Signal Processing 156 (2019) 41–45

Contents lists available at ScienceDirect

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

Short communication

Direction of Arrival estimation by modified Orthogonal Propagator

Method with linear prediction in low SNR scenarios
Meng Sun a, Yide Wang b, Jingjing Pan b,∗
a
Information Engineering College, Shanghai Maritime University, Shanghai, People’s Republic of China
b
Institut d’ Electronique et Télécommunications de Rennes (IETR), Université de Nantes, UMR CNRS 6164, Rue Christian Pauc BP 50609, Nantes 44306,
France

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

Article history: In propagator based methods for Direction of Arrival (DoA) finding, conventional approaches for eliminat-
Received 8 June 2018 ing the noise impact are based on signal enhancement or noise power estimation. In this paper, a new
Revised 17 October 2018
method is introduced from another point of view. When the noise is an additive spatially and temporally
Accepted 20 October 2018
white Gaussian noise, theoretically, it only impacts the diagonal elements of the data covariance matrix.
Available online 22 October 2018
Firstly, the proposed method utilizes the principle of Linear Prediction (LP) to rebuild the denoised diag-
Keywords: onal elements of the data covariance matrix. Then, the Orthogonal Propagator Method (OPM) is directly
Direction of Arrival (DoA) estimation applied for DoA estimation. Compared with the conventional OPM-based methods, the proposed method
Linear Prediction (LP) is more robust to the noise, especially in low Signal to Noise Ratio (SNR) scenarios. Simulation examples
Orthogonal Propagator Method (OPM) are provided to demonstrate the performance of the proposed method.
© 2018 Elsevier B.V. All rights reserved.

1. Introduction Loève Transformation (KLT). Stoica et al. [8] propose an eigenvalue-

based method to estimate the power of the noise. However, these
Direction of Arrival (DoA) estimation is of practical importance two methods require EVD. In addition, Marcos and Sanchez-Araujo
in many engineering applications, for example, radar, sonar, and [6] and Stoica et al. [8] also propose two different linear operation
wireless communications [1,2]. A variety of methods have been methods for estimating the power of the noise without EVD.
proposed to estimate DoA. MUSIC [3] and ESPRIT [4] have asymp- In this paper, we propose a new way to eliminate the noise im-
totically infinite resolution but high computational burden, since pact. When the noise is an additive spatially and temporally white
they are based on the EigenValue Decomposition (EVD) of the data Gaussian noise, theoretically, it only affects the main diagonal el-
covariance matrix. Propagator-based methods, for example, Propa- ements of the data covariance matrix. In this case, the noise im-
gator Method (PM) [5,6] and Orthogonal Propagator Method (OPM) pact can be removed by restoring the denoised main diagonal el-
[7], have high resolution and accuracy by using only linear opera- ements. In the configuration of Uniform Linear Array (ULA), the
tions without any EVD operation, which make real time processing data covariance matrix of totally uncorrelated incoming signals is a
possible. Toeplitz and Hermitian matrix. In theory, the elements in the first
Nevertheless, propagator-based methods are developed with column (line) of the data covariance matrix are linear combina-
signal models under ideal condition and without the consideration tions of complex exponentials, which can be modeled by a Linear
of noise. Thus, their performance is degraded in low SNR scenar- Prediction (LP) like model [11]. Therefore, we can restore the de-
ios. There are three possible solutions to solve this problem. Firstly, noised main diagonal elements by using LP and then reconstruct a
the least squares method can be applied to estimate the propaga- denoised covariance matrix.
tor (the principle of PM and OPM). However, this method only re- The rest of this paper is organised as follows: the received sig-
duces the noise impact but does not eliminate it [5]. The other two nal model is presented in Section 2. Section 3 describes the pro-
possible solutions are either to estimate the power of the noise posed method. Simulation results and a discussion on the perfor-
[8,9] or to apply signal enhancement techniques [10]. In [10], Li mance of the conventional OPM, OPM-MPEN (method of propaga-
et al. propose to enhance the received signals by using Karhunen– tor with elimination of noise) [6], OPM-EVD [8], OPM-KLT [10] and
the proposed method are given in Section 4. Conclusions and per-
spectives are drawn in Section 5.
∗
Corresponding author.
E-mail addresses: [email protected] (M. Sun), [email protected]
nantes.fr (J. Pan).

https://doi.org/10.1016/j.sigpro.2018.10.013
0165-1684/© 2018 Elsevier B.V. All rights reserved.
42 M. Sun et al. / Signal Processing 156 (2019) 41–45

Fig. 1. Uniform Linear Array configuration.

