DOA Estimation in Compressed Sensing Framework by Adaptively Exploiting Prior
Support Information
Jiao Yang, Zhi Zheng, Bo Yan, Yan Ge, and Kehong Liu
School of Communication and Information Engineering
University of Electronic Science and Technology of China
Chengdu, China
Email: zz@uestc.edu.cn
Abstract—In this paper, we propose a novel greedy pursuit
algorithm for direction-of-arrival (DOA) estimation of multiple
moving sources with time-variant DOAs. A prior support set
of DOAs and a metric information are assumed to be available.
Different from the traditional greedy pursuit algorithms which
don't exploit prior support but select entries over the entire
signal index set, or exploit the prior statically, we adaptively
exploit the previously estimated prior support set by firstly
selecting directly entries from the prior support based on the
metric information. The exact number and values of
unchanged DOAs are not required. Simulation results prove
that the proposed method obtains more accurate estimates
even for a small number of snapshots and low SNR.
Keywords-Direction-of-arrival (DOA) estimation; compress-
ed sensing; prior support; adaptive algorithm.
I. INTRODUCTION
Direction-of-arrival (DOA)estimation of multiple sources
plays a fundamental role in array signal processing due to its
wide applications in target tracking, microphone array
systems, mobile communication systems and smart antennas
[1]. Classical DOA estimation algorithms can be divided into
two categories, namely, nonparametric and parametric.
Beamforming, Capon and some subspace-based methods
such as MUSIC [2] and ESPRIT [3], are all nonparametric
methods. Maximum likelihood (DML) and stochastic
maximum likelihood (SML) belong to parametric methods.
Generally, parametric methods can obtain higher perfor-
mance with higher computational complexity. All these
methods rely on statistical properties of the received data and
a sufficiently large number of samples are required for an
accurate estimation.
Compressed sensing theory [4-6] provides a rather
different framework for DOA estimation by formulating it as
a sparse signal recovery problem, In doing so, the DOAs of
sources can be estimated by less samples even a single
sample, especially the family of greedy pursuit algorithms
[7-9]. OMP (orthogonal matching pursuit) in [7] and SOMP
(Simultaneous Orthogonal Matching Pursuit) in [8] both
select only one index to augment the signal support at each
iteration, however, SP (subspace pursuit) in [9] choses
multiple indexes at each iteration. Encouraged by the
potential power of compressed sensing, a variety of
algorithms have been proposed to estimate DOA . Authors of
[10] and [11] utilized convex optimization method to solve
the problem, which has higher computational complexity
than greedy pursuit algorithms. In [12] and [13], single
snapshot estimation was achieved, namely, we can just
utilize one single sample to estimate source DOA.
However, most of the greedy algorithms are designed to
recovery original signals blindly without prior information or
exploit prior support statically. In some practical applications,
sparse structure of signals is correlated with time. For
instance, target tracking, with time-variant DOAs. Consider
multiple moving sources, for the case that all the source
directions don't change synchronously, some source DOAs
will maintain their values during two estimations, namely,
some signals will have fixed DOAs during two estimation
intervals.
In this paper, we propose a modified OMP (orthogonal
matching pursuit) method to adaptively exploit the prior
signal support, which has been estimated previously to
improve the recovery performance. A metric information
is introduced to measure the similarity between the
previously estimated support
and the current support
.
And the assumption that our proposed method used is similar
to MSP [14] method proposed by Rao for channel estimation
of massive MIMO with temporal correlation. In simulations,
we apply MSP to estimate DOA for comparison. Simulation
results demonstrate that the proposed method acquires higher
estimation accuracy compared to the existing methods.
II. COMPRESSED SENSING OVERVIEW
A signal vector
is said to be k- sparse, if
contains k nonzero entries at most with
. And a set
of indices
is the support of
, if the
nonzero entries of
are indexed by the entries in
,
which denotes the cardinality of
and equals to
,
where
is
pseudo-norm calculating the number of
nonzero entries. Compressed sensing theory proves that a N-
dimension sparse signal
can be recovered from the M-
dimension measurements
via the following formul-
ation:
0
ˆ
argmin subject to
N
x
xx Φxy
(1)
where
is called measurement matrix.
The model in (1) is called SMV (single measurement vector)