You are on page 1of 19

Signal Processing 209 (2023) 109007

Contents lists available at ScienceDirect

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

Estimating the target DOA, range and velocity using subspace methods
in a MIMO OFDM DFRC system
Satwika Bhogavalli a,∗, K.V.S. Hari a, Eric Grivel b, Vincent Corretja c
a
Department of ECE, Indian Institute of Science, Bangalore, 560012, India
b
Bordeaux University-INP Bordeaux, IMS - UMR CNRS, 5218, France
c
Thales Defence Mission Systems, Campus Thales Merignac, France

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

Article history: Dual-function Radar communication (DFRC) systems implement Radar detection and communication us-
Received 24 September 2022 ing the same hardware simultaneously. Significant attention has been paid to a multiple-input multiple-
Revised 6 March 2023
output DFRC system based on orthogonal frequency division multiplexing (OFDM). So far, the directions of
Accepted 7 March 2023
arrival of the targets have been well estimated using subspace methods. In this paper, the ranges and/or
Available online 14 March 2023
velocities are estimated using the subspace methods. To this end, the Radar waveform must be based
Keywords: on the data symbols that are replicated over a few sub-carriers and/or during a few OFDM symbols. To
DFRC avoid data replication, the target ranges and velocities can be estimated based on a least-squares (LS) or
Subspace methods total LS method. After giving a few practical considerations, simulation results show that the proposed
Least squares approaches outperform the existing ones based on the Fourier transform and/or Lasso algorithm. The per-
DOA formance of the proposed approaches is compared with the Cramer–Rao bound. Moreover, we show how
Range
the performance of the DFRC system (in terms of the accuracy of the estimated target parameters and
Velocity
data rate) evolves when modifying the system parameters, such as the number of sub-carriers and the
number of OFDM symbols.
© 2023 Elsevier B.V. All rights reserved.

1. Introduction ing DFRC systems is to design a dual-function waveform in order


to detect targets and transmit useful information simultaneously.
A great deal of interest has been paid to joint Radar and com- This can be done by designing either a constant-modulus wave-
munication systems especially in air-borne and autonomous ve- form [5,6] or a beamforming where the precoding matrix1 is op-
hicular applications for the last years. The dual-function Radar- timized with different Radar sensing and communication metrics
communication (DFRC) systems have been developed to integrate including the data rate, the BER, etc. One can also use a communi-
both Radar and communication systems onto a single hardware cation waveform such as the orthogonal frequency division multi-
platform by sharing the spectrum efficiently [1]. Sensing and com- plexing (OFDM) [7–9] or embed information in a Radar waveform
munication are performed by using the same antenna array, the [10–14]. In [15,16], the advantages and the limitations of the above-
same frequency band, and a dual-function waveform [2–4]. There- mentioned strategies are presented. Finally, the index modulation
fore, the DFRC system is a complex system, the design of which can be another strategy of signal embedding in DFRC systems [17–
requires experts working in different domains such as electronics, 19]. In [20], the authors use a carrier agile phased array Radar
antennas and signal processing. In the latter domain, the main is- where the randomness in selecting the carrier frequencies and al-
sue is the estimation of the target parameters such as the direc- locating these frequencies to the antennas is exploited in the form
tion of arrival (DOA), the range, and the velocity in the presence of of index modulation to convey the information to the communica-
disturbances (namely, the additive thermal noise and/or the clut- tion systems.
ter), but also the retrieval of the information bits with the re- To improve the performance of both Radar and communication
quired bit error rate (BER) and the required spectral efficiency in functionalities, multiple-input multiple-output (MIMO) techniques
the communication system. Moreover, a critical challenge when us- have been adapted in the DFRC systems for the last few years

∗ 1
Corresponding author. The transmitted signal is pre-processed so that the system performance is im-
E-mail address: bhogavallis@iisc.ac.in (S. Bhogavalli). proved.

https://doi.org/10.1016/j.sigpro.2023.109007
0165-1684/© 2023 Elsevier B.V. All rights reserved.
S. Bhogavalli, K.V.S. Hari, E. Grivel et al. Signal Processing 209 (2023) 109007

[21,22]. When dealing with MIMO systems, three domains can be Then, both the range and the velocity are estimated again using
considered: the spatial one that is defined by the number of an- the FT method.
tennas (from a small number of antennas to a massive system), In the above mentioned papers, although the estimation of the
the configuration of the antennas (namely a uniform linear array target parameters by computing the FT over the antenna index
(ULA), a co-prime array, etc.) and the locations of the transmit and m, the sub-carrier index g or the OFDM symbol index u is a pri-
receive antennas (leading to the monostatic and bi-static cases). ori appealing due to the low computational cost, it has several
The second one, which is the frequency domain, is also of impor- drawbacks. Thus, when dealing with a single target, estimating
tance where an OFDM waveform is used. In that case, the selection the target parameters amounts to searching for the maximum of
of the private and shared sub-carriers have to be addressed.2 The the absolute value of a ratio of sinuses. As the number of anten-
time domain is the third domain to be considered, especially to nas, sub-carriers and OFDM symbols respectively denoted as M, G,
manage the range ambiguity, the target illumination, etc. and U are limited to small values in our case for various reasons
For the last few decades, the target DOA has been estimated (space and frequency optimization), zero-padding must be used.
not necessarily in a DFRC system using different methods. The The larger the zero-padding, the more accurate the estimation. In
reader can refer to pioneering works such as Jaffer [23] or re- return, the computational cost increases. In addition, if an additive
cent overviews [24] where machine learning-based algorithms are thermal noise is considered, its impact on the modulus of the FT is
cited. One of the most popular families of methods is based on all the more important as its power is larger than the power of the
the Fourier transform (FT) and includes algorithms using the dis- signal under study, leading to the degradation of the estimation ac-
crete Fourier transform (DFT), the inverse discrete Fourier trans- curacy. When there are K targets, with K strictly larger than 1, the
form (IDFT) [25], the fast Fourier transform (FFT) [26], the interpo- windowing effect, i.e. computing the FT on a finite sequence, is at
lated DFT [27] and the fractional Fourier transform [28]. In [26], the origin of several problems: inability to discriminate the DOA of
the DFT of the received data for one frequency bin is first ex- two targets when their difference becomes small and risk of inter-
pressed as a ratio of two polynomials. When different frequency preting the K largest local maxima as the values of interest.3
bins are considered, the expression of the DFT can be written in In addition, some aspects are not well-suited for practical pur-
a matrix form and depends on the parameters of the polynomials. poses in the above mentioned papers: 1. The attenuation of the re-
The latter are then estimated using a least-squares (LS) criterion ceived signal magnitude is not necessarily considered. For instance,
based on noisy observations at different time instants. Finally, the in Xu et al. [19], the received power from all targets is often as-
DOA is deduced. Capon Beamforming [29–31] and subspace meth- sumed to be the same whatever the target position may be. To
ods such as multiple signal classification (MUSIC) [32] or some solve that issue when modelling such a system in air-borne ap-
of its variants such as Root-MUSIC [33] as well as the estimation plications, one can introduce a multiplicative term which can be
of signal parameter via rotational invariance technique (ESPRIT) defined from the Radar range equation and the Swerling model
[34] can be used to deduce the DOA. To reduce the computational [45] characterizing the statistical behaviour of the Radar cross sec-
cost, many variants of subspace methods have been proposed [35– tion (RCS) of the target. 2. An element-wise division is often re-
37] without a significant loss in performance. Moreover, the per- quired [7–9,19]. From an implementation point of view, this step
formance analysis of different subspace methods is available in Rao may be quite expensive and hence should be avoided. Moreover,
and Hari [38,39,40], Delmas [41]. depending on the approach, the numerator and the denominator of
In this paper, a MIMO DFRC system based on OFDM is stud- the element-wise division operation need to be estimated. There-
ied. The transmitter of the MIMO OFDM DFRC system is shown in fore, the propagation of the resulting error can have an impact on
Fig. 1. We first focus on the Radar part, and more particularly on the performance of the estimation method. 3. Private sub-carriers
the estimation of the target parameters. In [7–9] the concepts that are often used for the estimation step [7–9,19], but this may lead
are used are more or less similar. In these papers, an OFDM wave- to a data rate loss.
form with private sub-carriers is considered. The received signal In this paper, our contributions are the following: 1. Unlike [7–
can be expressed as a sum of products of three complex exponen- 9,19], private sub-carriers are avoided. Moreover, we suggest using
tials. Their arguments depend on the DOA and the antenna index a more realistic model of the system: the received signal has its
m or on the range and the sub-carrier index g or on the velocity magnitude that is now random and defined by the Radar equa-
and the OFDM index u. Firstly, an element-wise division between tion and the Swerling model. Unlike [8] and [19], an additive zero-
the received data and the transmitted one, assumed to be known mean white Gaussian noise whose variance is unknown is consid-
at the receiver, is performed to get a sequence corresponding to ered. 2. Our purpose is to design a Radar waveform based on the
the product of two complex exponentials that depend on the range transmitted data symbol to estimate not only the DOA but also the
and the velocity. At that stage, the range and the velocity are es- range and/or the velocity of each target by using subspace meth-
timated using a FT-based method. The DOA of each target is esti- ods. Depending on the objectives the user has set (estimation of
mated by applying MUSIC on the received signal. the DOA and the range, estimation of the DOA and the velocity,
Unlike the above papers, in Xu et al. [19] and Xu and Petrop- estimation of the three target parameters), we will define which
ulu [42], the OFDM waveform is based on both private and shared assumptions must be made at the system level. More particularly,
sub-carriers. Specifically in Xu et al. [19], the target DOA is first we will see that the Radar waveform must be based on the trans-
estimated by means of a FT-based method. After an element-wise mission of data symbols that are replicated and transmitted over
division between the numerator (requiring the identification of the a few sub-carriers and/or during a few OFDM symbols. Depend-
main peaks of the FT of the received signal) and the denomina- ing on the performance levels that the user would like to get both
tor (defined from the DOA estimate), the range and the velocity in terms of target parameter estimation accuracy and data rate,
are deduced by applying FT methods again. It should be noted the practitioner will have to tune different parameters such as the
that there is no additive thermal noise in the system model in Xu number of sub-carriers or the number of OFDM symbols dedicated
et al. [19]. In addition, the authors suggest using the Lasso method
[43] to refine the DOA estimate by using a few private sub-carriers.
3
To avoid this problem, we present an operation mode addressing the estima-
tions of the DOAs separately. The waveform is designed so that the magnitudes of
2
The private sub-carriers correspond to some sub-carriers of a particular antenna the signals back-scattered by the targets, except the one located in a specific zone,
used for signal transmission whereas shared sub-carriers are sub-carriers of various are reduced. One zone, whose limits depend on the choice of the constellation and
antennas used for the same purpose. the number of antennas, is scanned at one sub-carrier [44].