2. Signal model which can be rewritten as

⎛ ⎞
R (0 ) R(−1 ) ... R (1 − N )
Consider K far-field narrow-band signals impinging on a ULA ⎜ R (1 ) R (0 ) ... R(2 − N )⎟
R=⎝
... ⎠
with N isotropic antenna elements, as shown in Fig. 1. The received (5)
... ... ...
signals are supposed totally uncorrelated. The indexes of the an- R (N − 1 ) R (N − 2 ) ... R (0 )
tennas are set to 0, 1, . . . , N − 1. Then, the received signal at the
mth element can be written as 3. Modified OPM (OPM-LP) for DoA estimation
 

K
d
rm (t ) = sk (t ) exp j2mπ sin θk + nm (t ) (1) The objective of this work is to build a denoised covariance ma-
k=1
λ trix for the implementation of OPM. Since R is a Toeplitz and Her-
mitian matrix, the main diagonal elements are the same (equal to
where sk (t) is the complex envelope of the kth signal, with DOA θ k ,
R(0)). According to (4), the elements in R are a linear combina-
received at the 0th element; nm (t) is an additive spatially and tem-
tion of exponentials. Therefore, we can use LP for restoring R(0). A
porally white Gaussian noise at the mth element with zero mean
Toeplitz and Hermitian matrix is uniquely determined by its first
and variance σ 2 ; d denotes the distance between two adjacent el-
line (column). In the proposed method, the information of the first
ements and λ is the wavelength of the received signals. (1) can be
column is used in the estimation of R(0). The first column of R is
written in the following vector form:
partitioned into several overlapping sub-sequences and each sub-
r(t ) = As(t ) + n(t ) (2) sequence contains L elements, as shown in Fig. 2. The LP prediction
Rˆ(n ) for R(n) has the following form:
with the following notation definitions: r(t ) =
T
[r0 (t ) r1 (t ) . . . rN−1 (t )] is the (N × 1) received signal vec- 
L
Rˆ(n ) = ωl R ( n + l ), n = 0, . . . , N − L − 1 (6)
tor, the superscript T denotes the transpose operation;
l=1
A = [a(θ1 ) a(θ2 ) . . . a(θK )] is the (N × K) directional matrix;
a(θk ) = [1 exp( j2π λd sin θk ) . . . exp( j2(N − 1 )π λd sin θk )]T is the where ωl (l = 1, . . . , L ) is the prediction coefficient. Therefore, we
T can obtain the denoised element R(0) from the rest of the ele-
directional vector; s(t ) = [s1 (t ) s2 (t ) . . . sK (t )] is the (K × 1)
T ments of the first column with the predictions coefficients. The
source vector; n(t ) = [n0 (t ) n1 (t ) . . . nN−1 (t )] is the (N × 1)
prediction coefficients can be estimated using the sub-sequences
noise vector, with zero mean and covariance matrix σ 2 I, I is the
r2 , r3 , . . . rN−L . In this case, the LP expression can be formulated
(N × N) identity matrix.
as
According to signal model (2) and assuming that the noise is ⎡ ⎤⎡ ⎤
independent from the signals, the covariance matrix R can be writ- R (2 ) R (3 ) ... R (L + 1 ) ω1
ten as ⎢ .. .. .. ⎥ ⎢ .. ⎥
  ⎢ . . . ⎥⎢ . ⎥
⎣R(N − L ) R (N − L + 1 ) R (N − 1 )
⎦ ⎣ω ⎦
R = E r(t )rH (t ) = AAH + σ 2 I (3) ... L

