OF EXISTING WORKS Capon) algorithm which requires only a single dimensional
Several techniques have been described in the existing search, thus entailing a substantially lower computational
literature on MIMO Radar toward target detection and complexity. However, the constraint restricts only the data
parameter estimation. Bekkerman et al., in [13] proposed a received for first transmitted signal to be used to estimate
generalized likelihood ratio test (GLRT) for target detection the angle of arrival, which results in a performance loss.
and maximum likelihood (ML) estimation of target location. The authors in the treatise in [23] presented the novel Capon
However, the implementation of GLRT is computationally and Approximate Maximum Likelihood (AML) method,
complex, especially in a multi-target scenario, as the param- termed CAML, to improve the accuracy of RCS estimation.
eter search space becomes exponentially large. Moreover, It must be noted that all the above works consider the
the GLRT also requires prior information pertaining to targets to be located at an identical range, which is not a
the number of targets, which is infeasible in practice. realistic assumption in practical scenarios. In a rich scattering
To overcome this challenge, the authors in [14], proposed two environment, the multiple targets are often located at distinct
alternative schemes, i.e., the conditional GLRT (cGLRT) and angles and ranges [10]. Several schemes to tackle the
iterative GLRT (iGLRT), which only require searches over problem of identification, followed by estimation of the RCS,
one-dimensional spaces instead of the highly complex K - range and angle parameters were described in works such
dimensional search required by the GLRT, while achieving as [24], [25], [26], and [27] for stationary and mobile targets.
a performance close to that of the GLRT for parameter Yardibi et. al [24] proposed two non-parametric, viz., iterative
estimation. However, the cGLRT and iGLRT require a large adaptive approach (IAA) and its extension, termed IAA-ML,
number of snapshots for acceptable performance. Li et al. for RCS estimation coupled with target imaging. Roberts et.
in [15] proposed a range-angle generalized Likelihood al, in their treatise in [25], proposed the regularized IAA for
ratio test (RA-GLRT) for target detection with efficient mobile targets. However, these schemes incur a significant
clutter rejection. However, additional degrees of freedom computational burden owing to the large number of matrix
are needed in the range domain to distinguish targets of inversions required.
interest from the clutter. The authors in [16] proposed an Rawat et. al in [26] developed block least mean squares
ESPRIT-based angle of arrival estimation scheme to generate (BLMS) and fast BLMS (FBLMS) algorithms for the
highly resolvable images of the scattering environment. estimation of RCS coefficients and imaging in a MIMO
Their technique employed an intelligent scheme based on radar system. The fast FBLMS technique was clearly seen
the division of time-frequency resources to generate high to result in improved estimation and imaging performance,
quality radar snapshots. However, the scheme propounded together with faster convergence. However, the framework
by them can be computationally expensive since it requires considered in their work, and hence, the techniques described,
the eigenvalue decomposition of the spatial signal covariance were restricted to stationary targets and radar platform.
matrix. Yang et al. in [17] described a reduced-dimensional In [27], the authors extended the BLMS and FBLMS
ESPRIT algorithm to lower the dimensionality of the received for scenarios with mobile targets and radar platform. The
data prior to estimation of the angles of arrival in a multi- authors of [28] proposed recursive least squares (RLS)-
target scenario. Ngai et al. in [18], proposed a suitably refined based adaptive techniques for time-varying RCS coefficient
version of the ESPRIT algorithm, employing the Kalman estimation and 2D MIMO radar imaging in the presence of

VOLUME 10, 2022 132837

 M. Jafri et al.: Sparse Parameter Estimation and Imaging in mmWave MIMO Radar Systems

an unknown number of targets with unknown angles and mutual coupling among the antennas. However, the proposed
ranges. In [29], the authors proposed a two-stage parameter framework is unsuitable for colocated MIMO radar since
estimation technique for automotive MIMO radar. In their it involves a large number of computations. The authors
scheme, the first stage successfully performs low-complexity of [36] proposed a reduced-complexity SBL for colocated
three-dimensional peak detection, while the second stage MIMO radar toward DOA estimation. The array data
estimates the direction-of-arrival (DOA) via ESPRIT. Fur- obtained from matched filters is initially subjected to a low
thermore, the authors of [30] proposed a novel cost function complexity transformation, thereby removing redundancies
for joint estimation of the Doppler frequency and DOA of the in order to lower the computational complexity. However,
targets employing time division multiplexing (TDM) MIMO the authors therein consider only single dimensional spar-
radar toward resolving the velocity ambiguity for automotive sity in the angular domain. Zhao et al. [37] proposed
applications. However, the proposed framework therein is a framework for sparse parameter estimation in bistatic
only limited to the estimation of velocities lower than a MIMO radars based on l2 norm minimization employing
certain maximum velocity and the spectra of signals of targets the residual sum of square (RSS) technique. In their work,
with velocities higher than the velocity corresponding to the the constrained optimization model is initially transformed
Nyquist rate are still aliased, and hence cannot be accurately into an unconstrained optimization model with the aid of
estimated. To address this issue, Sohee et al. in [31] proposed Lagrange multipliers. The sparse solution is next obtained by
a method for velocity estimation while simultaneously minimizing this unconstrained model. However, the proposed
resolving the velocity ambiguity in a frequency-modulated solution requires the inversion of a dictionary matrix, which
continuous wave (FMCW) radar system. The authors of [32] incurs a high computational complexity. In [38], the authors
proposed a tensor generalized weighted linear predictor proposed a two-stage sparse parameter estimation procedure.
(TGWLP) for a frequency diverse array (FDA) MIMO radar In the first stage, the estimation problem is transformed
toward parallel estimation of radar parameters. into a sparse signal recovery problem, whose solution is
The significant difference of the previously published determined via a modified version of the sparse learning via
works [26], [27], and [28] with respect to the present iterative minimization (SLIM) technique. In the second stage,
work is that while the former consider only conventional a RELAX-based iterative algorithm is employed to refine
sub-6 GHz MIMO systems, the current work incorporates the estimates obtained via the SLIM algorithm. However,
mmWave MIMO technology, which is expected to play since the proposed method has two stages, and additionally
a crucial role in 5G communication networks due to its requires the RELAX algorithm to compute the sparse values
ultra-high data rates. This move to 5G mmWave MIMO of the target parameters, it can potentially be resource
technology leads to substantial challenges. For instance, intensive. Furthermore, the proposed algorithm also suffers
note that the conventional techniques, such as least squares from amplitude-related dynamic range issues, which hampers
(LS)/Linear minimum mean squared error (LMMSE), require its ability to efficiently exploit the sparsity of the radar
the transmission of a large number of probing signals, parameters. Zhimin et al., in [39], proposed an SBL varaint
particularly in high delay spread scenarios. This is because for phase errors-based DOA estimation, which exploits the
channel estimation using such conventional methods requires target sparsity in the spatial domain. Advantageously, the
an overdetermined system, thus leading to an ill-posed proposed framework does not require prior knowledge of the
problem when the number of probing signals is lower than phase errors.
the maximum number of channel taps in the finite impulse Although the works reviewed above focus on sparse
response (FIR) filter representing the frequency-selective parameter estimation in MIMO radar systems, they are
MIMO radar channel. However, all the above papers fail to based on the conventional fully-digital signal processing
exploit the inherent sparsity of the scattering environment architecture at the transmitter and receiver, which demands
at mmWave frequencies, arising from the presence of only a separate RF chain for each antenna. This poses significant
a few significant non-zero channel taps in the channel implementation challenges in the mmWave regime due to
response, which can play a crucial role in improving the the large number of antennas coupled with the high power
accuracy of target detection and RCS/location estimation. consumption of the high rate analog-to-digital converters
To overcome this challenge, other approaches such as (ADCs) [40]. To overcome this obstacle, novel hybrid
compressed sensing (CS), the convex-relaxation based least analog-digital beamforming architectures were advanced
absolute shrinkage and selection operator (Lasso) [33], and that successfully realized beamforming in mmWave MIMO
the focal underdetermined system solver (FOCUSS) [34], systems, especially in 5G, which require a much fewer
have been proposed for sparse parameter estimation in number of RF chains [41], [42], [43]. Thus, motivated by
ill-posed scenarios. However, the performance of Lasso these limitations of the above works in the existing literature,
depends critically on a user-defined regularization parameter, this paper conceives novel techniques for target identification
while that of FOCUSS suffers from convergence problems. via sparse parameter estimation in mmWave MIMO radar
As a further development, Peng et al. [35] proposed (mMR) systems using hybrid analog-digital beamforming,
sparse Bayesian learning (SBL) based direction of arrival considering stationary as well as mobile targets and radar
(DOA) estimation in MIMO radar systems with unknown platform, which efficiently exploit the sparse nature of

132838 VOLUME 10, 2022

M. Jafri et al.: Sparse Parameter Estimation and Imaging in mmWave MIMO Radar Systems

the scattering environment toward improved performance.

Table 1 presents a visual comparison of the contributions of
the various works discussed above in relation to the current
one. The various contributions of our paper are described
below in an itemized fashion.