2
S. Bhogavalli, K.V.S. Hari, E. Grivel et al. Signal Processing 209 (2023) 109007

Fig. 1. Transmitter structure of a MIMO DFRC system based on OFDM, showing private and shared sub-carrier transmission for illustration purpose.

to the design of the Radar waveform. 3. To avoid data replication Moreover, the system parameters and the resulting notations
that reduces the data rate at the communication level, the range are described in Table 1. Finally, when it is of interest, the sub-
and the velocity can be deduced by deriving a method based on scripts θ , D, and v will be used to relate the vectors or the matrices
LS or total-LS (TLS) criterion. 4. We make a comparison between with the DOA, the range and the velocity of the targets. The pa-
the proposed approaches and the one presented in Xu et al. [19] as rameters of the targets are represented as {θk }k=1,...,K , {Dk }k=1,...,K ,
well as the Cramer–Rao bound (CRB). The influence of the system and {vk }k=1,...,K .
parameters such as the signal-to-noise ratio (SNR), the number of
sub-carriers and the number of OFDM symbols dedicated to the 2. System model
Radar part is studied in terms of parameter estimation accuracy
and/or data rate. Let us consider a mono-static MIMO Radar in a DFRC system
The remainder of this paper is organized as follows: with a ULA consisting of L transmit antennas and M receive anten-
In Section 2, the MIMO OFDM Radar system is presented. nas spaced apart by a distance d. Although the transmit and re-
Section 3 presents the way to estimate the range and the ve- ceive antenna array is the same, we use different notations for the
locity using subspace methods while Section 4 is dedicated to number of transmit and receive antennas so that it can be easier
methods based on LS and TLS criteria. Section 5 presents the to extend the work to a bi-static Radar system in the future.
strategies for designing the Radar waveform in practical scenarios. As shown in Fig. 1, the transmit waveform is based on an OFDM
It also describes how these strategies affect the data rate at the scheme using G sub-carriers with a sub-carrier spacing  f . All the
communication level. Section 6 deals with the simulation results. sub-carriers are assumed to be shared. The data symbols are mod-
It includes a comparative study with the approaches proposed in ulated over all the sub-carriers by performing an IDFT. After ap-
Xu et al. [19] and with the CRB method. In addition, the influences pending it with a cyclic prefix (CP) and converting the samples to
of the signal-to-noise ratio and the system parameters such as the an analog signal, the signal is transmitted. This is referred as an
number of sub-carriers and OFDM symbols on the performance OFDM symbol. The L transmit antennas transmit U OFDM symbols,
of the DFRC system are studied. Section 7 concludes the paper each of duration T seconds. This transmission structure is referred
and provides some perspectives. An appendix dealing with the as a single snapshot for the Radar part. Multiple snapshots can be
estimation of the DOA using subspace methods and applied to the transmitted in a single pulse of a Radar waveform, as shown in
system under study is also proposed. Fig. 2.
Let sl (t ) be the base-band equivalent of the transmitted signal
from the l th transmit antenna, with l = 0, . . . , L − 1. It is given by:
Notations


U−1 G−1  t − uT 
Bold lower-case letters denote column vectors while upper-case sl (t ) = s(g, u, l ) exp ( j2π g f t ) rect (1)
bold letters denote matrices. Superscripts (. )∗ and (. )T respectively u=0 g=0
T
represent the complex conjugate and the transpose operators. 
represents the Hadamard product operator. Ia is the identity matrix where s(g, u, l ) is the data symbol of the uth OFDM symbol mod-
of size a. diag(a ) denotes the diagonal matrix whose main diagonal ulated over the gth sub-carrier denoted by the frequency fg and
is a. a ∼ exp(. ) indicates that the distribution of the random vari- transmitted by the l th antenna. All the data symbols constitute
able a is exponential. E(. ) is the expectation and rect(t ) stands for the information to be transmitted at the communication level, but
a rectangular pulse equal to 1 for − 21 ≤ t ≤ 12 and zero elsewhere. some of them will be also used to define the Radar waveform.

3
S. Bhogavalli, K.V.S. Hari, E. Grivel et al. Signal Processing 209 (2023) 109007

Table 1
System parameters and the corresponding notations.

Parameters Symbols Parameters Symbols

Number of transmit antennas L Velocity of the kth target vk


Index of the transmit antenna l = 0, . . . , L − 1 Doppler frequency of the kth target f kd
Number of receive antennas M Constellation order for modulation B
Index of the receive antenna m = 0, . . . , M − 1 Number of sub-carriers G
Distance between two consecutive d Index of the sub-carrier g = 0, . . . , G − 1
transmit/receive antennas
Number of targets K Number of OFDM symbols U
Index of the target k = 1, . . . , K Index of the OFDM symbol u = 0, . . . , U − 1
DOA of the kth target θk Duration of the OFDM symbol T
Range of the kth target Dk Duration of the cyclic prefix TCP

Fig. 2. Transmission of N snapshots.

In addition, fc is the carrier frequency. It is related with fg as results in:5


fg = fc + g f .   

K L−1

Let us assume that there are K targets in the free space char- ( fc + g f )
y(g, u, m ) = αk s(g, u, l ) exp − j2π ldsinθk
acterized by their DOAs, ranges and velocities. In the following, K c
k=1 l=0
is assumed to be unknown. Then, the reflected signal from the K  
targets is received by the M antennas. The ideal baseband signal ( fc + g f )
× exp − j2π mdsinθk
ym (t ) received by the mth receive antenna is given by:4 c
L−1
 2 Dk



K  × exp − j2π g f exp j2π uT fkd
ym (t ) = αk sl (t − τk ) exp ( j2π fkd t ) (2) c
k=1 l=0

K

where αk is the multiplicative constant defined from the Radar = k (g, u, m ) (6)
k=1
range equation [45] and the Swerling model 1 and is given by:
where k (g, u, m ) is the received signal back-scattered by the kth
λc σk1/2
αk = (3) target.
(4π )3/2 D2k Equation (6) can give a few insights on how the target param-
eters impact the received modulation symbols y(g, u, m ):
where λc is the ratio between c, the velocity of light and fc , the
carrier frequency. In addition, σk is the RCS of the kth target. Given • For a fixed sub-carrier and a fixed OFDM symbol, the DOA of
Swerling model 1, σk ∼ exp(σavg ) with σavg the statistical mean of the target introduces a progressive phase shift in the expression
all values of RCS of the target. Therefore, σk1/2 is a Rayleigh dis- of the received signal, y(g, u, m ) in the spatial domain, i.e. along
tributed random variable whose scale parameter is equal to √ 1 . m.
2σavg
In addition, τk appearing in (2) and affecting the transmitted
• The target range introduces a phase shift in the expression
signal is the round-trip delay of the kth target. It is defined by: of the received signal, y(g, u, m ) in the frequency domain, i.e.
along g.
Dk • The Doppler frequency or the velocity of the target introduces a
τk ≈ 2 + (l + m )d sin θk (4)
c progressive phase shift in the expression of the received signal
y(g, u, m ) in the time domain, i.e., along u.
According to (4), both the DOA θk and the range Dk have an influ-
ence on the delay. It should be noted that the Doppler frequency However, in practice, the received data is disturbed by an addi-
in (2) does not vary much with the sub-carrier index and therefore tive noise. The data model can hence be expressed as:
can be approximated as:

K
vk f c y(g, u, m ) = k (g, u, m ) + η (g, u, m ) (7)
fkd ≈ 2 (5)
c k=1

Applying the DFT to the ideal received signal after converting the where {η (g, u, m )}g=0,...G−1; u=0,...,U−1; m=0,...,M−1 is assumed to be
signal from the analog to the discrete domain and removing the CP a set of independent and identically distributed (i.i.d.) zero-mean
Gaussian random variables with variance ση2 .
4
In theory, the signal back-scattered by each target at each frequency f c + g f
c σ 1/2
should be multiplied by αk,g = (4π )3/2 ( fk +g f )D2 . However, in this paper, as f c + g f 5
Although the term IDFT is used, the division by G is omitted for the sake of
c k
can be approximated by f c , one can consider that αk,0 ≈ αk,1 . . . ≈ αk,G−1 = αk . simplicity.

4
S. Bhogavalli, K.V.S. Hari, E. Grivel et al. Signal Processing 209 (2023) 109007

Given the noisy received signal, the target parameters such as The vector s(g, u, L ) is hence the same for g = Gmin , . . . , Gmin + Gc −
the DOA, the range and the velocity have to be estimated. Let us 1. As it no longer depends on the sub-carrier index, we suggest
analyze the way to do it. using the following notation: s(u, L ).
With both assumptions, y(u, m, Gc ) can be approximated as:
3. Estimation of target parameters using the subspace methods T
y(u, m, Gc ) = y(Gmin , u, m ) ... y(Gmin + Gc − 1, u, m )
In this section, we present the way to estimate all the target (14)
parameters using the subspace methods applied on the noisy re-
ceived signal. These methods are commonly used to estimate the
≈ AD (Gc )(u, m )ATθ s(u, L ) + η(u, m, Gc ) (15)
DOAs (See Appendix A), but it is not the case for the range and
the velocity. Therefore, we will show which assumptions must be with the corresponding additive-noise vector at the receiver given
made at the system level to be able to apply subspace methods to by:
estimate all the target parameters. They either deal with mathe- T
matical approximations or constraints on the data symbols defin- η(g, u, Gc ) = η (g, u, 0 ) η (g, u, 1 ) ... η (g, u, Gc − 1 )
ing the Radar waveform, the number of sub-carriers and the num-
(16)
ber of OFDM symbols dedicated to the Radar system. Note that K
the number of targets is assumed to be known in those subsec- and the matrix AD (Gc ) defined as follows:
tions. We will see in Section 5 how to combine all these assump-

T
AD (Gc ) = aD Gmin ... aD Gmin + Gc − 1
tions in practice.    
⎡ 2D 2DK ⎤
In the following, let us introduce a notation for the vector exp − j2π Gmin  f c 1 ... exp − j2π Gmin  f c
where the data can be stacked along a “domain”. For example, the ⎢ ⎥
⎢ . . ⎥
=⎢ . . ⎥
data received over the gth sub-carrier during the uth OFDM symbol ⎣  .   . ⎦

2D
2DK
from the M antennas can be stacked in the following vector: exp − j2π Gmin + Gc − 1  f c 1 ... exp − j2π Gmin + Gc − 1  f c
T
y(g, u, M ) = y(g, u, 0 ) ... y(g, u, M − 1 ) (8) (17)

This column vector depends on g and u. Its size is M. The correlation matrix of y(u, m, Gc ) of size Gc × Gc satisfies:

3.1. Estimation of the range of the targets using subspace methods RD (u, m, Gc ) = E[y(u, m, Gc )yH (u, m, Gc )] (18)
As the noise vector η(u, m, Gc ) is i.i.d. and of zero-mean, the cor-
To estimate the range, Eq. (7) can be rewritten as follows: relation matrix RD (u, m, Gc ) can be approximated as follows:
y(g, u, m ) = aTD (g)(g, u, m )ATθ (g)s(g, u, L ) + η (g, u, m ) (9)
D (Gc ) + ση IGc
RD (u, m, Gc ) ≈ AD (Gc )(u, m )ATθ Rs (u, L )A∗θ H (u, m )AH 2

where aD (g) is given by: (19)


T
aD (g) = exp(− j2π g f 2D1
c
) ... exp(− j2π g f 2DK
c
) (10) where Rs (u, L ) = E[s(u, L )sH (u, L )]. As the transmitted symbols cor-
Let us define respond to a sequence of i.i.d. zero-mean random variables with
fc + g f variance σs2 , Rs (u, L ) is the diagonal matrix σs2 IL independent of g
ωk (g) = 2π d sin θk (11) and u.
c
Therefore, (19) becomes:
Given the expression of ωk (g), the matrix (g, u, m ) of size K × K
is defined by: RD (u, m, Gc ) ≈ σs2 AD (Gc )(u, m )ATθ A∗θ H (u, m )AH
D ( G c ) + ση I G c
2

(g, u, m ) = diag α1 exp j2π exp (−jmω1 (g) ) · · · αK


uT f1d (20)


× exp j2π uT fK exp (−jmωK (g) )
d
(12) Let us look at the right hand side of the Eq. (20): (u, m ) is a di-
agonal matrix and the columns of both AD (Gc ) and Aθ are linearly
In the following, let us consider the vector y(u, m, Gc ) stacking
independent. Consequently, the matrices (u, m ), AD (Gc ) and Aθ
the data y(g, u, m ) over any consecutive Gc sub-carriers, where
have full rank equal to K. The signal subspace associated with the
Gc ∈ [K + 1, G]. Note that the subscript c refers to “consecutive”.
correlation matrix equal to σs2 AD (Gc )(u, m )ATθ A∗θ H (u, m )AH
D ( Gc )
Thus, the index g of the sub-carrier varies between Gmin and
is of dimension K. Therefore, the subspace methods like MUSIC,
Gmin + Gc − 1 where Gmin ∈ {0, 1, . . . , G − Gc }. The remaining G − Gc
Root-MUSIC and ESPRIT are applicable to estimate the ranges of
sub-carriers are used for the communication part. Moreover, the
the K targets. We propose to apply them on the correlation matrix
index u will refer to the index of the OFDM symbols within the
RD (u, m, Gc ).
subset U1 , which consists of U1 OFDM symbols that are not neces-
Similarly to the process used to estimate the DOA, the corre-
sarily consecutive among the U OFDM symbols available.
lation matrices RD (u, m, Gc ) can be averaged over the U1 OFDM
Our motivation is to write the vector y(u, m, Gc ) in a matrix
symbols and the M antennas. The subspace methods can then be
form well-suited to apply subspace methods. To this end, let us
applied to the averaged correlation matrix given by:
look at the two assumptions one has to make to estimate the range
1 
with subspace methods. M−1

Assumption 1 to use subspace methods: Let us approximate fg by RD ( G c ) = RD (u, m, Gc ) (21)


U1 M
fc ∀g = 0, 1, . . . , G − 1. This approximation is valid as g f << fc u∈U1 m=0

for any g. In that case, (g, u, m ) and ATθ (g) no longer depend on g. From the correlation matrix RD (Gc ), the target range is estimated
Assumption 2 to use subspace methods: Given the l th transmit an- using subspace methods such as MUSIC, Root-MUSIC and ESPRIT.
tenna, the data symbols transmitted during the uth OFDM symbol Gc , Gmin and U1 are the three parameters that can be tuned by
are assumed to be the same for g = Gmin , . . . , Gmin + Gc − 1. Math- the practitioner to get the set of K range estimates.
ematically, this means for l = 0, 1, . . . , L − 1: The user will have to take into account the following points:

s(Gmin , u, l ) = s(Gmin + 1, u, l ) . . . = s(Gmin + Gc − 1, u, l ) (13) • Gc must be chosen larger than the number of targets.

