Gridless Postprocessing for Sparse Signal
Reconstruction based DOA Estimation
Xiaohuan Wu
∗
, Wei-Ping Zhu
∗†
and Jun Yan
∗
∗
Institute of Signal Processing and Transmission,
Nanjing University of Posts and Telecommunications, Nanjing, China
Email: {2013010101, zwp, yanj}@njupt.edu.cn
†
Department of Electrical and Computer Engineering,
Concordia University, Montreal, Canada
E-mail: weiping@ece.concordia.ca
Abstract—Recently, many sparse signal reconstruction (SSR)
based methods have been proposed for direction-of-arrival
(DOA) estimation. However, these methods often suffer from the
off-grid problem caused by the discretization of the potential
angle space. Most of them employ iterative grid refinement
(IGR) method to alleviate this problem. However, IGR requires a
high computational load and may not comply with the restricted
isometry property (RIP) condition. In this paper, we propose a
novel postprocessing scheme named as gridless postprocessing
(GPP) for the SSR-based DOA estimation. GPP solves a convex
optimization problem with an alternate procedure to obtain
the bias estimate. To accelerate the convergence, a closed-form
expression is derived for the bias estimation. The proposed
scheme enjoys much smaller computational load than IGR while
provides comparable performance. Furthermore, by avoiding
further dividing the grids, the GPP is superior to IGR in the
correlated signal scenario. Simulations are carried out to verify
the performance of our proposed method.
Index Terms—Direction-of-arrival (DOA) estimation, sparse
signal representation (SSR), iterative grid refinement (IGR).
I. INTRODUCTION
Direction-of-arrival (DOA) estimation with sensor arrays
has been an active topic in array signal processing over the
past three decades. Various applications like localization or
tracking of the transmitting sources, channel estimation, and
wireless communications require the estimation of the DOAs
of the interested signals. A large number of subspace-based
algorithms for DOA estimation have been proposed in the
literature such as [1]–[3]. These powerful algorithms show
a better resolution as compared to beamforming technique
exploiting the orthogonality between the signal subspaces
and the noise subspaces. However, the accuracies of these
subspace-based algorithms are highly dependent on the esti-
mation of the covariance matrix of the array output. Hence,
their performances deteriorate sharply in the cases of low
signal-to-noise-ratio (SNR), limited snapshots and coherent
signals. The maximum likelihood (ML) method owns the best
asymptotic performance [4] and enjoys the excellent adaptive
ability to the aforementioned scenarios, but suffers from
heavy computational complexity due to multi-dimensional
searching. In addition, the methods based on subspace and
ML require the number of incident signals as a priori to
achieve a satisfying performance. A false number can result
in the failure of DOA estimation or spurious spectral peaks.
Recently, the sparse signal reconstruction (SSR) technique
has been introduced for the first time to DOA estimation by
Malioutov et al. [5] and soon adopted by many researchers
to develop a number of DOA estimation methods [6]–[9].
These SSR-based estimators are often immune to the lack of
prior signal number imformation and able to keep the merits
of the subspace-based ones. One typical assumption of the
SSR-based estimators is that the true DOAs of the sources lie
exactly on the predefined finite discrete grids. Since the true
DOAs of the signals belong to the continuous angle space and
cannot be represented by a finite number of grids precisely,
this assumption fails in practice. To alleviate this problem,
most SSR-based algorithms use an iterative grid refinement
(IGR) procedure [5]. The IGR begins with an initial coarse
grid and uses an SSR-based algorithm to obtain first a coarse
DOA estimate. It then further divides the grid containing the
coarse DOA estimate into a finer grid and repeat the SSR-
based algorithm to achieve a better DOA solution. Since
the SSR-based DOA estimation algorithm is called in each
iteration of IGR, the computational load of this method is very
high. Moreover, a too dense grid would contradict the use of
restricted isometry property (RIP) condition of compressive
sensing (CS) [10], and hence may not further improve the
estimation performance. Therefore, the off-grid model has
been studied recently in [11]–[13], which introduces a bias
parameter to modify the SSR-based signal model in order to
avoid an iterative grid refinement.
In this paper, we propose a novel postprocessing method
named as gridless postprocessing (GPP) to obtain a fine DOA
estimate. After obtaining the coarse DOA, we formulate the
bias estimation problem as an optimization problem, which is
then solved by using a linear approximation and a simple up-
dating procedure. To accelerate the convergence, we derive a
closed-form solution for the bias estimation, thus avoiding the
use of time-consuming convex optimization toolboxes such as
CVX or SeDuMi. Without requiring a further refinement of
the grid, GPP is immune to the correlation of the columns
of the manifold dictionary, as opposed to the pure SSR-based
DOA algorithm where the dense grid refinement contradicts
684978-1-4799-8058-1/15/$31.00 ©2015 IEEE.