B. CONTRIBUTIONS
This paper focuses on RCS, location estimation and radar
imaging for stationary and mobile targets/radar platforms in
mMR systems.
1) Initially, a model is developed for mMR systems with
stationary targets and radar platform with a specific
focus on the sparsity of the mmWave MIMO channel.
Subsequently, the orthogonal matching pursuit (OMP)-
based mmWave MIMO Radar (OmMR) algorithm is FIGURE 1. Block diagram of mmWave MIMO radar signal processing.
proposed to estimate the RCS coefficients for multiple
targets of interest in mMR systems.
the CRBs pertaining to estimation of the RCS coefficients and
2) Next, the proposed model is extended to include
Doppler velocities of the multiple targets. Finally, Section VI
mobility of the targets and radar platform with mobility.
presents our simulation results to illustrate the performance
For this system, the simultaneous orthogonal match-
of the proposed algorithms, followed by the conclusion in
ing pursuit (SOMP)-based mmWave MIMO Radar
Section VII. The intermediate steps in the derivation of the
(SmMR) technique is developed for joint estimation
CRBs are presented in Appendix A.
of the RCS coefficients, angle/range parameters as
well as Doppler velocities of the multiple targets using
multiple snapshots. Once again, taking advantage of 1) NOTATIONS
the sparsity leads to a substantial improvement in the The following notation is used throughout this paper. Vectors
estimates. Furthermore, radar imaging algorithms of and matrices are denoted by boldfaced lowercase a and upper-
the scattering environment are presented in the angle case A, respectively. The quantity diag (a1 , a2 , · · · , aN )
and range dimensions as well as the Doppler and range represents a diagonal matrix with a1 , a2 , · · · , aN on the
dimensions. principle diagonal and superscripts (·)T , (·)H , (·)∗ , (·)−1 and
3) Furthermore, closed-form analytical expressions are (·)† denote the transpose, Hermitian, conjugation, inverse
derived for the Cramér-Rao bounds (CRBs) of the and pseudoinverse of a matrix or a vector, respectively. The
error covariance matrices pertaining to the estimation notation 0M ×N denotes matrix of zeros of size M × N . The
of the RCS coefficients and Doppler velocities for the quantity vec(.) denotes the vector obtained by stacking the
stationary and mobile target/radar platform scenarios. columns of a matrix. The quantity [a]i denotes the ith element
4) Exhaustive simulation results are presented to char- of vector a. The quantities |·| and 6 (·) represent the magnitude
acterize the RCS coefficient and Doppler velocity and phase of a complex quantity. The matrix Kronecker
estimation performance of the proposed techniques in product is denoted by ⊗ whereas the l2 -norm and the l0 norm
mMR systems. Results demonstrate that the proposed are represented by k · k2 and k · k0 , respectively. Cardinality
schemes exploit the inherent sparsity of the clutter of A is denoted as |A|. The quantity w ∼ CN (α, 6) denotes
environment to yield a significant improvement in a circularly symmetric complex Gaussian random vector w
comparison to the conventional linear minimum mean with mean vector α and covariance matrix 6. The statistical
squared error (LMMSE) estimator [44] that does not expectation operator is denoted by E{·}.
leverage the sparsity, as well as the focal under-
determined system solver (FOCUSS) [34] that is not II. mmWave MIMO RADAR SYSTEM MODEL
as efficient in utilizing the sparsity. Consider a co-located mMR system with NT transmit
The rest of the paper is organized as follows. The mMR antennas (TAs), NR receive antennas (RAs), NRF T transmit

system model for stationary targets is developed in section II, R

radio-frequency chains (RFCs) and NRF receive RFCs,
followed by the RCS coefficient estimation model. Section III where NRF T  NT and NRF R  NR . The hybrid
describes the proposed OmMR algorithm for sparse estima- analog-digital beamforming (HAD) architecture leads to a
tion of the RCS coefficient vector, followed by radar imaging. significant reduction in the number of RF chains required
In Section IV, we extend the stationary system model of for beamforming. The schematic block-diagram of the mMR
Section II to incorporate target as well as radar platform system is given in Fig. 1. We commence by considering the
mobility. The SmMR algorithm is subsequently developed system model for stationary targets and a stationary radar
for joint RCS coefficient and Doppler velocity estimation. platform. A comprehensive system model for mobile targets
Section V derives the closed form analytical expressions for and a mobile radar platform is considered in section IV.

VOLUME 10, 2022 132839

 M. Jafri et al.: Sparse Parameter Estimation and Imaging in mmWave MIMO Radar Systems

TABLE 1. Comparison of the salient contributions of existing and present works.

where v̄(d) ∈ CNR ×1 represents the Gaussian noise vector

at the receiver with covariance matrix σ̄v2 INR . The symbol ∗
denotes the linear convolution between the RCS coefficients
α̃(q, d) and probing signal vector x̄(d). The mMR transmitter
T
consists of the RF transmit precoder (TPC) FRF ∈ CNT ×NRF .
Similarly, the receiver end is equipped with the RF receive
R
combiner (RC) WRF ∈ CNR ×NRF . Note that the RF TPC
FRF and RF RC WRF , respectively, comprise digitally
controlled phase-shifters with a constant magnitude for each
element. Let λ denote the signal wavelength and dT and dR
denote the spacings between the antenna elements of the
FIGURE 2. mMR frame structure for stationary targets.
transmitter and receiver arrays, respectively. The steering
array response vectors aT (q) ∈ CNT ×1 and aR (q) ∈
CNR ×1 , at the transmitter and receiver, respectively, are
The scattering environment can be segmented into Q given by
angular bins and R range bins. Consider L targets to be
randomly dispersed within the scattering environment at h iT
aT (q) = 1 e−j λ dT (1) sin θq · · · e−j λ dT (NT −1) sin θq ,
2π 2π
some unidentified angles and ranges from the radar. Assume
all the targets to be in the far field of the radar system, i.e., (2)
the distance between the targets and the radar platform is h iT
substantially larger than the dimension of the antenna array. aR (q) = 1 e λ dR (1) sin θq
−j 2π
··· e λ dR (NR −1) sin θq
−j 2π
.
The signal echoes are assumed to bounce off all the targets (3)
of interest and reach the radar transceiver (TRX). Thus, the
signals reaching the targets from the TAs and the signal echos
arriving at RAs can be assumed to be parallel. Let α̃(q, r) To simplify the signal processing in a wideband mMR
denote the RCS coefficient associated with a generic target system, this study utilizes a frequency-domain (FD) equal-
located at an angle θq and range Rr from the radar platform. ization technique [45]. One can obtain the corresponding
The transmitted signal is reflected back by a target located at FD representation by using the well-known overlap-and-add
a range Rr , and received at the radar platform after a round principle. The model to estimate the RCS coefficients is
trip delay of τr = 2Rc /c, where c is the speed of light. The developed next.
RCS coefficient α̃(q, r) is set to zero for the (q, r)th bin if no
target is present in the bin. We consider a practical scenario A. RCS COEFFICIENTS ESTIMATION MODEL
where prior information regarding the number of targets and The probing signal matrix consisting of D snapshots X(d) =
T
their angle and range bins is unknown to the radar transceiver [x̄(0), · · · , x̄(D − 1)] ∈ CNRF ×D is grouped into M sub-
T
(TRX). Let x̄(d) ∈ CNRF ×1 0 ≤ d ≤ D − 1 be the initial blocks, where the mth block, 0 ≤ m ≤ M − 1, is of
T
probing signal vector at time instant d sent by the radar TRX length D0 < D as shown in Fig. 2. Let x̄(m, d 0 ) ∈ CNRF ×1
with covariance matrix 6 x = E{x̄(d)x̄H (d)}. The received 0 0 0
represent the d th, 0 ≤ d ≤ D − 1, probing vector of
R
echo ȳ(d) ∈ CNRF ×1 reflected by the L targets and summed the mth block. The ‘‘Add ZP’’ block in Fig. 1 pads all the
over all Q angular bins at time instant d can be written as M blocks with R − 1 zeros. The new block length of the
zero-padded (ZP) probing sequence equals K = D0 + R −
−1
Q−1
X 1 with x̄(m, l)K l=0 denoting the probing sequences. The RCS
RF aR (q)aT (q)FRF (x̄(d) ∗ α̃(q, d))
WH H
ȳ(d) = coefficient parameter sequence α̃(q, d 0 ) of length R is padded
q=0 with D0 − 1 zeros to obtain the ZP sequence α̃(q, l)K −1
l=0 . One
+WH
RF v̄(d), (1) can represent the ZP probing sequence and RCS coefficient

132840 VOLUME 10, 2022

M. Jafri et al.: Sparse Parameter Estimation and Imaging in mmWave MIMO Radar Systems

sequence as where elements of γ̃ [k] ∈ CQ×1 denote the kth component

−1 of the K -point FFT of the RCS coefficients α̃(q, l). Thus, (8)
{x̄(m, l)}K
l=0 can be re-written as
= {x̄(m, 0), x̄(m, 1), · · · , x̄(m, D0 − 1), 0, · · · , 0}, (4)
| {z } y[m, k] = (m)
¯ γ̃ [k] + ṽ[m, k], (12)
R−1
{α̃(q, l)}K −1
l=0
where ṽ[m, k] = WH
 v[m, k], which
RF,m  follows the
= {α̃(q, 0), α̃(q, 1), · · · , α̃(q, R − 1), 0, · · · , 0}. (5) Gaussian distribution CN 0N R ×1 , σ̄v 6N R , with 6N R =
2
| {z } RF RF RF
D0 −1 K WH RF,m WRF,m as its covariance matrix. One can now
The received signal vector after RF combining ȳ(m, l) ∈ horizontally concatenate y[m, k] across all the K sub-
R
CNRF ×1 , 0 ≤ l ≤ K − 1 for the mth block, corresponding carriers, to form the output receive matrix Y(m) =
R
to the time index l, is given by [y[m, 0] y[m, 1] · · · y[m, K − 1]] ∈ CNRF ×K . The sparse
parameter estimation model for the mMR system can be
Q−1
X reformulated as
RF,m aR (q)aT (q)FRF,m (x̄(m, l) ~K α̃(q, l))
WH H
ȳ(m, l) =
q=0 Y(m) = (m)0
¯ + V(m), (13)
+WH
RF,m v̄(m, l), (6)
 
where V(m) = ṽ[m, 0] ṽ[m, 1] · · · ṽ[m, K − 1] ∈
R
CNRF ×K and 0 = γ̃ [0] γ̃ [1] · · · γ̃ [K − 1] ∈ CQ×K
 