5
S. Bhogavalli, K.V.S. Hari, E. Grivel et al. Signal Processing 209 (2023) 109007

• The accuracy of the range estimation should a priori be im- As the noise vector, η(g, m, Uc ) is i.i.d. and of zero-mean, the cor-
proved by increasing Gc . relation matrix Rv (g, m, Uc ) becomes:
• As G − Gc sub-carriers are dedicated to the communication sys-
tem, selecting a large value for Gc should reduce the data rate. Rv (g, m, Uc ) = σs2 Av (Uc )χ (g, m )ATθ A∗θ χH (g, m )AH
v (Uc ) + ση IUc
2

In Sections 5 and 6, we will see the influence of the parameters (30)


such as Gc and U1 on the target parameter estimation accuracy and since Rs (g, L ) = E[s(g, L )sH (g, L )] = σs2 IL .
the data rate in more detail. In the next subsection, let us address In (30), χ(g, m ) is a diagonal matrix and the columns of both
the estimation of the velocity. Av (Uc ) and Aθ are linearly independent. The matrices χ(g, m ),
Av (Uc ) and Aθ have consequently a full rank, equal to K. The
3.2. Estimation of the velocity of the targets using subspace methods signal subspace associated with the correlation matrix equal to
σs2 Av (Uc )χ(g, m )ATθ A∗θ χH (g, m )AHv (Uc ) is of dimension K. Therefore,
To estimate the target velocity, the Eq. (7) is written as: the subspace methods like MUSIC, Root-MUSIC and ESPRIT are ap-
y(g, u, m ) = aTv (u )χ (g, m )ATθ s(g, u, L ) + η (g, u, m ) (22) plicable to estimate the velocity of K targets from the full-rank cor-
relation matrix Rv (g, m, Uc ).
where av (u ) is given by:
Following the same reasoning as above, the correlation matri-
T ces Rv (g, m, Uc ) can be averaged over the G1 sub-carriers and the
av (u ) = exp( j2π uT f1d ) ... exp( j2π uT fKd ) (23)
M antennas. Then the subspace methods can be applied to the av-
Given ωk whose expression is given in (11), the matrix χ(g, m ) of eraged correlation matrix given by:
size K × K is defined as:
   1 
M−1
2D
χ(g, m ) = diag α1 exp − j2π g f 1 exp (−jmω1 (g) ) · · · αK Rv (Uc ) = Rv (g, m, Uc ) (31)
c G1 M g∈G
m=0
 2DK
  1

exp − j2π g f exp (−jmωK (g) ) (24) From the correlation matrix Rv (Uc ), the target velocity is esti-
c
mated using subspace methods such as MUSIC, Root-MUSIC and
Now, one aims at writing the vector y(g, m, Uc ) in a matrix form. ESPRIT.
It is the starting point to apply subspace methods. This vector As one can see, there are three degrees of freedom to get the
stacks the data y(g, u, m ) during Uc consecutive OFDM symbols, set of K velocity estimates: Uc , Umin and G1 .
where Uc ∈ [K + 1, U]. Thus, in the following, the index u of the When tuning these parameters, the practitioner can consider
OFDM symbol varies between Umin and Umin + Uc − 1 where Umin ∈ the following points:
{0, 1, . . . , U − Uc }. The remaining U − Uc OFDM symbols are used
for the communication part. The index g will refer to the index • The value for Uc must be chosen larger than the number of the
of the sub-carriers within the subset G1 that consists of G1 sub- targets.
carriers that are not necessarily consecutive among the G sub- • One may expect that the accuracy of the velocity estimation
carriers available. should be improved with the size of the correlation matrix
At this stage, let us look at the assumption one could make to Rv (Uc ). Therefore, the higher Uc is, the better the accuracy
use subspace methods to estimate the velocity. should be.
Assumption to use subspace methods: Given the l th transmit an- • As U − Uc OFDM symbols are dedicated to the communication
tenna, the data symbols transmitted over the sub-carrier frequency system, increasing Uc should reduce the data rate a priori.
fg are assumed to be the same during Uc consecutive OFDM sym-
In Sections 5 and 6, we will examine the influence of the pa-
bols i.e., from Umin , . . . , Umin + Uc − 1. Mathematically, this means
rameters such as Uc and G1 on both target parameter estimation
for l = 0, 1, . . . , L − 1:
accuracy and the data rate.
s(g, Umin , l ) = s(g, Umin + 1, l ) . . . = s(g, Umin + Uc − 1, l ) (25)
With the above assumption, the vector s(g, u, L ) is the same for 3.3. Using multiple snapshots for subspace methods
u = Umin , . . . , Umin + Uc − 1. As it no longer depends on the OFDM
symbol index, we suggest using the following notation: s(g, L ). For subspace methods, the estimation of the correlation matrix
Therefore, y(g, m, Uc ) is given as: is a key step to estimate the target parameters accurately. In (A.12),
T (21) and (31) we already suggested averaging the correlation ma-
y(g, m, Uc ) = y(g, Umin , m ) ... y(g, Umin + Uc − 1, m ) trices over the sub-carrier index, the OFDM symbol index or the
= Av (Uc )χ (g, m )ATθ s(g, L ) + η(g, m, Uc ) (26) antenna index. Moreover, multiple snapshots can be transmitted. In
that case, the target parameters are assumed to remain the same
during the N transmitted snapshots. The correlation matrices for
= Av (Uc )χ (g, m )ATθ s(g, L ) + η(g, m, Uc ) (27)
each snapshot are averaged before the estimation of the target pa-
where the matrix Av (Uc ) is defined as: rameters using subspace methods.
T From Sections 3.1 and 3.2, the data need to be replicated over a
Av (Uc ) = av (Umin ) ... av (Umin + Uc − 1 ) few sub-carriers and during some OFDM symbols in order to apply



exp j2π UminTf 1 d
... exp j2π UminTf dK the subspace methods. This results in a loss in the data rate. To
⎢ . . ⎥ avoid data replication, we present another method to estimate the
=⎣ . . ⎦
.
.
target range and the target velocity using LS or TLS criterion in the
exp j2π (Umin + Uc − 1 )T f1d ... exp j2π (Umin + Uc − 1 )T fKd next section.
(28)
4. Estimation of the target parameters using LS/TLS
In addition, the correlation matrix of y(g, m, Uc ) of size Uc × Uc is
given as:
First, the estimation of the target range is addressed. Then the
Rv (g, m, Uc ) = E[y(g, m, Uc )yH (g, m, Uc )] (29) estimation of the target velocity is presented.

6
S. Bhogavalli, K.V.S. Hari, E. Grivel et al. Signal Processing 209 (2023) 109007

4.1. Estimation of the range using LS/TLS For each target, G estimates of the range are hence available. For
this reason, they are averaged:
By introducing bk (g, u, m ) defined as follows: G−1
1
L−1
  Dˆ k = Dˆ k (g) for k = 1, 2, . . . , K (40)
 ( fc + g f ) G
bk (g, u, m ) = s(g, u, l ) exp − j2π (md + ld ) sin θk g=1
c
l=0 This LS method does not take into account the fact that
(32) ˆ (g, 0 )aD,α (g) was an approximation of y(g, 0 ) since the DOA is
B
estimated. For this reason, a TLS method is considered now.
the noisy received signal can be rewritten this way:


K  2D

4.1.2. Total least-squares method to estimate the range
y(g, u, m ) = αk bk (g, u, m ) exp − j2π g f k exp j2π uT fkd For g = 1, . . . , G − 1, Eq. (35) can be rewritten as follows:
c
k=1

+ η (g, u, m ) (33) y(g, 0 ) − η(g, 0 ) = B


ˆ (g, 0 ) + B
˜ (g, 0 ) aD,α (g) (41)

where B ˜ (g, 0 ) = B(g, 0 ) − B


ˆ (g, 0 ). Our purpose is to estimate
To make the estimation of the range independent of the target
aD,α ( g) that minimizes the Frobenius norm of the augmented ma-
velocity, we propose to consider u = 0 since exp( j2π uT fkd ) reduces
trix B ˜ (g, 0 ) −η(g, 0 ) . To this end, let us rewrite (41) as follows:
to 1. One obtains:
 

K   aD,α (g)
y(g, 0, m ) =
2D
αk bk (g, 0, m ) exp − j2π g f k + η (g, 0, m )
ˆ (g, 0 ) + B
B ˜ (g, 0 ) y(g, 0 ) − η(g, 0 ) = 0M×1 (42)
−1
c
k=1
(34) The goal is to find B ˜ (g, 0 ) −η(g, 0 ) that reduces the rank

of B ˆ (g, 0 ) y(g, 0 ) by 1. To address this issue, the SVD of
The information of interest, namely the range Dk , lies in the ar-
ˆ (g, 0 ) y(g, 0 ) can be considered and is given by:
B
gument of the complex exponentials that depends on g. Conse-
quently, only the sub-carrier index g equal to 1, . . . , G − 1 will be
ˆ (g, 0 )
B y(g, 0 ) = PQ (43)
considered in the following.
Given the gth sub-carrier and using aD (g) defined in (10), stack- where P and Q are unitary matrices and  is a matrix whose
ing y(g, 0, m ) for all values of m i.e., m = 0, 1, . . . , M − 1 results in: size is the same as the ones of the matrix B ˆ (g, 0 ) y(g, 0 ) and
whose diagonal
elements are the singular values of the augmented
y(g, 0 ) = B(g, 0 )aD,α (g) + η(g, 0 ) (35) matrix B ˆ (g, 0 ) y(g, 0 ) arranged in the descending order. From
[46], the vector estimate can be expressed this way:
where: ⎡ ⎤
⎡ ⎤ q1,K+1
b1 (g, 0, 0 ) ... bK (g, 0, 0 ) ⎢q2,K+1 ⎥
1 ⎢ . ⎥ provided qK +1,K +1 = 0
B(g, 0 ) = ⎣ .. .. ⎦ aˆ D,α (g) = − (44)
. . (36) qK +1,K +1 ⎣ .. ⎦
b1 (g, 0, M − 1 ) ... bK (g, 0, M − 1 ) qK,K+1
Re-calling aD (g) from equation (10), aD,α (g) is given as: where {qb,K+1 }K+1 are elements of the (K + 1 )th singular vector
T b=0
in the matrix Q. The estimated ranges of the targets denoted as
aD,α (g) = α1 α2 ... αK  aD ( g ) (37)
{Dˆ k }Kk=1 can be obtained from Eqs. (39), (40) in the Section 4.1.1.
Given the matrix formulation (34)–(36), our goal is to estimate In both LS and TLS methods, as the range takes only positive
aD,α (g) which consists of all the target ranges. As the true matrix values, ∠aˆDk (g) should be between [0, 2π ) to avoid any ambigu-
B(g, 0 ) is unknown, it is replaced by its estimate Bˆ (g, 0 ) obtained ity of the phase values. For g = 1, . . . , G − 1, this condition puts a
with the estimates of the DOAs. First of all, a LS criterion is used. restriction on the product of the sub-carrier index and the target
Then, as an error is introduced by the estimation of the DOA, a TLS range as follows:
criterion is considered. ∠aˆDk (g) gDˆ k (g) f 1
0≤ = < (45)
4π c 2
4.1.1. Least-squares method to estimate the range
or equivalently:
We propose to search the vector aD,α (g) of the target ranges
that minimizes the following criterion: ||y(g, 0 ) − B
ˆ (g, 0 )aD,α (g)||2 . c
0 ≤ Dˆ k (g) < (46)
For g = 1, . . . , G − 1, this leads to: 2g f