where E{:} denotes the expectation,  = E {s(t )sH (t )}

is the (K × K)      
X ω
dimensional covariance matrix of the source vector s(t) and the su- ⎡ ⎤
perscript H denotes the conjugate transpose operation. Therefore, R (1 )
the elements of R can be expressed as ⎢ .. ⎥
=⎢ . ⎥
⎣R(N − L − 1 )⎦ . (7)
R(m − n ) = R(m, n )
⎧ K
⎪    
⎪
⎪ Psk + σ 2 m=n
⎨ y
k=1
=   (4) Since X and y are known, the coefficient vector ω can be estimated
⎪
⎪ K
d
⎪
⎩ Psk exp j2π (m − n ) sin θk m = n by the least-squares approach:
k=1
λ ω = (XH X )−1 XH y = X+ y. (8)
where m, n ∈ [1, . . . N], Psk is the power of the kth signal. As shown where X+
is the Moore–Penrose inverse of X. One should notice
in (4), the noise is presented only in the diagonal elements of R, that ω has a unique solution only when N − L − 1 ≥ L. Therefore, in
 M. Sun et al. / Signal Processing 156 (2019) 41–45 43

Fig. 2. Set of the overlapping sub-sequences of the first line of R.

Fig. 3. Case a, pseudo-spectrums of the OPM-LP, conventional OPM and OPM-MPEN for DoA estimation, the two DoAs are θ1 = −5◦ and θ2 = 4◦ .

Fig. 4. Case b, pseudo-spectrums of the OPM-LP, conventional OPM and OPM-MPEN for DoA estimation, the two DoAs are θ1 = −1◦ and θ2 = 2◦ .
44 M. Sun et al. / Signal Processing 156 (2019) 41–45

Fig. 5. Case b, RMSE of DoA estimation versus SNR.

Fig. 6. RMSE of DoA estimation in terms of angle separation.

Fig. 7. RMSE of DoA estimation as a function of L.

1 4. Simulations and discussion

order to resolve K sources, the constraint for L is K ≤ L ≤ ( N − 1 ).
2
Then, we can calculate Rˆ(0 ) with the estimated ω: In this section, the performance of OPM-LP is tested on 4 dif-
ferent numerical simulations. A ULA with 15 isotropic elements is

L considered, along with two far-field uncorrelated narrow-band sig-
Rˆ(0 ) = ωl R ( l ). (9) nals with equal power, that is N = 15, K = 2. The number of ele-
l=1 ments in each sub-sequence L is equal to 6. The distance between
two adjacent elements is half of the wavelength.
Since the noise is removed, OPM [7] can be directly applied on the In the first experiment, the pseudo-spectrums of the proposed
denoised data covariance matrix for DoA estimation. method, conventional OPM and OPM-MPEN are estimated. Two
 M. Sun et al. / Signal Processing 156 (2019) 41–45 45

cases are taken into account with different DoAs of the incoming the angle separation increases. Similar to the second simulation,
signals: Case a. θ1 = −5◦ and θ2 = 4◦ ; Case b. θ1 = −1◦ and θ2 = 2◦ . the proposed method offers better performance than that of OPM-
SNR is fixed at −5 dB. The data covariance matrix is estimated MPEN, OPM-KLT and OPM-EVD.
from 10 0 0 independent snapshots. The pseudo-spectrum search In the last simulation, the performance of the proposed method
is performed over [−20◦ , 20◦ ] with step size 0.001°. Figs. 3 and versus the length of sub-sequence L is studied with 500 indepen-
4 show the pseudo-spectrums of the proposed method and the dent runs. The angles in Case b are applied, SNR = −5 dB. L ∈ [2, 7]
conventional OPM and OPM-MPEN for Cases a and b, respectively. 1
due to K ≤ L ≤ (N − 1 ). It can be seen from Fig. 7, when L = 5,
For Case a, both the proposed method and OPM-MPEN can detect 2
the RMSE reaches its minimum. Therefore, in this situation, L = 5 is
the true DOAs of the incoming signals, as shown in Fig. 3. The re-
optimal value for DoA estimation. However, the RMSE shows simi-
sult of the conventional OPM is biased. When the angle separation
lar value with different L.
becomes small, in Case b, the conventional OPM fails to estimate
the DoAs while OPM-MPEN estimates the DoA with bias. However,
5. Conclusion
the proposed method remains accurate, even when the angle sep-
aration is small.
This paper proposes a new method OPM-LP for DoA estimation
In the second simulation, the performance of the proposed
in low SNR scenarios. This new technique utilizes the principle of
method versus SNR is assessed with a Monte-Carlo process of 500
LP, which offers a new way to eliminate the noise impact. Simula-
independent runs. The Root Mean Square Error (RMSE) of the esti-
tion results show the stability, robustness and efficiency of the pro-
mated DoA is defined as follows:
 posed method. In perspective, the proposed method will be tested
 K  2
1 
J
 with real measurements.
RMSE =  θˆk j − θk (10)
KJ References
j=1 k=1