T
NT ×NRF R
NR ×NRF
where FRF,m ∈ C and WRF,m ∈ C represent
the RF TPC and RF RC, for the mth block, respectively, denote the concatenated noise matrix and the RCS coefficient
and ~K denotes the K-point circular convolution between matrix in the frequency-domain, respectively.
the sequences x̄(m, l) and α̃(q, l). The mth frequency domain Utilizing the zero-padding done in the time domain, one
−1 R ×K can express the frequency-domain RCS coefficient matrix 0
(FD) output block {y[m, k]}K k=0 ∈ C
NRF is obtained as
as 0 = 0̄ 0Q×(D0 −1) F where F ∈ CK ×K represents the

 
{y[m, k]}K −1 K −1 discrete Fourier transform (DFT) matrix. This
T can be further
k=0 = FFT {ȳ(m, l)}l=0
simplified as 0 = 0̄ 0Q×(D0 −1) FT1 FT2 = 0̄F1 , where
 
= {y[m, 0], y[m, 1], · · · , y[m, K − 1]}.
0̄ ∈ CQ×R and F1 ∈ CR×K represent the RCS coefficient
(7) matrix in the time-domain and DFT matrix corresponding
The received signal vector for the mth block and kth to the non ZP component, respectively. The RCS coefficient
subcarrier can be written as matrix 0̄ in the time-domain is as follows
α̃(0, 0) · · · α̃(0, R − 1)
 
Q−1
.. ..
X
y[m, k] = WH H
RF,m aR (q)aT (q)FRF,m x[m, k]α[q, k] 0̄ = 

. . ,

(14)
q=0
α̃(Q − 1, 0) · · · α̃(Q − 1, R − 1)
+WH
RF,m v[m, k], (8)
where α̃(q, r) represents the RCS coefficient of the target
T
where v[m, k] ∈ CNR ×1 and x[m, k] ∈ CNRF ×1 represent the present in the (q, r)th bin corresponding to angle θq and range
kth FFT-point of the noise vector v(m, l) and probing signal Rr with respect to the radar TRX. The coefficient α̃(q, r) is
vector x̄(m, l), respectively. Also, note that variance of each either zero or non-zero depending on absence or presence,
element of v[m, k] is σv2 = K σ̄v2 . The TD probing signal respectively, of a target in the (q, r)th bin. Substituting 0 =
vector for the mth block is given as 0̄F1 in (13), the parameter estimation model for the mth block
can be written as
l=0 = x(m) × IFFT(s[m, k])k=0 ,
−1 K −1
x̄(m, l)K (9)
Y(m) = (m)
¯ 0̄F1 + V(m), (15)
where {s[m, k]} represents a known random sequence of
symbols with s[m, k] ∈ {+1, −1}. After multiplying both Applying the vec(·) operator, i.e., y(m) = vec (Y(m)) to (15),
R
sides of (8) with s[m, k], one can define ω̄(m, q) ∈ CNRF ×1 as one can reformulate the above model as

ω̄(m, q) = WH H
RF,m aR (q)aT (q)FRF,m x(m). (10) y(m) = (FT1 ⊗ (m))
¯ γ̄ + v(m), (16)
| {z }
Q−1 (m)
Next, concatenate {ω̄(m, q)}q=0 for all the angular bins to
R R K ×QR
obtain the common sensing matrix (m) ¯ ∈ CNRF ×Q across where (m) ∈ C NRF = (FT1 ⊗ (m)),
¯ while γ̄ ∈
R
all the subcarriers of the mth block as CQR×1 = vec(0̄) and v(m) ∈ CNRF K ×1 = vec(V(m))
represent the vectorized form of the RCS coefficients and
(m)
¯ = [ω̄(m, 0) ω̄(m, 1) · · · ω̄(m, Q − 1)] . (11)
noise matrix, respectively. Next, one can stack the received
The RCS coefficient vector for kth subcarrier is represented output vectors y(m) for all the M blocks resulting in the final
as estimation model
γ̃ [k] = [α[0, k] α[1, k] · · · α[Q − 1, k]]T , y =  γ̄ + v, (17)

VOLUME 10, 2022 132841

 M. Jafri et al.: Sparse Parameter Estimation and Imaging in mmWave MIMO Radar Systems

Algorithm 1 OmMR for Sparse RCS Estimation

Input: Sensing matrix , output vector y, dictionary matrices
AR (8) and AT (8), stopping parameter t
Initialization: Index set I = [ ], I = [ ], residue vectors
r−1 = 0N R KN ×1 , r0 = y, b̄
γ OmMR = 0QR×1 , counter i = 0
 RF
while k ri−1 k22 − k ri k22  ≥ t do


1) i ← i + 1 
i−1 (:, k)

2) j = arg max rH 
k=1,··· ,QR
3) I = I ∪ j
4) I = (:,
 I)†
i FIGURE 3. mMR frame structure for mobile targets and radar platform.
γ LS = I y
5) b̄
i
6) ri = y − I b̄
γ LS
end while columns indexed by the set I. Next, Step-5 obtains the
i
γ OmMR (I) = b̄
b̄ γ LS intermediate least squares (LS) solution b̄
i
γ LS using the
0 OmMR = vec−1 b̄
γ OmMR


Output: b̄ submatrix I . Finally, the associated residue vector ri is
computed using I in Step-6. These steps are iterated until
the difference between the successive residuals becomes
R R
sufficiently small, i.e., k ri−1 k22 − k ri k22  < t , where

where  ∈ CNRF KM ×QR and v ∈ CNRF KM ×1 are obtained
by vertically stacking the sensing matrices (m) and noise t is a suitable threshold. Finally, the estimate of the RCS

0 OmMR ∈ C 0 OmMR =
vectors v(m) for all m. Note that the RCS coefficient vector coefficient matrix b̄ Q×R is obtained as b̄

γ̄ has a sparse structure, i.e., most of the entries of γ̄ are vec −1 γ OmMR . The central advantages of the proposed
b̄
close to zero, with only a few of them taking significant non- OmMR algorithm are that it yields a sparse solution and has a
zero values. One can now observe that (17) represents a well- low computational complexity. It is important to note that the
known sparse signal recovery problem. It is worth noting that choice of the stopping parameter t plays a prominent role in
the sparsity of the RCS coefficient vector was not utilized determining the convergence of the OmMR algorithm.
in [27]. To address this shortcoming, in this work, we develop The radar image can now be obtained as follows. The
the appropriate system model to fully exploit sparsity toward element b αOmMR (q, r) of the matrix b̄ 0 OmMR corresponds to
efficient and accurate estimation of the RCS parameters. the estimate of the RCS coefficient of the target present in the
(q, r)th bin. Plotting the magnitudes of the RCS coefficients
III. OmMR-BASED RCS ESTIMATION AND RADAR across the angle/range bins yields the radar intensity image.
IMAGING One can also estimate the number of targets as the number
For the radar cross-section (RCS) coefficient estimation of non-zero entries of the RCS coefficient matrix. The next
model of (17), one can now formulate the following section develops the system model for RCS and Doppler
optimization problem to determine a sparse estimate of the velocity estimation for a scenario with mobile targets and
RCS coefficient vector γ̄ as radar platform.
min kγ̄ k0
γ̄ IV. RCS ESTIMATION AND RADAR IMAGING WITH
MOBILITY
s.t. ky − γ̄ k22 ≤ t , (18)
The model developed for stationary targets in the previous
where k·k0 represents the l0 -norm, which equals the number section can be readily extended for mobile targets incorporat-
of non-zero elements in the vector [46], and t represents ing also mobility of the radar platform. The frame-structure
a tunable parameter, which can be adjusted to vary the for the mobile target and radar platform is given in Fig. 3. The
observational error. Note that, the optimization problem in relative Doppler velocity υl associated with the lth target can
(18) is non-convex [47], which renders it challenging to solve be related to radar platform’s linear velocity υr , and target’s
using conventional optimization techniques. Therefore, the linear velocity ύl , as υl = (υr − ύl ) cos θl , where θl is the
OmMR procedure is described next toward sparse estimation. angle between the lth target and radar platform. Consider that
the pth pulse, x̄(m, p, l), 0 ≤ p ≤ P − 1 in the mth block
The key steps of the OmMR technique are described to be subjected to a Doppler shift of 2π υl Tm m/λ. The block
next. These are also summarized in Algorithm 1. In Step- pulse repetition interval Tm is related to the symbol period T
2, the algorithm selects the column of the sensing matrix as Tm = KPT . Hence, one can recast the system model of
 that attains the maximum projection along residue ri−1 . (16) for the mobile scenario as
Step-3 updates the index-set I by incorporating the index
j determined in Step-2. Subsequently, Step-4 builds the y(m, p) = (m, p) diag (d(m)) γ̄ +v(m, p), (19)
submatrix I of the sensing matrix  by extracting the
| {z }
h(m)