aD,α (g) = (B
ˆ H (g, 0 )B
ˆ (g, 0 ))−1 B
ˆ H (g, 0 )y(g, 0 ) (38) One can notice that the maximum distance that can be estimated
c
depends on the sub-carrier index. When g = 1, it is equal to 2 f
;
Therefore, the estimation of aD,α (g) can be expressed from the c
if g = 2, it is equal to 4 and so on. Consequently, if all the sub-
f
pseudo-inverse of the matrix B ˆ (g, 0 ). It corresponds to the orthog- c
carriers contribute to the range estimation, 2(G−1 ) f is the largest
onal projection of y(g, 0 ) onto the space spanned by the columns
distance that can be estimated a priori. Therefore, the way to esti-
of B ˆ (g, 0 ). Instead of using the pseudo inverse, an expression based
mate the range is dependent on the sub-carriers that are chosen.
on the singular value decomposition (SVD) of B ˆ (g, 0 ) can be ob-
In the next section, the estimation of the last target parameter,
tained.
namely the velocity, is addressed.
Since {αk }Kk=1 is a real quantity, Dˆ k the estimated range of the
kth target, can be deduced from the phase of the kth element of
4.2. Estimation of the velocity using LS/TLS
aˆ D,α (g), i.e. ∠aˆDk (g), for k = 1, . . . , K and for g = 1, . . . , G − 1:

c∠aˆDk (g) Let us start from Eqs. (32), (33). To estimate the velocity inde-
Dˆ k (g) = − (39) pendently of the target range, we propose to consider g = 0 since
4π g f

7
S. Bhogavalli, K.V.S. Hari, E. Grivel et al. Signal Processing 209 (2023) 109007

2Dk
exp(− j2π g f c ) reduces to 1 in (33). One obtains: 4.4. Summary


K
On the one hand, the estimation of the range and the veloc-
y(0, u, m ) = αk bk (0, u, m ) exp( j2π uT fkd ) + η (0, u, m ) (47)
ity using LS/TLS criterion a priori requires the target DOA esti-
k=1
mates. Given (34)–(37) and (47)–(50), the range and the velocity
In (47), as the arguments of the complex exponentials depend on are inherently associated with the DOA. On the other hand, the
the velocity vk but also on u, all the OFDM symbol indices except estimations of the range and the velocity using subspace meth-
the one equal to 0 will be used to estimate the velocity. Given ods do not require the DOA estimates. Hence, the accuracy of the
the uth OFDM symbol and recalling av (u ) defined in (23), stack- range and the velocity estimates do not depend on the accuracy
ing y(0, u, m ) for all values of m i.e., m = 0, 1, . . . , M − 1, results in of the target DOA estimates, but an a posteriori association of the
the following equality: target parameters must be done. Let us look at the case when
y(0, u ) = B(0, u )av,α (u ) + η(0, u ) (48) the three parameters are estimated. Given the sets of estimated
DOAs, ranges and velocities, the possible number of associations
where:
⎡ ⎤ of the target parameters is equal to K 3 . One solution is the fol-
b1 (0, u, 0 ) ... bK (0, u, 0 ) lowing: For a given estimated DOA and a given estimated range
⎣ .. .. ⎦ T
B ( 0, u ) = . . (49) and by referring to (35), the vector α1 α2 . . . αK can be
b1 (0, u, M − 1 ) ... bK (0, u, M − 1 ) estimatedin the LS sense from the vector y(g, 0 ) and the matrix

and B(g, 0 )  aˆ D (g) × 1 ... 1 , where aˆ D (g) is the vector sim-
T ilarly defined as in (10), but the range is replaced by the esti-
av,α (u ) = α1 α2 ... αK  av ( u ) (50)
mated one. Then, for each velocity estimate available and based
Given this matrix formulation (47)–(49), our purpose is to esti- on Eqs. (48)–(50), the difference between the 2-norm of the vector
mate av,α (u ) which consists of all the target velocities. As the true y ( 0, u ) − B
ˆ (0, u )aˆ v,α (u ) and the estimation of the variance of the
matrix B(0, u ) is unknown, we propose to replace it by its esti- additive noise is computed. This criterion is then used to finalize
mate B ˆ (0, u ) obtained with the estimates of the DOAs. Similarly to the associations.
the procedure suggested in Section 4.1.1, the vector av,α (u ) is esti- Table 2 summarizes the assumptions one has to make and dif-
mated in the LS sense and leads to aˆ v,α (u ). Then, vˆ k the velocity of ferent comments on each proposed method depending on the ob-
the kth target, can be estimated from the phase of the kth element jectives set by the practitioner. The name of the approach is de-
in aˆ v,α (u ), i.e. ∠aˆvk (u ), for k = 1, . . . , K and for u = 1, . . . , U − 1: fined by the acronym of the methods used to estimate the DOA,
c∠aˆvk (u ) the range and/or the velocity. When subspace methods are used
vˆ k (u ) = (51) for the three parameters, the approach is labelled SS-SS-SS. When
2π uT fc
subspace methods are used for DOA and LS (or TLS) is used for
For each target, U estimates of the velocity are available. For this range, the approach is labelled as SS-LS (or SS-TLS).
reason, they are averaged over all the U OFDM symbols and one
obtains:
5. Data rate of the communication part of the DFRC system
1 
U−1
vˆ k = vˆ k (u ) for k = 1, 2, . . . , K (52)
U In Appendix A and in Sections 3 and 4, we focused our atten-
u=1
tion on the Radar part of the DFRC system. More particularly, the
As the LS method does not take into account the fact that B(0, u )
estimations of the three target parameters were addressed inde-
was an approximation (as the DOA estimates are involved in it),
pendently from a theoretical point of view. Depending on the ob-
the TLS method can be considered. After estimating the vector
jective of the practitioner (estimation of the DOA and the range,
aˆ v,α (u ) using TLS as suggested in Section 4.1.2, the target velocity
estimation of the DOA and the velocity or estimation of the three
is estimated from (51), (52).
target parameters) and the methods used to estimate the target
In both LS and TLS estimates, as the velocity can take both pos-
parameters (subspace methods or TLS/LS method), the design of
itive and negative values, ∠aˆvk (u ) should be between [−π , π ) to
the Radar waveform is different. It has hence an incidence on the
avoid any ambiguity of the phase values. This condition puts a re-
data rate of the communication system. As many scenarios can be
striction on the product of the OFDM symbol index and the target
derived, we propose to study two of them in this section. The first
velocity as follows:
one deals with subspace methods to estimate the three target pa-
−1 ∠aˆvk (u ) uT fc vˆ k (u ) 1 rameters while the second one is based on the approach using sub-
≤ = < (53)
2 2π c 2 space methods for the DOA and TLS methods for range and veloc-
Following the same reasoning as stated at the end of the ity.
Section 4.1.2, from (53) the detectable velocity depends on the
OFDM symbols chosen. 5.1. When subspace methods are used to estimate all the target
parameters (SS-SS-SS)
4.3. Using multiple snapshots for methods based on LS/TLS criterion.
5.1.1. Design of the Radar waveform in that scenario
The method based on LS/TLS criterion is suitable even if there
As mentioned in Sections 3.1 and 3.2, the waveform is designed
is one snapshot. For multiple snapshots, the transmitted data sym-
so that the same data is transmitted over Gc sub-carriers during U1
bols are the same. Therefore, given the expression (7) of the re-
OFDM symbols to estimate the range and G1 sub-carriers during
ceived signal, the only difference from one snapshot to another is
Uc OFDM symbols to estimate the velocity. Let us introduce Gi and
the additive noise. To exploit all the snapshots, we propose to av-
Ui , the number of sub-carriers and OFDM symbols that are used
erage the data received from all the snapshots. In that case, the
for both the estimation of the range and the velocity.6 To design a
additive-noise variance is divided by the number of snapshots.
waveform, let us now select the values of the system parameters,
Remark: For the methods based on FT and Lasso, if there are
multiple snapshots, the received data symbols can be averaged
over the snapshots. 6
The subscript i stands for “intersection”.

8
S. Bhogavalli, K.V.S. Hari, E. Grivel et al. Signal Processing 209 (2023) 109007

Table 2
The assumptions and degrees of freedom for the parameters to be estimated depending on the method used for estimation.

Parameters to Methods Assumptions/approximations for the method to be Degrees of Comments


be estimated used applicable freedom

DOA SS No M, G, U 1. M > K.
2. In scheme 2 (See Appendix A), G is chosen so that
f c + g f ≈ f c ∀g ∈ {0, . . . , G − 1}
DOA and SS-SS 1. The same data should be transmitted over G, U 1. M > K.
range Gc consecutive sub-carriers during U1 (≥ 1 ) OFDM Gc , U1 , M 2. In scheme 2 (See Appendix A), G is chosen so that
symbols. f c + g f ≈ f c ∀g ∈ {0, . . . , G − 1}
2. f c + g f ≈ f c ∀g ∈ {0, . . . , G − 1} 3. G ≥ Gc > K
DOA and SS-LS No G, U 1. M > K.
range SS-TLS 2. In scheme 2 (See Appendix A), G is chosen so that
f c + g f ≈ f c ∀g ∈ {0, . . . , G − 1}
3. The maximum range that can be estimated can be
deduced from:
0 ≤ gDk (cg) f < 12
ˆ

DOA and SS-SS The same data should be transmitted during G, U 1. M > K.
velocity Uc consecutive OFDM symbols over G1 (≥ 1 ) Uc , G1 , M 2. In scheme 2 (See Appendix A), G is chosen so that
sub-carriers. f c + g f ≈ f c ∀g ∈ {0, . . . , G − 1}
3. U ≥ Uc ≥ K
DOA and SS-LS No G, U 1. M > K.
velocity SS-TLS 2. In scheme 2 (See Appendix A), G is chosen so that
f c + g f ≈ f c ∀g ∈ {0, . . . , G − 1}
3. The maximum velocity that can be estimated can be
deduced from:
−1
2
≤ uT fc vcˆ k (u ) < 12
DOA, range SS-SS-SS 1. The same data should be transmitted over G, U 1. M > K.
and velocity Gc consecutive sub-carriers during U1 (≥ 1 ) OFDM Gc , U1 2. In scheme 2 (See Appendix A), G is chosen so that
symbols. Uc , G1 , M f c + g f ≈ f c ∀g ∈ {0, . . . , G − 1}
2. The same data should be transmitted during 3. G ≥ Gc > K
Uc consecutive OFDM symbols over G1 (≥ 1 ) 4. U ≥ Uc > K
sub-carriers.
3. f c + g f ≈ f c ∀g ∈ {0, . . . , G − 1}
DOA, range SS-SS-LS 1. The same data should be transmitted over G, U 1. M > K.
and velocity SS-SS-TLS Gc consecutive sub-carriers during U1 (≥ 1 ) OFDM Gc , U1 , M 2. In scheme 2 (See Appendix A), G is chosen so that
symbols. f c + g f ≈ f c ∀g ∈ {0, . . . , G − 1}
2. f c + g f ≈ f c ∀g ∈ {0, . . . , G − 1} 3. G ≥ Gc > K
4. The maximum velocity that can be estimated can be
deduced from:
−1
2
≤ uT fc cvˆ k (u ) < 12
DOA, range SS-LS-SS The same data should be transmitted during G, U 1. M > K.
and velocity SS-TLS-SS Uc consecutive OFDM symbols over G1 (≥ 1 ) Uc , G1 , M 2. In scheme 2 (See Appendix A), G is chosen so that
sub-carriers. f c + g f ≈ f c ∀g ∈ {0, . . . , G − 1}
3. The maximum range that can be estimated can be
deduced from:
0 ≤ gDk (cg) f < 12
ˆ