[1] H. Krim, M. Viberg, Two decades of array signal processing research: the para-
where θˆk j denotes the estimated DoA of the kth incoming signal metric approach, IEEE Signal Process. Mag. 13 (4) (1996) 67–94.
for the jth run, and θ k the true value. SNR varies from −10 to 5 dB. [2] T.E. Tuncer, B. Friedlander, Practical aspects of design and application of direc-
The angles in Case b are applied here. The proposed method is tion-finding systems, in: Classical and Modern Direction-of-Arrival Estimation,
Elsevier, 2009, pp. 53–92.
compared with OPM-MPEN, OPM-EVD [8], OPM-KLT [10], and also
[3] R.O. Schmidt, Multiple emitter location and signal parameter estimation, IEEE
Cramér–Rao Bound (CRB). Fig. 5 plots the RMSE of DoA estimation Trans. Antennas Propag. 34 (3) (1986) 276–280.
with respect to SNR. For all the compared methods, the RMSE is [4] R. Roy, T. Kailath, ESPRIT-estimation of signal parameters via rotational in-
variance techniques, IEEE Trans. Acoust. Speech Signal Process. 37 (7) (1989)
continuously decreasing when SNR increases. It can be seen from
984–995.
Fig. 5 that the RMSE of the proposed method is smaller than that [5] J. Munier, G.Y. Delisle, Spatial analysis using new properties of the cross-spec-
of OPM-MPEN and OPM-EVD, especially in low SNR scenarios. Al- tral matrix, IEEE Trans. Signal Process. 39 (3) (1991) 746–749.
though OPM-KLT has smaller RMSE than the proposed method at [6] S. Marcos, J. Sanchez-Araujo, Méthodes linéaires haute résolution pour
l’estimation de directions d’arrivée de sources. performances asymptotiques et
lower SNR, the proposed method has a more significant decrease of complexité, Traitement du Signal 14 (2) (1997) 99–116.
RMSE as SNR increases. Moreover, the computational complexity of [7] S. Marcos, A. Marsal, M. Benidir, The propagator method for source bearing
the proposed denoising procedure (LP) is O(L2 (N − L − 1 )), which estimation, Signal Process. 42 (2) (1995) 121–138.
[8] P. Stoica, T. Söderström, V. Šimonyt, On estimating the noise power in array
is similar to MPEN (O(N 2 (K + 1 ) + NK 2 )) and is much smaller than processing, Signal Process. 26 (2) (1992) 205–220.
that of EVD and KLT (O(N3 )). [9] A. Mariani, A. Giorgetti, M. Chiani, Effects of noise power estimation on energy
In the third simulation, the statistical performance of the pro- detection for cognitive radio applications, IEEE Trans. Commun. 59 (12) (2011)
3410–3420.
posed method versus the angle separation between two incoming [10] J. Li, C.L. Bastard, Y. Wang, G. Wei, B. Ma, M. Sun, Enhanced GPR signal for
signals is assessed with 500 independent runs of the algorithm. layered media time-delay estimation in low-SNR scenario, IEEE Geosci. Remote
One of the incoming signals is fixed at θ1 = 0◦ while the other Sens. Lett. 13 (3) (2016) 299–303.
comes from θ2 = θ1 + θ . SNR is fixed at −5 dB and θ varies [11] S.L. Marple, S.L. Marple, Digital Spectral Analysis: With Applications, 5, Pren-
tice-Hall, Englewood Cliffs, NJ, 1987.
from 1° to 10°. As shown in Fig. 6, when the angle separation is
small (1° to 2°), all methods fail to detect the true DoAs of the in-
coming signals. Moreover, the RMSE continuously decreases when