132842 VOLUME 10, 2022

M. Jafri et al.: Sparse Parameter Estimation and Imaging in mmWave MIMO Radar Systems

where d(m) ∈ CQR×1 is the Doppler shift vector with Algorithm 2 SmMR for Sparse Scattering Scene Estimation
the (q + rQ)th element defined as ej2π υq,r Tm m/λ , where υq,r in Mobile mMR Systems
indicates the relative Doppler velocity of the target present in Input: Dictionary matrix , observation matrix Y and
the (q, r)th bin. The scattering scene vector h(m) ∈ CQR×1 is stopping parameter s
defined as the element-wise product of the RCS coefficient Initialization: I = [ ], residue R0 = Y, R−1 = 0 H
bSmMR =
I
vector γ̄ and Doppler shift vector d(m) which can be stated 0QR×M , counter k = 0  = [ ]
e
as h(m) = diag (d(m)) γ̄ = diag (γ̄ ) d(m). Assume that the while k Rk−1 k22 − k Rk k22  ≥ s do


scattering scene vector h(m) is the same for all the P pulses 1) k ← k + 1
in the mth block. This assumption is justified for an mMR 2) 9 = H Rk−1 
3) i(k) = arg max 99 H l,l

system due to the small symbol period T . Keeping the sensing
matrix (m, p) identical for all the M blocks, the received eIk =  eIk−1 |ω(i(k))
h i
R 4) 
signal vector y(m, p) ∈ CNRF K ×1 in the mth block for the pth −1
eIk )H 
eIk eIk )H Y

pulse is given by bSmMR = (
5) H (
eI H
Y−
y(m, p) = (p) h(m) + v(m, p). k SmMR
b
(20) 6) Rk = eI
kY−k HSmMR kF
b
end while
Stacking the received signal y(m, p) across all the p =
Return: H
1, · · · , P pulses, the received signal vector in the mth block
bSmMR
R
i.e., y(m) ∈ CNRF KP×1 can be obtained as
(1)
     
y(m, 1) v(m, 1) where k·kF represents the Frobenius-norm. Algorithm 2
 ..   .   . 
 .  =  ..  h(m) +  ..  , describes the proposed SmMR technique for estimation of
the matrix H. Step-1 performs a projection of the dictionary
y(m, P) (P) v(m, P)
| {z } | {z } | {z } column, followed by greedy selection of the columns of 
R
y(m)∈CNRF KP×1
R
∈CNRF KP×QR
R
v(m)∈CNRF KP×1
that has the maximum projection along the residue, similar
in principle to the OmMR technique. Step-3 updates the
which can be compactly written as submatrix  eI using indices found in Step-2. Subsequently,
y(m) =  h(m) + v(m). (21) the LS estimate of the scattering scene matrix H bSmMR
for the kth iteration is obtained in Step-4. Finally, Step-5
The overall received signal matrix Y can be written as obtains the associated normalized residue Rk . These steps are
repeated for 1 ≤ k ≤ QR, after the completion of which
Y =  H + V, (22) one obtains the scattering scene matrix estimate H bSmMR .
R R To detect the presence of the target in a particular angle-
where Y ∈ CNRF KP×M , H ∈ CQR×M and V ∈ CNRF KP×M ,
range bin, an appropriate threshold ηth ( 1) is chosen.
which are determined as
The target is considered to be present in the (q, r)th bin if
1 PM −1 b̄
Y = [y(0) y(1) · · · y(M − 1)] , M m=0 |[hSmMR (m)]q+rQ | ≥ ηth .
H = [h(0) h(1) · · · h(M − 1)], One can now extract the RCS coefficients and Doppler
velocities from the estimated scattering scene matrix H
bSmMR .
V = [v(0) v(1) · · · v(M − 1)], (23)
Note that the mth column ĥSmMR (m) of the matrix H bSmMR ,
denote the received signal matrix, scattering scene matrix corresponding to the mth block, represents an estimate of
and noise matrix obtained after horizontal concatenation, the scattering scene vector h(m) = diag (γ̄ ) d(m). The
respectively. The scattering scene matrix H has a unique relation between the RCS coefficient vector γ̄ and columns
simultaneous sparse structure. Each column of H has the of scattering scene matrix H can be expressed as
same sparsity profile which results in several zero rows and M −2
1 X
a few non-zero rows. The SmMR-based technique can be (diag (h(m + 1)))−m (diag (h(m)))m+1
efficiently employed to obtain an estimate of the scattering M −1
m=0
scene matrix H, which is discussed next. M −2
1 X
= (diag (γ̄ ))−m (diag (d(m + 1)))−m
A. SmMR-BASED JOINT ESTIMATION OF RCS M −1
m=0
COEFFICIENTS AND DOPPLER VELOCITIES × (diag (d(m)))m+1 (diag (γ̄ ))m+1 = diag (γ̄ ) ,
The optimization problem for estimation of the scattering
scene matrix H for a mobile radar and target scenario can be where the last equality stems from the relation
stated as follows
(diag(d(m + 1)))−m (diag(d(m)))m+1 = IQR .
arg min kY − HkF
Thus, one can relate the estimate of the RCS coefficient vector
s.t. diag(HHH ) = QR, (24) γ SmMR to the columns of estimated scattering scene matrix

b̄
0

VOLUME 10, 2022 132843

 M. Jafri et al.: Sparse Parameter Estimation and Imaging in mmWave MIMO Radar Systems

H
bSmMR as Lemma 1: The CR bounds for the RCS coefficient vector
M −2  γ̄ and target Doppler velocity vector υ for a system with
1 X  −m
γ SmMR = mobility can be expressed as


diag b̄ hSmMR (m + 1)
diag b̄
M −1
m=0
CRB γ̄ ef


  m+1 
× diag b̄hSmMR (m) . (25)
i−1
h
= F γ̄ ef , γ̄ ef − F γ̄ ef , υ F −1 (υ, υ) F υ, γ̄ ef ,

Toward estimation of the Doppler velocity υl of lth target (30)

from the scattering scene matrix, consider the ratio CRB (υ)
[ γ̄ ]l ej2π υl Tm (m+1)/λ
i−1
  h
h̄(m + 1) l = F (υ, υ) − F υ, γ̄ ef F −1 γ̄ ef , γ̄ ef F γ̄ ef , υ



,
= = ej2π υl Tm /λ . (26)
[ γ̄ ]l ej2π υl Tm m/λ
 
h̄(m) l (31)
Thus, one can estimate the Doppler velocity υl , 0 ≤ l ≤
L̃ − 1 as where the matrix F(γ̄ ef , γ̄ ef ) ∈ R2QR×2QR can be derived as
h i   
M −1 P−1
M −2 hSmMR (m + 1)
b̄
λ
X X
F γ̄ ef , γ̄ ef = 2 I2 ⊗ diag(d(m)) 6 c diag(d(m)) .


1 X q̃ +r̃ Q 
υl =
b 6 
 h i l l ,
2π Tm M − 1 h
b̄ (m) m=0 p=0
m=0 SmMR
q̃l +r̃l Q (32)
(27)
Further, the components of the Fisher information matrices
where (q̃l , r̃l ) denotes the estimated angle and range index F(γ̄ ef , υ) ∈ R2QR×L and F(υ, γ̄ ef ) ∈ RL×2QR are given as
pair of the lth target and b L = |Ĩ| is the total number of
M −1 P−1
estimated targets. Note that the estimates of the RCS coeffi- X X
F γ̄ R , υl = 2 diag(d(m)) 6 c diag(ḋl (m))γ̄ R ,



cients and the Doppler velocities are non-linear functions of
H
bSmMR . Therefore, minor errors in locations of the targets can m=0 p=0
result in large mean squared errors (MSEs) corresponding to M
X −1 P−1
X
F γ̄ I , υl = 2 diag(d(m)) 6 c diag(ḋl (m))γ̄ ,



estimation of the RCS coefficients and the Doppler velocities. I
However, the SmMR technique can overcome these issues m=0 p=0
efficiently due to its superior ability for sparse recovery. (33)
The next section derives the Cramér-Rao bounds pertaining ∂d(m)
where the vector ḋl (m) = ∈
∂ul CQR×1
is given as
to the MSEs of RCS and target Doppler velocity estimation
in mMR systems. T
ḋl (m) = 0, · · · , j 2π
λ mTm [d(m)]ql +rl Q , · · · , 0 , (34)


V. Cramér-RAO BOUNDS and the non-zero entry in the vector ḋl (m) is at the location
The CRBs are computed in this section for estimation of the ql + rl Q for the lth target corresponding to the angle-
RCS coefficient vector γ̄ and Doppler velocity vector υ = range bin (ql , rl ). Note that F(υl , γ̄ R ) = (F(γ̄ R , υl ))T and
[υ1 , · · · , υL ]T , for mMR systems with and without mobility, F(υl , γ̄ I ) = (F γ̄ I , υl ))T . The Fisher Information matrix
as derived in equations (16) and (21). Let the effective F(υ, υ) ∈ RL×L has elements F(υl , υl 0 ), 1 ≤ l, l 0 ≤ L,
RCS coefficient vector be defined as γ̄ ef = [γ̄ TR , γ̄ TI ]T ∈ which can be written as
R2QR×1 , where the real and imaginary components of the M −1 P−1
RCS coefficient vector are given by γ̄ R and γ̄ I , respectively.
X X
F (υl , υl 0 ) = 2γ̄ H diag(ḋ∗l (m)) 6 c diag(ḋl 0 (m))γ̄ .
The expression to obtain the Fisher information matrix from m=0 p=0
the log-likelihood function L(y; 2) is given as (35)
∂ ∂L(y; 2)
  