4. U ≥ Uc > K
DOA, range SS-L S-L S No G, U 1. M > K.
and velocity SS-TLS- 2. In scheme 2 (See Appendix A), G is chosen so that
TLS f c + g f ≈ f c ∀g ∈ {0, . . . , G − 1}
3. The maximum range that can be estimated can be
deduced from:
0 ≤ gDk (cg) f < 12
ˆ

4. The maximum velocity that can be estimated can be


deduced from:
−1
2
≤ uT fc cvˆ k (u ) < 12

namely Gc , Uc , G1 , U1 , Gi and Ui , by considering a few practical as- U − 1 respectively. In practice, if M the number of receive anten-
pects. nas is assumed to be smaller than G and U, the maximum number
When the ranges and velocities of the targets are estimated of targets that can be characterized by the DOA, the range and the
with subspace methods, Gc and Uc should be larger than the num- velocity is limited by the number of receive antennas. Therefore,
ber of targets. A first solution would be to estimate the number of we suggest the following choice:
targets and the DOAs as presented in Appendix A, provided M is
larger than the number of targets. Then, a second waveform would
Gc ≥ M and Uc ≥ M (54)
be transmitted to estimate the ranges and velocities. However, this The next step is to define G1 and U1 . For the sake of simplicity, we
strategy based on the transmission of two different consecutive can consider:
waveforms may be questionable for various reasons: the number
of targets can vary over time, their DOAs can evolve over time, etc. G1 = Gc and U1 = Uc (55)
Another strategy is to define the values of Gc and Uc by con-
Figure 4 shows an example of allocation of sub-carriers and OFDM
sidering the maximum number of targets that can be detected
symbols dedicated to the Radar and communication systems for
when subspace methods are used. Thus, from Appendix A and
different values of Gi and Ui .
Sections 3.1 and 3.2, the maximum number of target DOAs, ranges
Let us now look at the influence of the system parameters on
and velocities that can be estimated is equal to M − 1, G − 1 and
the data rate.

9
S. Bhogavalli, K.V.S. Hari, E. Grivel et al. Signal Processing 209 (2023) 109007

Fig. 3. Processing chain for MIMO OFDM DFRC when SS-SS-SS approach is used.

5.1.2. Data rate analysis over all the sub-carriers from all the transmit antennas during all
The waveform containing the data symbols is transmitted from OFDM symbols. This waveform is transmitted during the first snap-
the DFRC system towards a standard MIMO OFDM communication shot and the same is repeated for the remaining N − 1 snapshots.
receiver as shown in Fig. 3. At the communication receiver, the Figure 5 shows the processing chain of the Radar system when
signal has the form of pulses where each pulse has N snapshots, the approach is SS-TLS-TLS. The waveform containing the informa-
as shown in Fig. 2. The data symbols transmitted during the first tion bits is transmitted from the DFRC system towards a standard
snapshot are repeated during the remaining (N − 1 ) snapshots, as MIMO OFDM communication receiver, as shown in Fig. 6. For this
mentioned in Section 1. In this case, the data rate Rd that can be design of the waveform, the data rate that can be offered is equal
obtained is equal to: to:

Rd = (GU − GcU1 − Uc G1 + GiUi + 1 )L log2 B bits per N snapshots Rd = GUL log2 B bits per N snapshots (61)
(56)
Given the above analysis on the system parameter selection for
different approaches, let us now analyze the performance of the
(55 )
= (GU − 2GcUc + GiUi + 1 )L log2 B bits per N snapshots (57) Radar system.

where B is the constellation order.


6. Simulation results
The larger the product GiUi , the larger Rd . Therefore, the max-
imum data rate is achievable when Gc = G1 = Gi and Uc = U1 = Ui .
In this section, the simulation protocol is presented in
In that case, one has:
Section 6.1. Then in Sections 6.2 and 6.3, we compare different
Rd = (GU − GcUc + 1 )L log2 B bits per N snapshots (58) approaches. In Section 6.4, we compare the proposed approaches
with the CRB. Finally, in Section 6.5, the trade-off between the
Radar and the communication systems is presented.
(54 )

≤ GU − M2 + 1 L log2 B bits per N snapshots (59)


6.1. Simulation protocol
Since L = M, Rd can be written as:

Let us first summarize the assumptions on the transmitted data
Rd ≤ GU − L2 + 1 L log2 B bits per N snapshots (60)
required for the proposed approaches in Table 3. This will help us
Therefore, the waveform is designed so that the same data is trans- define the simulation protocol presented in this paper. The data
mitted over Gc sub-carriers during Uc OFDM symbols in order to model used in Xu et al. [19] is also recalled.
estimate the range and the velocity. The LS method is not indicated in Table 3 for the sake of
compactness. Indeed, wherever the TLS method is applicable, the
5.2. When subspace methods are used to estimate DOA and TLS LS method can be also used. We a priori suggest using the TLS
methods are used to estimate range and velocity (SS-TLS-TLS) method as it takes into account different types of approximations.
In addition, the approaches such as SS-TLS-TLS and FT-FT-FT do not
When the SS-TLS-TLS approach is used to estimate the target require any assumption on the transmitted data. Therefore, they
parameters, data replication is no longer needed over any sub- can be applicable to all the data models listed in Table 3.
carrier during any OFDM symbol. Therefore, the waveform can For the simulations, the following steps are considered to es-
be designed in such a way that different data are transmitted timate the DOA, the range and the velocity using the subspace

10
S. Bhogavalli, K.V.S. Hari, E. Grivel et al. Signal Processing 209 (2023) 109007

Fig. 4. Illustration of the allocation of sub-carriers or OFDM symbols for Radar and communication systems for different values of Gi and Ui with G = 6, U = 4, Gc = G1 =
Gi = 3, Uc = U1 = Ui = 2.

Table 3
Assumptions on the transmitted data and corresponding usable approaches among the proposed and the existing approaches.

Index Assumption on transmitted data Interpretation Potential approaches that can be used to
estimate DOA-range-velocity

Data Model 1 No assumption on s(g, u, l )∀g, u and l The transmitted data is different over SS-TLS-TLS (proposed)
allsub-carriers during all OFDM symbols FT-FT-FT [19]
acrossall antennas. All the sub-carriers are
shared.
Data Model 2 s(Gmin , u, l ) = s(Gmin + 1, u, l ) . . . = The transmitted data is the same SS-SS-TLS (proposed)
s(Gmin + Gc − 1, u, l ) ∀u ∈ U1 and l overconsecutive Gc sub-carriers during SS-TLS-TLS (proposed)
U1 OFDM symbols. All the sub-carriers are FT-FT-FT [19]
shared.
Data Model 3 s(g, Umin , l ) = The transmitted data is the same SS-TLS-SS (proposed)
s(g, Umin + 1, l ) · · · =s(g, Umin + Uc − 1, l ) duringconsecutive Uc OFDM symbols over SS-TLS-TLS (proposed)
∀g ∈ G1 and l G1 sub-carriers. All the sub-carriers are shared. FT-FT-FT [19]
Data Model 4 s(Gmin , u, l ) · · · = s(Gmin + Gc − 1, u, l )= The transmitted data is the same overGc SS-SS-SS (proposed)
s(g, Umin , l ) · · · = s(g, Umin + Uc − 1, l )∀g ∈ sub-carriers during U1 OFDM symbolsand over SS-SS-TLS (proposed)
G1 , u ∈ U1 , and l G1 sub-carriers during Uc OFDMsymbols. All SS-TLS-SS (proposed)
the sub-carriers are shared. SS-TLS-TLS (proposed)
FT-FT-FT [19]
Data Model 5 Let G be the set of private sub-carriers. For A set of L sub-carriers among G Lasso-FT-FT [19]
g ∈ G,s(g, u, l ) = 0 if the gth sub-carrier is sub-carriersare private while remaining FT-FT-FT [19]
paired tothe l th transmit antenna. Otherwise, sub-carriersare shared. Data transmitted over SS-TLS-SS (proposed)
s(g, u, l ) = 0.For g ∈/ G, no assumption on allsub-carriers during all OFDM symbols SS-TLS-TLS (proposed)
s(g, u, l )∀u and l . acrossall antennas is different.

11
S. Bhogavalli, K.V.S. Hari, E. Grivel et al. Signal Processing 209 (2023) 109007

Table 4
System parameters and target parameters.

Symbol Value Symbol Value

fc 2 GHz K 2
f 20 KHz d λc /2
L 16 N 10
M 16 TCP 3.125μs
G 64 (θ1 , D1 , v1 ) (39.32◦ , 153.73 m, 31.72 m/s )
Fig. 5. Processing chain for the Radar system for the approach SS-TLS-TLS . U 64 (θ2 , D2 , v2 ) (49.32◦ , 903.73 m, 81.72 m/s )

method. Once the data y(g, u, m ) is received, the correlation matri-


ˆ θ (g, u, M ), R
ˆ D (u, m, Gc ) and R
ˆ v (g, m, Uc ) are estimated to de- all values of RCS of the target, to be equal to 1. The system param-
ces R
eters and target parameters used in the simulations are presented
duce the DOA, the range and the velocity as stated in Appendix A,
in Table 4. The target parameters are chosen randomly to illustrate
3.1 and 3.2 respectively. These estimated correlation matrices are
our work.
then averaged across N snapshots along with the corresponding
The signal-to-noise ratio (SNR) for each target at the receiver in
degree of freedom parameter and are given by:
a trial is calculated as follows: First of all, given (7), let us stack the
N G−1 U−1
1  samples k (g, u, m ) in a vector for g = 0, . . . , G − 1, u = 0, . . . , U − 1
ˆ θ (M ) =
R ˆ θ (g, u, M )
R (62)
NGU and m = 0, . . . , M − 1. This results in:
n=1 g=0 u=0 T
zk = k (0, 0, 0 ) . . . k (G − 1, 0, 0 ) . . . k (G − 1, U − 1, M − 1 )
where Rˆ θ (g, u, M ) is the estimation of the correlation matrix de-
fined in Eq. (A.9). (65)

1 
N  
M−1 Similarly, the noise samples η (g, u, m ) are stacked in a vector for
ˆ D ( Gc ) =
R ˆ D (u, m, Gc )
R (63) g = 0, . . . , G − 1, u = 0, . . . , U − 1 and m = 0, . . . , M − 1, leading to:
NU1 M
n=1 u∈U1 m=0

where Rˆ D (u, m, Gc ) is the estimation of the correlation matrix in- η = [η ( 0 , 0 , 0 ) . . . η ( G − 1 , 0 , 0 ) . . . η ( G − 1 , U − 1 , 0 ) . . .


troduced in Eq. (18).
η (G − 1, U − 1, M − 1 )]T (66)
1 
N M−1
ˆ v (Uc ) =
R ˆ v (g, m, Uc )
R (64) The sample SNR for the kth target is then defined by:
NG1 M
 = 10 log zk
2
n=1 g∈G1 m=0
SNR (67)
where R ˆ v (g, m, Uc ) is the estimation of the correlation matrix in-
k 10 2 η
troduced in Eq. (29). In the following, the simulation results will be given for a fixed
Then, given these correlation matrices, the Root-MUSIC method  and a fixed SNR
SNR  . To this end, for each trial, given σ and the
1 2 1
is used to estimate the target parameters. set of transmitted symbols s(g, u, l ), the noise-variance is adjusted
In this section, the simulations are carried out for an air-borne to get SNR1 . Then, σ2 is adjusted to get SNR2 . To present the per-
scenario with two targets for 100 trials in the presence of an addi- formance of the different approaches, the estimated target parame-
tive white Gaussian noise. We consider σavg , the statistical mean of ter values, i.e. (θˆk , Dˆ k , vˆ k ) for k = 1, . . . , K, are provided for different

Fig. 6. Processing chain for MIMO OFDM DFRC when SS-TLS-TLS approach is used.