F(w, x) = −E , (28) Proof: The
 log likelihood L(y; 2) of the
T received output
∂w ∂xT R KPM ×1
vector y = y(0) , y(1) , · · · , y(M − 1)
T T T ∈C NRF

where 2 = [γ̄ Tef , υ T ]T ∈ R2QR+L×1 is the vector of is given as

unknown parameters obtained via concatenating γ̄ ef and υ.
Thus, the derived CRBs are applicable for the scenarios L(y; 2)
without and with mobility, as is appropriate for the particular = −κ
system. The CRB for joint estimation of the RCS coefficients M
X −1 P−1
X
in γ̄ and Doppler velocities in υ is given as − kyR (m, p) − R (p)hR (m) + I (p)hI (m)k2

−1 m=0 p=0
F γ̄ ef , γ̄ ef
F γ̄ ef , υ


CRB (2) = . (29) M −1 P−1
F υ, γ̄ ef F (υ, υ)
X X
− kyI (m, p) − I (p)hR (m) − R (p)hI (m)k2 ,
The following Lemma derives the equations for the compo- m=0 p=0
nent Fisher information matrices. (36)

132844 VOLUME 10, 2022

M. Jafri et al.: Sparse Parameter Estimation and Imaging in mmWave MIMO Radar Systems

R
MPKNRF ∂ L(y;2)
where the constant κ = 2 ln π. The quantities ∂ γ̄ R ,
∂ L(y;2) ∂ L(y;2)
∂ γ̄ I , ∂υl can be determined as shown in (37), (38) and
(39), as shown at the bottom of the page, respectively.
Other pertinent terms can be obtained as follows
!
∂ ∂L(y; 2)
∂ γ̄ R ∂ γ̄ TR
!
∂ ∂L(y; 2)
= (41)
∂ γ̄ I ∂ γ̄ TI
M
X −1 P−1
X
= −2 diag (d(m)) T (p)(p)diag (d(m)) ,
m=0 p=0
!
∂ ∂L(y; 2)
∂ γ̄ R ∂ γ̄ TI FIGURE 4. Radar image showing the true target locations, i.e., angle and
range bins for a stationary scenario.
!
∂ ∂L(y; 2)
= = 0QR,QR . (42)
∂ γ̄ I ∂ γ̄ TR
M −1 P−1
∂ ∂L(y; 2) X 
  X 
− diag ḋl (m) T (p)v(m, p)


= −2
∂ γ̄ R ∂υl m=0 p=0
I
!!T
∂∂L(y; 2)  


= + diag (d(m)) T (p)(p)diag ḋl (m) γ̄ . (44)
∂υl ∂ γ̄ TR I
M −1 P−1
X  The detailed procedure to derive these expressions is given
X 
− diag ḋl (m) T (p)v(m, p)


= −2
R in Appendix A. Lastly, the second-order derivative of L(y; 2)
m=0 p=0
  with respect to the target Doppler velocity υl is given in (40),


+ diag (d(m)) T (p)(p)diag ḋl (m) γ̄ , (43) as shown at the bottom of the page, where d̈l (m) ∈ CQR×1 is
R given as
∂ ∂L(y; 2)
 

∂ γ̄ I ∂ul ∂ 2 d(m)
d̈l (m) =
!!T ∂ 2 υl
∂ ∂L(y; 2)
= 2π
∂υl ∂ γ̄ TI = [0, · · · , (j mTm )2 [d(m)]ql +rl Q , · · · , 0]T . (45)
λ

M −1 P−1
∂L(y; 2) X X     
= −2 diag (d(m))  (p)(p)diag (d(m)) γ̄
T
− diag (d(m))  (p)y(m, p)
T
, (37)
∂ γ̄ R R R
m=0 p=0
M −1 P−1
∂L(y; 2) X X     
= −2 diag (d(m)) T (p)(p)diag (d(m)) γ̄ − diag (d(m)) T (p)y(m, p) , (38)
∂ γ̄ I I I
m=0 p=0
M −1 P−1
∂L(y; 2) X X
dH (m)diag γ̄ ∗ H (p)(p)diag (γ̄ ) ḋl (m) − 2yH (m, p)(p)diag (γ̄ ) ḋl (m)


=−
∂υl
m=0 p=0

+ḋl (m)diag γ̄  (p)(p)diag (γ̄ ) d(m) ,
H ∗

H
(39)
M −1 P−1
∂ ∂L(y; 2)
  X X
l (m)diag γ̄  (p)(p)diag (γ̄ ) ḋl 0 (m), for l 6= l 0 ,
∗
ḋH

H
= −2
∂υl 0 ∂υl
m=0 p=0
M
X −1 P−1
X 
ḋH γ̄ ∗
 (p)(p)diag (γ̄ ) ḋl (m) − v (m, p)(p)diag (γ̄ ) d̈l (m) , for l = l 0 .
H H


= −2 l (m)diag
m=0 p=0
(40)

VOLUME 10, 2022 132845

 M. Jafri et al.: Sparse Parameter Estimation and Imaging in mmWave MIMO Radar Systems

FIGURE 5. Stationary scenario: (a) normalized mean squared error vs. SNR for the proposed OmMR, FOCUSS and LMMSE techniques. (b) Impact
of number of snapshots M on the estimation of RCS coefficients using the proposed OmMR, FOCUSS and LMMSE algorithms.

FIGURE 6. Stationary scenario: reconstructed radar images obtained using (a) proposed OmMR, (b) FOCUSS, (c) LMMSE techniques at
SNR = −10 dB.

Substituting the various quantities derived above, one can Q = 32 angular bins and R = 32 range bins. The widths of
readily obtain the Fisher Information matrices which in turn the angular and range bins are 1Q = 1◦ and 1R = 1 m,
yield the CRBs in (30), (31). For the case of a stationary mMR respectively. It is assumed that the number of targets, RCS
system, the unknown parameter vector 2 = γ̄ eff ∈ R2QR×1 . coefficients, Doppler velocities, angular and range bins in
The CRB for the estimation of the RCS coefficient vector for which the targets are present are unknown to the radar TRX.
this scenario is given as The number of probing signal vectors D0 in a block is set as
1 D0 = 33 and the FFT size is K = D0 + R − 1 = 64. The
CRB(2) = (I2 ⊗ 6 −1
c ). (46) elements of the probing signal vector x(d) are drawn from
2M
an 8-PSK (phase shift keying) constellation with an average
power of unity. The RCS coefficients are randomly initialized
VI. SIMULATION RESULTS from a zero-mean and unit variance circularly symmetric
This section presents the simulation results to illustrate complex Gaussian distribution and the noise is considered
the performance of the proposed techniques for RCS and to be circularly symmetric complex Gaussian with mean
Doppler velocity estimation for stationary as well as mobile zero and unit variance. A comprehensive list of the various
targets and radar platforms. Consider an mMR TRX with simulation parameters and their values are given in Table 2.
NT = 8 transmit antennas, NR = 16 receive antennas, The results obtained for various scenarios are elaborated
T
NRF = 4 transmit RF chains and NRF R = 6 receive next.
RF chains. The inter-element spacings between the transmit
and receive antenna arrays are set to be dT = 0.7λ and A. STATIONARY TARGETS AND RADAR PLATFORM
dR = 0.9λ, respectively, where λ = 0.0107m for the For a stationary scenario, L = 10 targets were placed at
28 GHz radar frequency. The number of targets L is set some angle and range bins in the scattering environment by
equal to 10, and the scattering environment is divided into randomly selecting L grid points from the angle-range grid

132846 VOLUME 10, 2022

M. Jafri et al.: Sparse Parameter Estimation and Imaging in mmWave MIMO Radar Systems

FIGURE 7. Mobile mMR system: (a) True radar image of target RCS values; (b) True Doppler velocity image of scattering scene.

FIGURE 8. For a mobile mMR system: (a) NMSE of RCS estimation vs. SNR for various schemes; (b) NMSE of Doppler velocity estimates vs.
SNR for various schemes.

TABLE 2. Simulation parameters. range bins. The empty bins are indicated by white colors
while the colored bins show the RCS coefficient magnitudes
in dB, along with the colorbar that maps the intensity of color
to dB values given on the right. Fig. 5 shows the normalised
MSE (NMSE) performance of the proposed OmMR scheme
corresponding to M = 5 snapshots. It can be observed that
the NMSE of the RCS coefficient vector estimate decreases
with increasing signal to noise ratio (SNR), which is along
expected lines since probing signals of higher power result
in RCS estimates of higher accuracy. Furthermore, it can
also be noted that the proposed OmMR scheme performs
better than the existing FOCUSS [34] and conventional
LMMSE schemes for mMR systems. It may be observed
that for a low value of SNR = −15 dB, the performance
of the OmMR algorithm improves by approximately 20 dB
of size Q × R. Fig. 4 shows the true radar image obtained in comparison to the FOCUSS and conventional LMMSE
by plotting the RCS coefficient matrix against the angle and schemes. This is due to the fact that the OmMR efficiently

VOLUME 10, 2022 132847

 M. Jafri et al.: Sparse Parameter Estimation and Imaging in mmWave MIMO Radar Systems

FIGURE 9. For a mobile mMR system: Reconstructed RCS images obtained after thresholding with ηth = 0.1 using (a) proposed SmMR,
(b) FOCUSS, (c) LMMSE algorithms at SNR = −10 dB; reconstructed Doppler velocity images of the scattering obtained after
thresholding with ηth = 0.1 using (d) proposed SmMR, (e) FOCUSS, (f) LMMSE schemes at SNR = −10 dB.