12
S. Bhogavalli, K.V.S. Hari, E. Grivel et al. Signal Processing 209 (2023) 109007

Fig. 7. DOA, range and velocity estimates from SS-SS-SS, SS-SS-TLS, SS-TLS-SS, SS-TL S-TL S and FT-FT-FT based on data model 4 using the Swerling model for RCS for 100
trials and 10 snapshots at (SNR1 , SNR2 ) = (20, 20 ) dB. True targets parameters are (39.32◦ , 153.73 m, 31.72 m/s ) and (49.32◦ , 903.73 m, 81.72 m/s ).

values of SNR using 3-D scatter plots. In addition, when necessary, the the difference between two target DOAs. From Figs. 7, 8,
tables and figures will give the Relative Root Mean Square Error one can observe that the DOA, range and velocity estimates
(RRMSE), which is the performance metric given by: form a cluster around the true values. As we could expect,
 
the size of this cluster increases when the SNRs decrease. In
100
i=1 |x − xˆi |2 /100 Fig. 7, the estimates obtained with the FT-FT-FT approach are
RRMSE = (68) far from the true values compared with all other approaches.
x
In Fig. 8, all the estimates obtained when FT-FT-FT is used, de-
where x and xˆi are the true target parameter value and its estimate viate from the true values and hence cannot be trusted. The
for the ith trial respectively. reason for this unwanted phenomenon has already been men-
We launched various sets of simulations based on the different tioned in the introduction of the paper. Besides, all the estimates
data models. For the sake of compatibility, we present the results obtained with the SS-SS-SS approach are clustered around the
obtained when data model 4 and data model 5 are used. Note that true values in both scenarios. With SS-SS-TLS, the velocity esti-
the subspace method Root-MUSIC is used for all the simulations. mates tend to spread around the true value when the SNR de-
When applying FT-FT-FT, the quantity d sin θ fc +cg f is estimated by creases. The same phenomenon is observed for the range estimates
looking at 256 values uniformly distributed between 0 and 1. Con- when using SS-TLS-SS. Consequently, both phenomena appear with
sequently, due to the sine function, the DOA is estimated using a SS-TLS-TLS.
non-uniform grid. When using Lasso-FT-FT, the DOA grid is uni- In Table 5, the smaller RRMSE values are indicated in bold. The
form and set at 512 candidate values. The grid sizes used to es- RRMSE value of the DOA is the same for all approaches as the sub-
timate the target range and the velocity are 2048 and 1024 re- space method is used to estimate the DOA. The RRMSE of the range
spectively. Finally, we present the performance of the Radar and is the same for SS-SS-SS and SS-SS-TLS since the subspace method
communication systems in terms of RRMSE values of the estimated is used to estimate the target range. Similarly, the RRMSE of the
target parameters and the data rate offered to the communication range is the same for SS-TLS-SS and SS-TLS-TLS. Similar comments
systems with respect to the system parameters. can be made for the RRMSE of the velocity estimates. For most
In what follows, we present the results and give some com- cases, the RRMSE value decreases when the number of snapshots
ments. increases. These results can be mainly explained by the benefit of
the averaging of the estimations of the correlation matrices and
6.2. Comparative study between SS-SS-SS, SS-SS-TLS, SS-TLS-SS, the received symbols over the snapshots for subspace methods and
SS-TLS-TLS and FT-FT-FT based on data model 4 LS/TLS methods respectively.
For the sake of space, we present results only for SS-SS-SS and
In this subsection, after presenting some illustrative results SS-TLS-TLS approaches. In Fig. 9 (a), the RRMSE values of the DOA
for all approaches, we analyse the relevance of the proposed ap- are the same for both SS-SS-SS and SS-TLS-TLS approaches because
proaches by reducing the number of snapshots, the SNR and the subspace method is used in both cases. The RRMSE of the

13
S. Bhogavalli, K.V.S. Hari, E. Grivel et al. Signal Processing 209 (2023) 109007

Fig. 8. DOA, range and velocity estimates with SS-SS-SS, SS-SS-TLS, SS-TLS-SS, SS-TL S-TL S and FT-FT-FT based on data model 4 using the Swerling model for RCS for 100
trials and 10 snapshots at (SNR1 , SNR2 ) = (0, 0 ) dB. True targets parameters are (39.32◦ , 153.73 m, 31.72 m/s ) and (49.32◦ , 903.73 m, 81.72 m/s ).

Table 5
RRMSE of target parameters vs. the number of snapshots for the approaches based on data model 4, using the Swerling model for RCS. (SNR1 , SNR2 ) = (0, 0 ) dB. True
target parameters are: (39.32◦ , 153.73 m, 31.72 m/s ) and (49.32◦ , 903.73 m, 81.72 m/s ).

N SS-SS-SS SS-SS-TLS SS-TLS-SS SS-TLS-TLS

Tgt 1 Tgt 2 Tgt 1 Tgt 2 Tgt 1 Tgt 2 Tgt 1 Tgt 2

RRMSE of θˆk at (SNR1 , SNR2 ) = (0, 0 ) for (θ1 , θ2 ) = (39.32◦ , 49.32◦ )


3 4.94e-04 4.67e-04 4.94e-04 4.67e-04 4.94e-04 4.67e-04 4.94e-04 4.67e-04
5 4.16e-04 4.88e-04 4.16e-04 4.88e-04 4.16e-04 4.88e-04 4.16e-04 4.88e-04
10 3.72e-04 4.44e-04 3.72e-04 4.44e-04 3.72e-04 4.44e-04 3.72e-04 4.44e-04
RRMSE of Dˆ k at (SNR1 , SNR2 ) = (0, 0 ) for (D1 , D2 ) = (153.73 m, 903.73 m )
3 2.51e-03 5.61e-04 2.51e-03 5.61e-04 2.63e-01 5.42e-02 2.63e-01 5.42e-02
5 2.61e-03 5.57e-04 2.61e-03 5.57e-04 2.52e-01 4.32e-02 2.52e-01 4.32e-02
10 2.84e-03 5.41e-04 2.84e-03 5.41e-04 2.00e-01 2.87e-02 2.00e-01 2.87e-02
RRMSE of vˆ k at (SNR1 , SNR2 ) = (0, 0 ) dB for (v1 , v2 ) = (31.72 m/s, 81.72 m/s )
3 5.83e-04 2.15e-04 1.24e-01 5.21e-02 5.83e-04 2.15e-04 1.24e-01 5.21e-02
5 5.17e-04 1.86e-04 1.05e-01 3.97e-02 5.17e-04 1.86e-04 1.05e-01 3.97e-02
10 3.91e-04 1.38e-04 7.49e-02 3.34e-02 3.91e-04 1.38e-04 7.49e-02 3.34e-02

range estimate in Fig. 9 (b) does not become smaller when SNR estimates obtained for the SS-SS-SS approach do not vary with θ
increases. This is probably because of the approximation used to since the estimation of range and velocity is independent of the
obtain the Eq. (14) for the estimation of the range ( fc + g f ≈ fc ). DOA estimation.
In addition, one can see that the RRMSE of the range when the Based on this comparative study, we can conclude that the ap-
SS-SS-SS approach is used is much smaller than the RRMSE of the proach SS-SS-SS is the most reliable. In the next subsection, let us
range obtained with the SS-TLS-TLS at any SNR. Thus, when the analyze the approaches based on data model 5.
SNRs are equal to 20dB, the difference between the two RRMSE is
equal to 9 dB. In Fig. 9 (c), the difference is even equal to 24dB. 6.3. Comparative study between Lasso-FT-FT, FT-FT-FT, SS-TLS-SS and
According to Fig. 10 (a), the RRMSE of the DOA estimates from SS- SS-TLS-TLS based on data model 5
SS-SS decreases when θ the difference between the DOA of the
two targets increases. From Fig. 10 (b), (c), the RRMSE of the range In Fig. 11, the estimates for Lasso-FT-FT and FT-FT-FT ap-
and the velocity when the SS-TLS-TLS approach is used increases proaches are far from the true values compared with the proposed
when θ decreases. This is due to the fact that the DOA estimates approaches. This can be due to the use of the Swerling model. The
are used to estimate the range and the velocity in that case. In con- Swerling model introduces some randomness in the amplitudes of
trast, from Fig. 10 (b), (c), the RRMSE of the range and the velocity the received signal from the targets which are no longer the same

14
S. Bhogavalli, K.V.S. Hari, E. Grivel et al. Signal Processing 209 (2023) 109007

Fig. 9. System parameters are G = 64, U = 64, L = M = 16, 4−QAM for the data model 4 using the Swerling model for RCS. True target parameters are:
(39.32◦ , 153.73 m, 31.72 m/s ), (49.32◦ , 903.73 m, 81.72 m/s ) and SNR1 = SNR2 for 100 trials and 10 snapshots.

Fig. 10. System parameters are G = 64, U = 64,L=M=16, 4-QAM. True target parameters are (49.32◦ − θ , 153.73 m, 31.72 m/s ),(49.32◦ , 903.73 m, 81.72 m/s ) for the data
model 4 using the Swerling model for RCS at SNR = (20, 20) dB for 100 trials and 10 snapshots.

(unlike [19]). There is a strong risk that the target DOA associated tion and deriving it with respect to the parameters of interest as
with the smaller power cannot be estimated properly due to the in Stoica and Nehorai [47], Stoica et al. [48].
windowing effect. Moreover, the difference in the grid size and its The SNR of the kth target for one snapshot, useful when consid-
uniformity or non-uniformity may have an impact on the result. ering the CRB, is given by:
Among all the approaches, SS-TLS-SS and SS-TLS-TLS perform bet- G−1 U−1 M−1
g=0 u=0 m=0 E k (g, u, m ) k∗ (g, u, m )
ter as they form a cluster around the true values. SNRk = G−1 U−1 M−1
Given the various simulations we launched, it is confirmed g=0 u=0 m=0 E[ η (g, u, m )η∗ (g, u, m )]
that the proposed approaches outperform the existing approaches λ2
GUM (4π )3 D4 σ L
c
λ 2
[19]. avg c
= k
= L (69)
GUMση 2 ( 4π ) 3 D4 σ σ 2
k avg η
The above expression is obtained by taking into account the sta-
6.4. Comparison of the proposed approaches with the Cramer–Rao  ,
tistical properties of the symbols s(g, u, l ) and σk . Note that SNR k
bound which is the sample SNR introduced in (67), can be seen as an ap-
proximation of the SNR when considering a single realization of
In this section, we suggest comparing the results obtained with the process. It is used to get the MSE of our approach.
our approaches with the CRB when the purpose is to estimate the In Fig. 12, the MSE is considered instead of the RRMSE so that
DOA, the range and the velocity in a MIMO OFDM DFRC system, the comparison can be done with the CRB. From Fig. 12 (a), the dif-
based on Eqs. (3), (6) and (7). To this end, the Fischer informa- ference between the MSE of the DOA obtained with SS-SS-SS and
tion matrix has been expressed and then its inverse has been com- the CRB is around 18 dB. In Fig. 12 (b), the difference between the
puted. Note that for the sake of space, we will not provide the ex- MSE of the range with SS-SS-SS and the CRB increases from 46 dB
pression of each element of the Fischer information matrix and its to 116 dB when SNR increases from 0 dB to 30 dB. It is probably
inverse, but it can be obtained by expressing the likelihood func- due to the approximation of fc + g f by fc ∀ g ∈ {0, . . . , G − 1} in

15
S. Bhogavalli, K.V.S. Hari, E. Grivel et al. Signal Processing 209 (2023) 109007

Fig. 11. DOA, range and velocity estimates with Lasso-FT-FT, FT-FT-FT, SS-TLS-SS and SS-TL S-TL S based on data model 5 using the Swerling model for RCS for 100 trials and
10 snapshots at (SNR1 , SNR2 ) = (20, 20 ) dB. True targets parameters are (39.32◦ , 153.73 m, 31.72 m/s ) and (49.32◦ , 903.73 m, 81.72 m/s ).