exploits the sparsity of the RCS coefficient vector, which B. MOBILE TARGETS AND RADAR PLATFORM
leads to a significant improvement in the estimation accuracy. For the scenario with mobile targets and radar platform, the
The performance of FOCUSS suffers because it is sensitive number of pulses in a block is set as P = 10. The vector
to the regularization parameter, which ultimately leads to containing Doppler velocities pertaining to the L targets is set
convergence deficiencies. Note that the LMMSE scheme as υ = [11, 13, −13, 4, 14, 9, 16, −9, −12, 8]T . Fig. 7a
that does not exploit sparsity has the worst performance. and Fig. 7b show the true RCS and Doppler images of the
Thus, the performance of FOCUSS and LMMSE schemes radar scattering scene with the side color bars mapping the
is not competitive. On the other hand, the NMSE of the respective image bin color intensities to dB values of their
proposed OmMR scheme is seen to be close to the associated magnitudes. The threshold ηth to determine the non-zero RCS
CRB derived in Section-V, which demonstrates its efficiency. coefficient estimates, and thus obtain an estimate of the total
Remarkably, Fig. 5b shows that the OmMR algorithm for number of targets present in the scattering environment, is set
even M = 5 snapshots yields an improved performance in as ηth = 0.1.
comparison to the FOCUSS and the conventional LMMSE Fig. 8a contrasts the NMSE performance of RCS estima-
schemes with M = 10, 15 snapshots. Thus, one can draw a tion of the proposed SmMR algorithm for the mobile system
clear conclusion that OmMR can provide accurate estimates with that of the FOCUSS and LMMSE techniques along with
even in scenarios with a significantly fewer snapshots of the the associated CRB. As seen for the stationary scenario, the
scattering environment. Fig. 6a-6c show the reconstructed proposed SmMR algorithm results in lower NMSE values
radar images obtained using the OmMR, FOCUSS and as compared to those obtained from the FOCUSS and
conventional LMMSE schemes at SNR = −10 dB. One conventional LMMSE schemes. The poor performance of
can visually observe that image obtained for OmMR in the LMMSE scheme is attributed to the fact that it fails to
Fig. 6a is noticeably more accurate when compared to exploit the inherent structural sparsity of scattering scene
the images obtained using the FOCUSS and conventional matrix H. Furthermore, the poor performance of FOCUSS is
LMMSE techniques in Fig. 6b and Fig. 6c, respectively. due to its sensitivity to the regularization parameter, which
In fact, one can observe that the reconstructed OmMR image ultimately leads to convergence deficiencies. Fig. 8b plots
in Fig. 6a accurately maps all the targets to their true angular the NMSE of the Doppler velocity estimates versus SNR.
and range bins shown in Fig. 4. On the other hand, the The SmMR is once again seen to result in a significantly
images obtained using FOCUSS and LMMSE contain several improved performance in comparison to the FOCUSS and
non-zero entries in the estimated RCS coefficient matrix LMMSE schemes. Furthermore, it can be observed that
as they fail to efficiently exploit the innate sparsity of the the minimum NMSE of Doppler velocity estimates that
RCS vector and thus are not able to accurately map the can be achieved using the SmMR is approximately 10−4 ,
targets. which is comparatively higher than NMSE of the RCS

132848 VOLUME 10, 2022

M. Jafri et al.: Sparse Parameter Estimation and Imaging in mmWave MIMO Radar Systems

coefficient estimates that is close to 10−6 for the same given in equation (37) are
setting. This is due to the fact that estimation of Doppler M −1 P−1
velocity involves the phase of a ratio which is non-linear ∂L(y; 2) X X
=− − 2diag (d(m))R TR (p)yR (m, p)
in nature. A minute error in estimation of the scattering ∂ γ̄ R
m=0 p=0
scene matrix H can lead to a high NMSE for the Doppler
+2diag (d(m))I TI (p)yR (m, p)
velocity estimates. It is also worth noting that the proposed
SmMR scheme achieves an NMSE performance close to the −2diag (d(m))R TI (p)yI (m, p)
associated CRBs in Fig. 8a and Fig. 8b, which clearly demon- −2diag (d(m))I TR (p)yI (m, p)
strates its superior estimation and imaging performance. +2diag (d(m))R TR (p)R (p)diag (d(m))R γ̄ R

Fig. 9a-9c and Fig. 9d-9f show the reconstructed images
+2diag (d(m))I I (p)I (p)diag (d(m))I γ̄ R
T
for RCS coefficients and relative Doppler velocities at
SNR = −10 dB. As before, proposed scheme outperforms (48)
the FOCUSS and conventional LMMSE schemes, a nat-
Similarly, ∂ L∂(y;2)
γ̄ I can also be obtained as given in equation
ural result of the improved NMSE of estimation of the
(38). The expressions in (41) and (42) can be easily obtained
former.
with the help of the first order derivatives ∂ L∂ γ̄(y;2) , ∂ L∂(y;2)
γ̄ I .
R
The intermediate steps  to determine
 the expressions for
VII. CONCLUSION ∂ ∂
 
∂
∂ γ̄
L (y;2)
∂υl , ∂υ∂ l L (y;2)
∂ γ̄ T
can be found at the top of the
This work developed novel parameter estimation and imaging R R
schemes for co-located mMR systems for stationary as well page in equations (49) and (50), shown at the top of the next
as mobile scenarios. To begin with, the sparse estimation page, respectively.
model and an OmMR based RCS coefficient, angle and range The expressions in equation  (44) can  be obtained likewise.
estimation algorithm were developed for a scenario with sta- Finally, the quantities ∂υ∂ 0 ∂ L∂υ (y;2)
l
, 1 ≤ l, l 0 ≤ L
l
tionary targets and radar platform. Next, the SmMR technique are given in equation (51), shown at the top of the
next page, where the d̈l (m) ∈ CQR×1 = ∂ ∂d(m)
2
was also developed to estimate the various parameters along 2υ =
l
with target Doppler velocities considering target and radar
[0, · · · (j λ mTm ) [d(m)]ql +rl Q , · ·
· , 0] . Hence,
2π 2 T

the Fisher

mobility. Subsequently, the relevant CRBs were also derived
information matrices F γ̄ ef , γ̄ ef , F γ̄ ef , υ , F υ, γ̄ ef
to rigorously characterize the error covariance for estimation
and F (υ, υ) required to compute the CRBs given in
of the RCS coefficients and Doppler velocities of multiple
equations (30) and (31) are as follows.
targets. Simulation results demonstrated that the proposed 1) To begin with,
the Fisher
information
matri-
techniques yield a significant improvement in performance ces F γ̄ R
, γ̄ R , F γ̄ I , γ̄ I , F γ̄ R , γ̄ I and
in comparison to other conventional and competing sparse F γ̄ I , γ̄ R can be found using equations (41) and (42)
techniques. given in section V.
F γ̄ R , γ̄ R = F γ̄ I , γ̄ I



APPENDIX A M −1 P−1
The various derivatives of the log likelihood function =2
X X
diag (d(m)) 6 c diag (d(m)) ,
L(y; 2) with respect to the unknown joint parameter
m=0 p=0
2 = [γ̄ TR , γ̄ TI , υ T ]T ∈ R2QR+L×1 as mentioned in (52)
Lemma 1 of section V are derived here. The log likelihood F γ̄ R , γ̄ I = F γ̄ I , γ̄ R = 0QR,QR ,



(53)
function L(y; 2) is given as
where 6 c = E{T (p)(p)}. Subsequently one
can derive the expression for F γ̄ ef , γ̄ ef given in


M
X −1 P−1
X
L(y; 2) = −κ − ky(m, p) − (p)h(m)k2 , (47) Lemma 1 of section V.


F γ̄ R , υl , F υl , γ̄ R and F γ̄ I , υl ,



m=0 p=0 2) The vectors
F υl , γ̄ I are obtained using the expressions (43)
MPKN R
and (44) respectively and are mentioned in Lemma
where κ = ln π and h(m) = diag(γ̄ ) d(m) =

V. Therefore, the matrices F γ̄ ef , υ and
RF


2 1 of section
diag(d(m)) γ̄ . The quantity ∂ L∂υ(y;2)
l
obtained by differ- F υ, γ̄ ef are obtained as
entiating the likelihood function with respect to υl is 

T
given in (39) in section V. Further, one can express the F γ̄ ef , υ = F γ̄ R , υ1 , · · · , F γ̄ R , υL ,



log likelihood function L(y; 2) in terms of the real and | {z }
imaginary parts of the vector y(m, p), h(m) and matrix F (γ̄ R ,υ )

(p) as shown in equation (36). Substituting the expression 
F γ̄ I , υ1 , · · · , F γ̄ I , υL


T
h(m)R = diag(d(m))R γ̄ R + diag(d(m))I γ̄ I and h(m)I = | {z }
diag(d(m))R γ̄ I + diag(d(m))I γ̄ R in the log likelihood F (γ̄ I ,υ )
function, the intermediate steps to obtain the vector ∂ L∂ γ̄(y;2)

T
R
= F υ, γ̄ ef . (54)

VOLUME 10, 2022 132849

 M. Jafri et al.: Sparse Parameter Estimation and Imaging in mmWave MIMO Radar Systems

M −1 P−1
∂ ∂L(y; 2)
  X   
X

− 2 diag ḋl (m)∗ H (p)y(m, p) + 2 diag d(m)∗ H (p)(p)diag ḋl (m) γ̄



=−
∂ γ̄ R ∂υl R R
m=0 p=0
  
∗
 (p)(p)diag (d(m)) γ̄
H
,


+2 diag ḋl (m) (49)
R

M −1 P−1
!
∂ ∂L(y; 2) X X 
 

=− − 2 y(m, p)T (p)diag ḋl (p) + 2 γ̄ T diag (d(m)) T (p)(p)diag ḋl (m)
∂υl ∂ γ̄ TR R R
m=0 p=0
  