Fig. 12. 10log10 (MSE) (in dB) of the 1st target parameter obtained with the proposed approaches and 10log10 (CRB) vs. SNR. System parameters are G = 64, U = 64, L = M =
16, 4−QAM. True target parameters are: (39.32◦ , 153.73 m, 31.72 m/s ), (49.32◦ , 903.73 m, 81.72 m/s ) for 100 trials and 10 snapshots.

order to estimate the range using subspace methods. The differ- crease. The RRMSE of the range (resp. the velocity) increases when
ence between the MSE of the range obtained with SS-TLS-TLS and Gc (resp. Uc ) increases because the dimension of the noise sub-
the CRB is around 130 dB at every SNR. Similarly, in Fig. 12 (c), the space also increases. The RRMSE of the range (resp. the veloc-
difference between the MSE of the velocity obtained with the pro- ity) increases when Uc (resp. Gc ) increases because the estima-
posed approaches (SS-SS-SS, SS-TLS-TLS) and the CRB is 3 dB and tion of the correlation matrix is based on larger set of data. In
110 dB respectively. Fig. 13 (a), an error floor occurs for all cases when the SNR is
higher than 18 dB. This is probably due to the fact that fc + g f is
6.5. Trade-off to be found between the performances of the Radar approximated by fc ∀ g = 0, . . . , G − 1. From Fig. 13 (b), the RRMSE
and the communication systems of the velocity becomes smaller when Gc and Uc increase. Note
that the improvement of the RRMSE of the velocity due to the
In this subsection, we present how the RRMSEs of the estimated increase in Uc is better than the improvement due to Gc . How-
target parameters and the data rate vary with the system parame- ever, when referring to Eq. (60), the data rate decreases when
ters such as Gc , Uc when the SS-SS-SS approach is used. the product of Gc and Uc increases. The larger the values of
The simulations are carried out using the strategy mentioned in Gc and Uc are, the smaller the RRMSE of target parameters the
Section 5.1, i.e., Gc = G1 = Gi ≥ L and Uc = U1 = Ui ≥ L7 larger the product of GcUc , the smaller Rd . Therefore, the selec-
According to Fig. 13 (a) and (b), the RRMSE of the range and tion of Gc and Uc depends on the priority the practitioner has
velocity estimates for the 1st target increases when Gc and Uc in- between Radar performance and data rate. A trade-off has to be
found.

7
Therefore, we will only refer to Gc and Uc .

16
S. Bhogavalli, K.V.S. Hari, E. Grivel et al. Signal Processing 209 (2023) 109007

Fig. 13. Variation of RRMSE vs. SNR for the data model 4 using the Swerling model for RCS when the system parameters are G = 64, U = 64, L = M = 16 and 4−QAM at SNR
= (20, 20) dB for 100 trials and 10 snapshots. True targets parameters are: (39.32◦ , 153.73 m, 31.72 m/s ),(49.32◦ , 903.73 m, 81.72 m/s ).

7. Conclusions and perspectives partially funded by the Prime Minister’s Research Fellows (PMRF)
scheme, Government of India.
This paper deals with the estimation of the range and the ve-
locity along with the DOA of several targets by using subspace Appendix A. Estimation of the DOA using subspace methods
methods in a MIMO OFDM DFRC system. Such an approach can
be done if the data symbols used to design the Radar waveform This appendix deals with the estimation of the DOA using sub-
are replicated over a few sub-carriers during a few OFDM symbols. space methods. Even if these methods have been widely used in
However, this has an impact on the data rate that can be offered previous works in the past, we recall the main steps of the ap-
for the communication system. To avoid data replication, the range proach by adjusting them to the context under study. Let us ex-
and the velocity can be estimated by using LS/TLS methods. press y(g, u, M ) defined in the Eq. (8) above:
According to the various simulations we did, subspace methods
y(g, u, M ) = Aθ (g, M )(g, u )ATθ (g, L )s(g, u, L ) + η(g, u, M ) (A.1)
provide better performance in terms of estimation accuracy, but
the association of the target parameters needs to be done a poste- where Aθ (g, M ) is the received steering matrix of dimension M × K
riori. Moreover, the achievable data rate is impacted because data given by:
replication is required. A trade-off between the accuracy of the tar- ⎡ ⎤
get parameter estimation and the data rate has to be found by the 1 ... 1
practitioner. It should be noted that TLS methods provide less ac- ⎢ exp(− jω1 (g) ) ... exp(− jωK (g) ) ⎥
Aθ (g, M ) = ⎢ . . ⎥
curate target parameter estimates, but they have the advantage of ⎣ . . ⎦
. .
including the association of the target parameters. Finally, all the
exp(− j (M − 1 )ω1 (g)) ... exp(− j (M − 1 )ωK (g))
proposed approaches outperform the existing approaches based on
the FT and Lasso methods when introducing the Swerling model in (A.2)
the system modelling. with for k = 1, . . . , K:

Declaration of Competing Interest fc + g f


ωk (g) = 2π d sin θk (A.3)
c
The authors declare that they have no known competing finan- Note that the transmit matrix Aθ (g, L ) is similarly defined as
cial interests or personal relationships that could have appeared to Aθ (g, M ). Moreover, both have a Vandermonde structure due to the
influence the work reported in this paper. ULA structure.
In (A.1), y(g, u, M ) also depends on the following square matrix
CRediT authorship contribution statement of size K.
  2D


Satwika Bhogavalli: Conceptualization, Methodology, Software, (g, u ) = diag α1 exp − j2π g f 1 exp j2π uT f1d · · · αK
Writing – original draft. K.V.S. Hari: Conceptualization, Methodol- c
 2 DK


ogy, Supervision, Writing – original draft. Eric Grivel: Conceptual- exp − j2π g f exp j2π uT fKd (A.4)
ization, Methodology, Supervision, Writing – original draft. Vincent c
Corretja: Supervision, Writing – original draft.
Moreover, the symbols modulated over the gth sub-carrier in the
Data availability uth OFDM symbol and transmitted by the L antennas are stored in
the following vector:
No data was used for the research described in the article. T
s(g, u, L ) = s(g, u, 0 ) s(g, u, 1 ) ... s(g, u, L − 1 ) (A.5)
Acknowledgements Finally, the corresponding additive-noise vector at the receiver is
defined by:
The work is the result of the collaboration between IISc, the T
University of Bordeaux and Thales DMS France. This work is η(g, u, M ) = η (g, u, 0 ) η (g, u, 1 ) ... η (g, u, M − 1 ) (A.6)

17
S. Bhogavalli, K.V.S. Hari, E. Grivel et al. Signal Processing 209 (2023) 109007

In the system under study, M = L. Therefore, one has: The number of targets K can be determined from the number
of predominant eigenvalues of the correlation matrix (A.10) by us-
Aθ (g, M ) = Aθ (g, L ) = Aθ (g) (A.7)
ing one of the criteria proposed in Konstantinides and Yao [49].
So, re-writing the Eq. (A.1) leads to: This question, which is not only of interest for DOA estimation
but in many other applications such as spectrum analysis [50] or
y(g, u, M ) = Aθ (g)(g, u )ATθ (g)s(g, u, L ) + η(g, u, M ) (A.8)
speech enhancement [51] when the signal is modelled by a sum of
Given (A.8), the DOA of each target has to be estimated. As the complex exponentials, was more recently addressed in Adali and
DOA of the kth target is related to ωk (g) (which is introduced in Haykin [52]. It should be noted that due to the size of the correla-
(11) and appears in the matrix Aθ (g)), one has to characterize the tion matrix, the maximum number of targets that can be estimated
space spanned by the columns of Aθ (g). To this end, the corre- should be less than the number of receive antennas, i.e. K < M.
lation matrix Rθ (g, u, M ) of the received data vector y(g, u, M ) is Similarly, the noise variance can be deduced from the smallest
studied. Since the additive noise is independent of the transmitted eigenvalues of the correlation matrix (A.10).
data, Rθ (g, u, M ) is given by:

Rθ (g, u, M )=E y(g,u,M )yH (g,u,M )
References
θ ( g ) + ση I M
= Aθ (g)(g, u )ATθ (g)Rs (g, u, L )A∗θ (g)H (g, u )AH 2
[1] A. Hassanien, M.G. Amin, E. Aboutanios, B. Himed, Dual-function radar com-
(A.9) munication systems: a solution to the spectrum congestion problem, IEEE Sig-
nal Process. Mag. 36 (5) (2019) 115–126, doi:10.1109/MSP.2019.2900571.
where Rs (g, u, L ) is the correlation matrix of the transmitted data [2] A. Martone, M. Amin, A view on radar and communication systems coexistence
s(g, u, L ) equal to σs2 IL . Therefore, (A.9) can be rewritten as follows: and dual functionality in the era of spectrum sensing, Digit. Signal Process. 119
(2021) 103135, doi:10.1016/j.dsp.2021.103135.
[3] J.A. Zhang, F. Liu, C. Masouros, R.W. Heath, Z. Feng, L. Zheng, A. Petropulu, An
overview of signal processing techniques for joint communication and radar
Rθ (g, u, M ) = σs2 Aθ (g)(g, u )ATθ (g)A∗θ (g)H (g, u )AH
θ ( g ) + ση I M
2
sensing, IEEE J. Sel. Top. Signal Process. 15 (6) (2021) 1295–1315, doi:10.1109/
(A.10) jstsp.2021.3113120.
[4] F. Liu, C. Masouros, A.P. Petropulu, H. Griffiths, L. Hanzo, Joint radar and com-
In (A.10), (g, u ) is a diagonal matrix and the columns of Aθ (g) are munication design: applications, state-of-the-art, and the road ahead, IEEE
Trans. Commun. 68 (6) (2020) 3834–3862, doi:10.1109/TCOMM.2020.2973976.
linearly independent. Both (g, u ) and Aθ (g) have consequently a [5] F. Liu, C. Masouros, H. Griffiths, Dual-functional radar-communication wave-
full rank, equal to K. The signal subspace associated with the corre- form design under constant-modulus and orthogonality constraints, in: 2019
lation matrix equal to σs2 Aθ (g)(g, u )ATθ (g)A∗θ (g)H (g, u )AH
θ (g) is
Sensor Signal Processing for Defence Conference (SSPD), 2019, pp. 1–5, doi:10.
1109/SSPD.2019.8751644.
of dimension K. Therefore, the subspace methods like MUSIC [32],
[6] B. Tang, H. Wang, L. Qin, L. Li, Waveform design for dual-function MIMO
Root-MUSIC [33] and ESPRIT [34] are applicable to estimate the radar-communication systems, in: 2020 IEEE 11th Sensor Array and Multichan-
DOA of K targets from the full rank correlation matrix Rθ (g, u, M ). nel Signal Processing Workshop (SAM), 2020, pp. 1–5, doi:10.1109/SAM48682.
2020.9104378.
Now, we can estimate the DOA using one of the following two
[7] Y.L. Sit, T.T. Nguyen, C. Sturm, T. Zwick, 2D radar imaging with velocity esti-
schemes for the estimation of the correlation matrix. mation using a MIMO OFDM-based radar for automotive applications, in: 2013
Scheme 1: As the matrix Aθ (g) does not change with the OFDM European Radar Conference, 2013, pp. 145–148.
symbol index u, the correlation matrices {Rθ (g, u, M )}u 0,...,U−1 [8] D. Kuswidiastuti, M. Rizky, P.H. Mukti, G. Hendrantoro, MIMO radar waveform
design using interleaved-OFDM technique, in: 2016 IEEE International Confer-
can be averaged over all OFDM symbols, resulting in the matrix ence on Communication, Networks and Satellite (COMNETSAT), 2016, pp. 53–
Rθ (g, M ). For the gth sub-carrier, with g varying from 0 to G − 1, 59, doi:10.1109/COMNETSAT.2016.7907416.
the DOA is estimated by applying a subspace method to the corre- [9] B. Nuss, L. Sit, M. Fennel, J. Mayer, T. Mahler, T. Zwick, MIMO OFDM radar sys-
tem for drone detection, in: 2017 18th International Radar Symposium (IRS),
lation matrix Rθ (g, M ). Then, the mean of the resulting DOA esti- 2017, pp. 1–9, doi:10.23919/IRS.2017.8008141.
mates for all the sub-carriers can be considered. [10] S.D. Blunt, M.R. Cook, J. Stiles, Embedding information into radar emissions
via waveform implementation, in: 2010 International Waveform Diversity and
1 
U−1
Design Conference, 2010, pp. 0 0 0195–0 0 0199, doi:10.1109/WDD.2010.5592502.
Rθ (g, M ) = Rθ (g, u, M ) (A.11) [11] J. Euzière, R. Guinvarc’h, M. Lesturgie, B. Uguen, R. Gillard, Dual function radar
U
u=0 communication time-modulated array, in: 2014 International Radar Confer-
ence, 2014, pp. 1–4, doi:10.1109/RADAR.2014.7060416.
However, the computational cost of this scheme can be quite high [12] A. Hassanien, M.G. Amin, Y.D. Zhang, F. Ahmad, A dual function radar-
as subspace methods are applied G times. To reduce it while still communications system using sidelobe control and waveform diversity, in:
taking advantage of the data transmitted over all the sub-carriers, 2015 IEEE Radar Conference (RadarCon), 2015, pp. 1260–1263, doi:10.1109/
RADAR.2015.7131188.
the following scheme can be considered. [13] A. Hassanien, M.G. Amin, Y.D. Zhang, F. Ahmad, Dual-function radar-
Scheme 2: Given the orders of magnitude of fc and  f , fc + communications using phase-rotational invariance, in: 2015 23rd European
g f can be approximated by fc for g = 0, 1, . . . , G − 1. In that case, Signal Processing Conference (EUSIPCO), 2015, pp. 1346–1350, doi:10.1109/
EUSIPCO.2015.7362603.
the matrix Aθ (g) does not vary much with the sub-carrier index
[14] A. Ahmed, Y.D. Zhang, Y. Gu, Dual-function radar-communications using QAM-
g. Therefore, in terms of notation, we suggest dropping the index based sidelobe modulation, Digit. Signal Process. 82 (2018) 166–174, doi:10.
g in Aθ (g). For any value of g and u, we assume that the signal 1016/j.dsp.2018.06.018.
[15] L. Zheng, M. Lops, Y.C. Eldar, X. Wang, Radar and communication coexistence:
subspace “approximately remains the same”. We can average the
an overview: areview of recent methods, IEEE Signal Process. Mag. 36 (5)
correlation matrices Rθ (g, u, M ), for all G sub-carriers and U OFDM (2019) 85–99, doi:10.1109/MSP.2019.2907329.
symbols and then apply subspace methods to the following result- [16] A. Hassanien, M.G. Amin, Y.D. Zhang, F. Ahmad, Signaling strategies for dual-
ing matrix. function radar communications: an overview, IEEE Aerosp. Electron. Syst. Mag.
31 (10) (2016) 36–45, doi:10.1109/MAES.2016.150225.
G−1 U−1
1  1 
U−1 [17] A. Hassanien, E. Aboutanios, M.G. Amin, G.A. Fabrizio, A dual-function MIMO
Rθ ( M ) = Rθ (g, u, M )= Rθ (g, M ) (A.12) radar-communication system via waveform permutation, Digit. Signal Process.
GU U 83 (2018) 118–128, doi:10.1016/j.dsp.2018.08.010.
g=0 u=0 u=0 [18] T. Huang, X. Xu, Y. Liu, N. Shlezinger, Y.C. Eldar, A dual-function radar
communication system using index modulation, in: 2019 IEEE 20th Interna-
From a theoretical point of view, summing or averaging is not nec-
tional Workshop on Signal Processing Advances in Wireless Communications
essarily mandatory. However, in practice, the correlation matrix of (SPAWC), 2019, pp. 1–5, doi:10.1109/SPAWC.2019.8815471.
the noisy received data has to be estimated. Averaging over the [19] Z. Xu, A. Petropulu, S. Sun, A joint design of MIMO-OFDM dual-function radar
OFDM symbols but also on the sub-carriers can be a priori useful communication system using generalized spatial modulation, 2021.
[20] T. Huang, N. Shlezinger, X. Xu, Y. Liu, Y.C. Eldar, Majorcom: a dual-function
to get better estimates of the basis of the signal and noise sub- radar communication system using index modulation, IEEE Trans. Signal Pro-
spaces. cess. 68 (2020) 3423–3438, doi:10.1109/TSP.2020.2994394.