+2 γ̄ diag ḋl (m)  (p)(p)diag (d(m))
T
,

T
(50)
R

M −1 P−1
∂ ∂L(y; 2)
  X X
− 2yH (m, p)diag (γ̄ ) d̈l (m) + 2dH (m)diag γ̄ ∗ (m)H (m)diag (γ̄ ) d̈l (m)


=−
∂υl0 ∂υl
m=0 p=0

+2ḋH γ̄ ∗
 (p)(p)diag (γ̄ ) ḋl (m) , for l = l 0 ,
H


l (m)diag

M
X −1 P−1
X 
H
(m, p)(p)diag (γ̄ ) d̈l (m) + 2ḋH γ̄ ∗
 (p)(p)diag (γ̄ ) ḋl (m) , for l = l 0 ,
H


=− − 2v l (m)diag
m=0 p=0
M
X −1 P−1
X 
− 2ḋH γ̄ ∗
 (γ̄ ) , for l 6= l 0 ,

H
=− l 0 (m)diag (p)(p)diag ḋl (m) (51)
m=0 p=0

MKN R
3) Finally, using the result for F (υl 0 , υl ), 1 ≤ l, l 0 ≤ L where κ = 2
RF
ln π. The individual components of the
Fisher information matrix F γ̄ ef ∈ R2QR×2QR are given as


given in Lemma 1 of section V, one obtains the Fisher
information matrix F (υ, υ) invoking the relation
∂ ∂ T
 
F γ̄ R , γ̄ R = −E L(y; γ̄ )


∂ γ̄ R ∂ γ̄ R
F (υ1 , υ1 ) · · · F (υ1 , υL )
 
M −1  
.. ..
X
F (υ, υ) = 

. . . (55)
 =2 E  (m)(m) = 2M 6 c ,
T

F (υL , υ1 ) · · · F (υL , υL ) m=0

∂ ∂ T
 
F γ̄ I , γ̄ I L(y; γ̄ )


= −E
∂ γ̄ I ∂ γ̄ I
The CRB for the estimation of the RCS coefficient vector
M −1 
γ̄ ef for the stationary mMR system is given as follows. The X 
=2 E  (m)(m) = 2M 6 c ,
T
log likelihood function L(y; γ̄ ef ) of the vector received at the
T R m=0
output y = yT (0), · · · , yT (M − 1) ∈ CNRF KM ×1 with

∂ ∂ T
 
F γ̄ R , γ̄ I = −E L(y; γ̄ ) = 0QR,QR ,


y(m) defined in equation (16) is given as
∂ γ̄ R ∂ γ̄ I
∂ ∂ T
 
F γ̄ I , γ̄ R = −E L(y; γ̄ ) = 0QR,QR .


L(y; γ̄ ef ) ∂ γ̄ I ∂ γ̄ R
= −κ Using these results, one can readily obtain CRB(γ̄ ef ) as
MX −1 P−1
X
− kyR (m, p) − R (p)γ̄ R (m) + I (p)γ̄ I (m)k2 1  
CRB γ̄ ef = I2 ⊗ 6 −1 .


c (57)
m=0 p=0 2M
M −1 P−1
X X REFERENCES
− kyI (m, p) − I (p)γ̄ R (m) − R (p)γ̄ I (m)k2 ,
[1] X. Li, X. Wang, Q. Yang, and S. Fu, ‘‘Signal processing for TDM
m=0 p=0 MIMO FMCW millimeter-wave radar sensors,’’ IEEE Access, vol. 9,
(56) pp. 167959–167971, 2021.