18
S. Bhogavalli, K.V.S. Hari, E. Grivel et al. Signal Processing 209 (2023) 109007

[21] D. Ma, T. Huang, N. Shlezinger, Y. Liu, X. Wang, Y.C. Eldar, A DFRC system [36] F.-G. Yan, M. Jin, S. Liu, X.-L. Qiao, Real-valued MUSIC for efficient direction
based on multi-carrier agile FMCW MIMO radar for vehicular applications, estimation with arbitrary array geometries, IEEE Trans. Signal Process. 62 (6)
in: 2020 IEEE International Conference on Communications Workshops (ICC (2014) 1548–1560, doi:10.1109/TSP.2014.2298384.
Workshops), 2020, pp. 1–7, doi:10.1109/ICCWorkshops49005.2020.9145355. [37] F.-G. Yan, L. Shuai, J. Wang, J. Shi, M. Jin, Real-valued root-MUSIC for DOA es-
[22] S. Zhu, F. Xi, S. Chen, A. Nehorai, A low-complexity MIMO dual function radar timation with reduced-dimension EVD/SVD computation, Signal Process. 152
communication system via one-bit sampling, in: 2021 IEEE International Con- (2018) 1–12, doi:10.1016/j.sigpro.2018.05.009.
ference on Acoustics, Speech and Signal Processing (ICASSP), 2021, pp. 8223– [38] B.D. Rao, K.V.S. Hari, Performance analysis of root-music, IEEE Trans. Acoust.,
8227, doi:10.1109/ICASSP39728.2021.9414051. Speech, Signal Process. 37 (12) (1989) 1939–1949, doi:10.1109/29.45540.
[23] A.G. Jaffer, Maximum likelihood direction finding of stochastic sources: a sep- [39] B.D. Rao, K.V.S. Hari, Performance analysis of ESPRIT and TAM in determining
arable solution, in: ICASSP-88., International Conference on Acoustics, Speech, the direction of arrival of plane waves in noise, IEEE Trans. Acoust., Speech,
and Signal Processing, vol. 5, 1988, pp. 2893–2896, doi:10.1109/ICASSP.1988. Signal Process. 37 (12) (1989) 1990–1995, doi:10.1109/29.45548.
197258. [40] B.D. Rao, K.V.S. Hari, Statistical performance analysis of the minimum-norm
[24] P.K. Eranti, B.D. Barkana, An overview of direction-of-arrival estimation meth- method, in: International Conference on Acoustics, Speech, and Signal Process-
ods using adaptive directional time-frequency distributions, Electronics 11 (9) ing„ 1989, pp. 2760–2763 vol.4, doi:10.1109/ICASSP.1989.267040.
(2022) 1321. [41] J.P. Delmas, Performance bounds and statistical analysis of DOA estimation, in:
[25] Z. Zhang, C. Zhou, Y. Gu, J. Zhou, Z. Shi, An IDFT approach for coprime array Academic Press Library in Signal Processing, vol. 3, Elsevier, 2014, pp. 719–764,
direction-of-arrival estimation, Digit. Signal Process. 94 (2019) 45–55, doi:10. doi:10.1016/B978- 0- 12- 411597- 2.0 0 016-3.
1016/j.dsp.2019.05.006. [42] Z. Xu, A. Petropulu, A bandwidth efficient dual-function radar communication
[26] W. Chen, Y. Li, S. Li, F. Zhang, K. Wang, Efficient FFT based multi source system based on a MIMO radar using OFDM waveforms, IEEE Trans. Signal Pro-
DOA estimation for ULA, in: 2021 IEEE 6th International Conference on Sig- cess. 71 (2023) 401–416, doi:10.1109/TSP.2023.3241779.
nal and Image Processing (ICSIP), 2021, pp. 739–743, doi:10.1109/ICSIP52628. [43] R. Muthukrishnan, R. Rohini, Lasso: a feature selection technique in predictive
2021.9688584. modeling for machine learning, in: 2016 IEEE International Conference on Ad-
[27] H. Wen, C. Li, W. Yao, Power system frequency estimation of sine-wave cor- vances in Computer Applications (ICACA), 2016, pp. 18–20, doi:10.1109/ICACA.
rupted with noise by windowed three-point interpolated DFT, IEEE Trans. 2016.7887916.
Smart Grid 9 (5) (2018) 5163–5172, doi:10.1109/TSG.2017.2682098. [44] S. Bhogavalli, E. Grivel, K.V.S. Hari, V. Corretja, Waveform design to improve
[28] X. Fang, Z. Cao, R. Min, Y. Pi, Radar maneuvering target detection based on two the estimation of target parameters using the fourier transform method in a
steps scaling and fractional Fourier transform, Signal Process. 155 (2019) 1–13, MIMO OFDM DFRC system, in: ICASSP 2023, IEEE, 2023, pp. 1–5.
doi:10.1016/j.sigpro.2018.09.017. [45] M.I. Skolnik, Introduction to Radar Systems, McGraw-Hill Book Co., New York,
[29] J. Capon, High-resolution frequency-wavenumber spectrum analysis, Proc. IEEE 1980.
57 (8) (1969) 1408–1418, doi:10.1109/PROC.1969.7278. [46] G.H. Golub, C.F.V. Loan, An analysis of the total least squares problem, SIAM J.
[30] P. Handel, P. Stoica, T. Soderstrom, Capon method for DOA estimation: accuracy Numer. Anal. 17 (6) (1980) 883–893. http://www.jstor.org/stable/2156807
and robustness aspects, in: IEEE Winter Workshop on Nonlinear Digital Signal [47] P. Stoica, A. Nehorai, Music, maximum likelihood, and Cramer–Rao bound, IEEE
Processing, 1993, pp. P7.1–P7.5, doi:10.1109/NDSP.1993.767766. Trans. Acoust., Speech, Signal Process. 37 (5) (1989) 720–741, doi:10.1109/29.
[31] P. Stoica, P. Händel, T. Söderström, Study of capon method for array signal 17564.
processing, Circuits, Syst. Signal Process. 14 (6) (1995) 749–770, doi:10.1007/ [48] P. Stoica, E.G. Larsson, A.B. Gershman, The stochastic CRB for array processing:
BF01204683. a textbook derivation, IEEE Signal Process. Lett. 8 (5) (2001) 148–150, doi:10.
[32] R. Schmidt, Multiple emitter location and signal parameter estimation, IEEE 1109/97.917699.
Trans. Antennas Propag. 34 (3) (1986) 276–280, doi:10.1109/TAP.1986.1143830. [49] K. Konstantinides, K. Yao, Statistical analysis of effective singular values in ma-
[33] A. Barabell, Improving the resolution performance of eigen structure-based trix rank determination, IEEE Trans. Acoust., Speech, Signal Process. 36 (5)
direction-finding algorithms, in: ICASSP’83 IEEE International Conference on (1988) 757–763, doi:10.1109/29.1585.
Acoustics, Speech, and Signal Processing, vol. 8, 1983, pp. 336–339, doi:10. [50] C.W. Therrien, Discrete Random Signals and Statistical Signal Processing, Pren-
1109/ICASSP.1983.1172124. tice Hall PTR, 1992.
[34] R. Roy, T. Kailath, ESPRIT-Estimation of signal parameters via rotational invari- [51] S.H. Jensen, P.C. Hansen, S.D. Hansen, J.A. Sørensen, A signal subspace ap-
ance techniques, IEEE Trans. Acoust., Speech, Signal Process. 37 (7) (1989) 984– proach for noise reduction of speech signals, in: Signal Processing VII: Theories
995, doi:10.1109/29.32276. and Applications, European Association for Signal Processing (EURASIP), 1994,
[35] W. Wang, X. Wang, H. Song, Y. Ma, Conjugate ESPRIT for DOA estimation in pp. 1174–1177.
monostatic MIMO radar, Signal Process. 93 (7) (2013) 2070–2075, doi:10.1016/ [52] T. Adali, S. Haykin, Subspace Tracking for Signal Processing, 2010, pp. 211–270.
j.sigpro.2013.01.007. 10.1002/9780470575758.ch4

19

You might also like