132850 VOLUME 10, 2022

M. Jafri et al.: Sparse Parameter Estimation and Imaging in mmWave MIMO Radar Systems

[2] C. Vasanelli, R. Batra, A. D. Serio, F. Boegelsack, and C. Waldschmidt, [25] W. Roberts, P. Stoica, J. Li, T. Yardibi, and F. A. Sadjadi, ‘‘Iterative adaptive
‘‘Assessment of a millimeter-wave antenna system for MIMO radar appli- approaches to MIMO radar imaging,’’ IEEE J. Sel. Topics Signal Process.,
cations,’’ IEEE Antennas Wireless Propag. Lett., vol. 16, pp. 1261–1264, vol. 4, no. 1, pp. 5–20, Feb. 2010.
2017. [26] A. Rawat, S. Dwivedi, and A. K. Jagannatham, ‘‘BLMS and FBLMS-based
[3] K. Sakaguchi, R. Fukatsu, T. Yu, E. Fukuda, K. Mahler, R. Heath, adaptive time varying RCS estimation and 2D-imaging for monostatic
T. Fujii, K. Takahashi, A. Khoryaev, S. Nagata, and T. Shimizu, ‘‘Towards MIMO radar systems,’’ in Proc. 23rd Nat. Conf. Commun. (NCC),
mmWave V2X in 5G and beyond to support automated driving,’’ 2020, Mar. 2017, pp. 1–6.
arXiv:2011.09590. [27] S. Dwivedi, P. Aggarwal, and A. K. Jagannatham, ‘‘Fast block LMS
[4] F. Liu and C. Masouros, ‘‘A tutorial on joint radar and communi- and RLS-based parameter estimation and two-dimensional imaging
cation transmission for vehicular networks—Part I: Background and in monostatic MIMO RADAR systems with multiple mobile tar-
fundamentals,’’ IEEE Commun. Lett., vol. 25, no. 2, pp. 322–326, gets,’’ IEEE Trans. Signal Process., vol. 66, no. 7, pp. 1775–1790,
Feb. 2021. Apr. 2018.
[5] K. V. Mishra, M. R. B. Shankar, V. Koivunen, B. Ottersten, and [28] A. Rawat, S. Dwivedi, S. Srivastava, and A. K. Jagannatham, ‘‘RLS-
S. A. Vorobyov, ‘‘Toward millimeter-wave joint radar communications: A based adaptive time-varying RCS estimation and imaging in MIMO radar
signal processing perspective,’’ IEEE Signal Process. Mag., vol. 36, no. 5, systems,’’ in Proc. Nat. Conf. Commun. (NCC), Feb. 2020, pp. 1–6.
pp. 100–114, Sep. 2019. [29] Y. C. Lin, T. S. Lee, Y. H. Pan, and K. H. Lin, ‘‘Low-complexity high-
resolution parameter estimation for automotive MIMO radars,’’ IEEE
[6] K. V. Mishra and Y. C. Eldar, ‘‘Sub-Nyquist channel estimation over IEEE
Access, vol. 8, pp. 16127–16138, 2020.
802.11ad link,’’ in Proc. Int. Conf. Sampling Theory Appl. (SampTA),
[30] A. B. Baral and M. Torlak, ‘‘Joint Doppler frequency and direction of
Jul. 2017, pp. 355–359.
arrival estimation for TDM MIMO automotive radars,’’ IEEE J. Sel. Topics
[7] J. Hatch, A. Topak, R. Schnabel, T. Zwick, R. Weigel, and C. Waldschmidt,
Signal Process., vol. 15, no. 4, pp. 980–995, Jun. 2021.
‘‘Millimeter-wave technology for automotive radar sensors in the 77 GHz
[31] S. Lim, J. Jung, J. Kim, S.-C. Kim, and J. Choi, ‘‘Enhanced velocity esti-
frequency band,’’ IEEE Trans. Microw. Theory Techn., vol. 60, no. 3,
mation based on joint Doppler frequency and range rate measurements,’’
pp. 845–860, Mar. 2012.
in Proc. 13th Int. Conf. Ubiquitous Future Netw. (ICUFN), Jul. 2022,
[8] M. Lesturgie, ‘‘T06—MIMO radar,’’ in Proc. IEEE Radar Conf., pp. 217–221.
May 2014, p. 33. [32] C. Wen, Y. Xie, Z. Qiao, L. Xu, and Y. Qian, ‘‘A tensor gen-
[9] J. Li and P. Stoica, ‘‘MIMO radar with colocated antennas,’’ IEEE Signal eralized weighted linear predictor for FDA-MIMO radar parameter
Process. Mag., vol. 24, no. 5, pp. 106–114, Sep. 2007. estimation,’’ IEEE Trans. Veh. Technol., vol. 71, no. 6, pp. 6059–6072,
[10] J. L. P. Stoica, MIMO Radar Signal Processing. Hoboken, NJ, USA: Wiley, Jun. 2022.
2009. [33] R. Tibshirani, ‘‘Regression shrinkage and selection via the lasso,’’ J. Roy.
[11] D. W. Bliss and K. W. Forsythe, ‘‘Multiple-input multiple-output (MIMO) Statist. Soc., B, Methodolog., vol. 58, no. 1, pp. 267–288, 1996.
radar and imaging: Degrees of freedom and resolution,’’ in Proc. IEEE [34] I. F. Gorodnitsky and B. D. Rao, ‘‘Sparse signal reconstruction from limited
Conf. Rec. 37th Asilomar Conf. Signals, Syst. Comput., vol. 1, Nov. 2003, data using FOCUSS: A re-weighted minimum norm algorithm,’’ IEEE
pp. 54–59. Trans. Signal Process., vol. 45, no. 3, pp. 600–616, Mar. 1997.
[12] E. Fishler, A. Haimovich, R. Blum, L. Cimini, D. Chizhik, and [35] P. Chen, Z. Cao, Z. Chen, and X. Wang, ‘‘Off-grid DOA estimation
R. Valenzuela, ‘‘Spatial diversity in radars—Models and detection per- using sparse Bayesian learning in MIMO radar with unknown mutual
formance,’’ IEEE Trans. Signal Process., vol. 54, no. 3, pp. 823–838, coupling,’’ IEEE Trans. Signal Process., vol. 67, no. 1, pp. 208–220,
Mar. 2006. Jan. 2019.
[13] I. Bekkerman and J. Tabrikian, ‘‘Target detection and localization using [36] T. Liu, F. Wen, L. Zhang, and K. Wang, ‘‘Off-grid DOA estimation for
MIMO radars and sonars,’’ IEEE Trans. Signal Process., vol. 54, no. 10, colocated MIMO radar via reduced-complexity sparse Bayesian learning,’’
pp. 3873–3883, Oct. 2006. IEEE Access, vol. 7, pp. 99907–99916, 2019.
[14] L. Xu and J. Li, ‘‘Iterative generalized-likelihood ratio test for MIMO [37] X. Zhao, C. Guo, and W. Peng, ‘‘Joint estimation of multiple parameters
radar,’’ IEEE Trans. Signal Process., vol. 55, no. 6, pp. 2375–2385, in bistatic MIMO radar based on sparse signal restoration,’’ in Proc. Int.
Jun. 2007. Conf. Microw. Millim. Wave Technol. (ICMMT), May 2018, pp. 1–3.
[15] S. Li, L. Zhang, N. Liu, J. Zhang, and S. Zhao, ‘‘Range-angle dependent [38] C.-Y. Wu, T. Zhang, J. Li, and T. F. Wong, ‘‘Parameter estimation in PMCW
detection for FDA-MIMO radar,’’ in Proc. CIE Int. Conf. Radar (RADAR), MIMO radar systems with few-bit quantized observations,’’ IEEE Trans.
Oct. 2016, pp. 1–4. Signal Process., vol. 70, pp. 810–821, 2022.
[16] X. Zhang and D. Xu, ‘‘Low-complexity ESPRIT-based DOA estimation [39] Z. Chen, W. Ma, P. Chen, and Z. Cao, ‘‘A robust sparse Bayesian learning-
for colocated MIMO radar using reduced-dimension transformation,’’ based DOA estimation method with phase calibration,’’ IEEE Access,
Electron. Lett., vol. 47, no. 4, pp. 283–284, 2011. vol. 8, pp. 141511–141522, 2020.
[17] M. Yang, B. Chen, G. Zheng, and X. Dang, ‘‘Reduced-dimensional [40] R. W. Heath, N. González-Prelcic, S. Rangan, W. Roh, and A. M. Sayeed,
ESPRIT algorithm for MIMO radar,’’ in Proc. IEEE CIE Int. Conf. Radar, ‘‘An overview of signal processing techniques for millimeter wave MIMO
Oct. 2011, pp. 347–349. systems,’’ IEEE J. Sel. Topics Signal Process., vol. 10, no. 3, pp. 436–453,
Feb. 2016.
[18] N. Liu, L. R. Zhang, J. Zhang, and D. Shen, ‘‘Direction finding of MIMO
[41] F. Sohrabi and W. Yu, ‘‘Hybrid digital and analog beamforming design for
radar through ESPRIT and Kalman filter,’’ Electron. Lett., vol. 45, no. 17,
large-scale antenna arrays,’’ IEEE J. Sel. Topics Signal Process., vol. 10,
pp. 908–910, Aug. 2009.
no. 3, pp. 501–513, Apr. 2016.
[19] J. Capon, ‘‘High-resolution frequency-wavenumber spectrum analysis,’’ [42] O. El Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, and R. W. Heath, Jr.,
Proc. IEEE, vol. 57, no. 8, pp. 1408–1418, Aug. 1969. ‘‘Spatially sparse precoding in millimeter wave MIMO systems,’’ IEEE
[20] E. G. Larsson, J. Li, and P. Stoica, ‘‘High-resolution nonparamet- Trans. Wireless Commun., vol. 13, no. 3, pp. 1499–1513, Jan. 2014.
ric spectral analysis: Theory and applications,’’ in High-Resolution [43] A. Alkhateeb, G. Leus, and R. W. Heath, ‘‘Limited feedback hybrid
Robust Signal Processing. Boca Raton, FL, USA: CRC Press, 2017, precoding for multi-user millimeter wave systems,’’ IEEE Trans. Wireless
pp. 151–252. Commun., vol. 14, no. 11, pp. 6481–6494, Nov. 2015.
[21] A. Jakobsson and P. Stoica, ‘‘Combining Capon and APES for estimation [44] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation
of spectral lines,’’ Circuits, Syst., Signal Process., vol. 19, no. 2, Theory. Upper Saddle River, NJ, USA: Prentice-Hall, 1993.
pp. 159–169, Mar. 2000. [45] K. Venugopal, A. Alkhateeb, N. González-Prelcic, and R. W. Heath,
[22] X. Zhang and D. Xu, ‘‘Angle estimation in MIMO radar using reduced- Jr., ‘‘Channel estimation for hybrid architecture-based wideband mil-
dimension Capon,’’ Electron. Lett., vol. 46, no. 12, pp. 860–861, Jun. 2010. limeter wave systems,’’ IEEE J. Sel. Areas Commun., vol. 35, no. 9,
[23] L. Xu, J. Li, and P. Stoica, ‘‘Target detection and parameter estimation for pp. 1996–2009, Sep. 2017.
MIMO radar systems,’’ IEEE Trans. Aerosp. Electron. Syst., vol. 44, no. 3, [46] D. P. Wipf and B. D. Rao, ‘‘Sparse Bayesian learning for basis selection,’’
pp. 927–939, Jul. 2008. IEEE Trans. Signal Process., vol. 52, no. 8, pp. 2153–2164, Aug. 2004.
[24] T. Yardibi, J. Li, and P. Stoica, ‘‘Nonparametric and sparse signal [47] M. Rossi, A. M. Haimovich, and Y. C. Eldar, ‘‘Spatial compressive sensing
representations in array processing via iterative adaptive approaches,’’ in in MIMO radar with random arrays,’’ in Proc. 46th Annu. Conf. Inf. Sci.
Proc. 42nd Asilomar Conf. Signals, Syst. Comput., Oct. 2008, pp. 278–282. Syst. (CISS), Mar. 2012, pp. 1–6.

VOLUME 10, 2022 132851

 M. Jafri et al.: Sparse Parameter Estimation and Imaging in mmWave MIMO Radar Systems

MEESAM JAFRI (Student Member, IEEE) SANA ANWER received the B.Tech. degree
received the B.E. degree in electronics and in electronics and communication engineering
communication engineering from Jamia Millia from the J. K. Institute of Applied Physics
Islamia, New Delhi, India, in 2016, and the and Technology, University of Allahabad, Praya-
M.Tech. degree in communication and infor- graj, India, in 2019, and the M.Tech. degree
mation technology from the National Institute in electrical engineering with specialization in
of Technology Srinagar, India, in 2018. He is signal processing, communication and networking
currently pursuing the Ph.D. degree with the from the Indian Institute of Technology Kanpur,
Department of Electrical Engineering, Indian Kanpur, India, in 2022. She is currently a Systems
Institute of Technology Kanpur, Kanpur, India. His Engineer with Qualcomm, Hyderabad, India. Her
research interests include coordinated beamforming in 5G wireless systems, research interests include joint sensing and communication systems, radar
mmWave communication, orthogonal time-frequency space (OTFS), radar signal processing, orthogonal time-frequency space (OTFS), and mmWave
signal processing, and joint radar and communication (RadCom). He was communications.
awarded Qualcomm Innovation Fellowship (QIF) from Qualcomm, in 2022.

ADITYA K. JAGANNATHAM (Senior Member,

SURAJ SRIVASTAVA (Member, IEEE) received IEEE) received the bachelor’s degree from the
the M.Tech. degree in electronics and commu- Indian Institute of Technology, Bombay, and the
nication engineering from the Indian Institute M.S. and Ph.D. degrees from the University of
of Technology Roorkee, India, in 2012, and the California at San Diego, San Diego, CA, USA.
Ph.D. degree in electrical engineering from the From April 2007 to May 2009, he was employed
Indian Institute of Technology Kanpur, Kanpur, as a Senior Wireless Systems Engineer with
India, in 2022. From July 2012 to November Qualcomm Inc., San Diego, where he was a part
2013, he was employed as a Staff-I Systems of the Qualcomm CDMA Technologies (QCT)
Design Engineer with Broadcom Research India Division. He is currently a Professor with the
Pvt. Ltd., Bengaluru, and from November 2013 to Department of Electrical Engineering, IIT Kanpur, where he also holds
December 2015, he was employed as a Lead Engineer with Samsung the Arun Kumar Chair Professorship. His research interests include next-
Research India, Bengaluru, where he worked on developing layer-2 of generation wireless cellular and WiFi networks, with a special emphasis on
the 3G UMTS/WCDMA/HSDPA modem. His research interests include various 5G technologies, and such as massive MIMO, mmWave MIMO,
applications of sparse signal processing in 5G wireless systems, mmWave FBMC, NOMA, and full duplex. He has been twice awarded the P. K.
and terahertz communication, orthogonal time-frequency space (OTFS), Kelkar Young Faculty Research Fellowship for excellence in research, the
joint radar and communication (RadCom), and optimization and machine Qualcomm Innovation Fellowship (QInF), the IIT Kanpur Excellence in
learning. He was awarded the Outstanding Ph.D. Thesis and Outstanding Teaching Award, the CAL(IT)2 Fellowship at the University of California
Teaching Assistant Awards from the IIT Kanpur. He was awarded Qualcomm at San Diego, and the Upendra Patel Achievement Award at Qualcomm.
Innovation Fellowship (QIF) from Qualcomm, in 2018 and 2022.

132852 VOLUME 10, 2022