International Journal of Future Computer and Communication, Vol. 2, No.

6, December 2013

Comparison of Direction of Arrival (DOA) Estimation

Techniques for Closely Spaced Targets
Nauman Anwar Baig and Mohammad Bilal Malik

x(n) = A(φ )s(n) + n(n) (1)

Abstract—This paper deals with different beamforming
techniques for DOA estimation. High resolution techniques
Here, A(φ ) ( N × M ) is the steering matrix where N is the
such as Multiple Signal Classification (MUSIC) and Estimation
of Signal Parameters via Rotational Invariance Techniques number of sensors and M is the number of sources. s(n) is the
(ESPRIT) which are subspace based techniques are also
signals vector and n(n) is additive white Gaussian noise [2].
discussed. Finally, further resolution improvement is achieved
through the use of directional sensors. The computational The number of snapshots of the signals are K, n=1, 2, 3,….,K.
complexities of beamforming techniques are also compared.
y Source

Index Terms—Beamforming, computations,

direction-of-arrival, multiple, radar, resolution.

I. INTRODUCTION
Beamforming uses an array of receivers (sensors) for
directional transmission/reception. The main goal for x
formation of an array and array processing is to combine the
sensors outputs so that the SNR can be enhanced, Fig. 1. Multiple signals incident on ULA
information about the number of sources/targets and
x1 (n)
direction of each can be determined, various parameters of w1
the incident signals can be estimated. x2 (n)
The main requirement in many source localization w2 y (n)
∑
applications e.g. Radar, Sonar is to estimate the direction of x3 (n)
w3
arrival without errors. When two sources have a small
angular distance between them in space, angular resolution is x4 (n)
w4
an important area to concern about; otherwise some
sources/targets would not be detected. The objective of Fig. 1. Beamformer
paper is to discuss DOA estimation algorithms with focus on It is also assumed that the signal and noise both are zero
the enhancement of angular resolution. The requirement is mean, noise variance at each of the sensors is σ 2 . The
achieve a higher resolution with minimum number of
sample correlation matrix is given by
computations.
Propagating signals contain much information about the ˆ = 1
K

sources that produce them. Not only does each signal’s

R
K
∑ x( n) x
n =1
H
( n) (N × N ) (2)
waveform express the nature of the source, its temporal and
spatial characteristics combined with the laws of physics
allow us to determine the source’s location [1]. For III. BEAMFORMING TECHNIQUES
propagating signals, more is needed; spatiotemporal filtering Beamforming techniques can be divided into spectral
must be employed to separate signals according to their estimation techniques and subspace based methods [3]. In
directions of propagation and their frequency content. spectral estimation, a spectrum-like function of the parameter
of interest e.g. the DOA is formed. The locations of the
highest (separated) peaks of the function are the DOA
II. MODEL FOR INCIDENT WAVES estimates. The idea is to steer the array in one direction at a
A uniform linear array (ULA) is used for beamforming. time and measure the output power. The steering locations
The sources/targets are assumed to be in the far-field region which result in maximum power yield the DOA estimates.
so that the waves coming from them to the ULA can be The beamformer’s output power is [3]
considered to be plane waves. The data from the ULA is ˆ
P (w ) = w H Rw (3)

A. Conventional Delay and Sum Beamformer (CBF)

Manuscript received October 29, 2012; revised February 15, 2013.
The authors are with Department of Electrical Engineering, College of It maximizes the power of the beamformer output for a
E&ME, National University of Sciences and Technology, Islamabad, given input signal. For CBF, weight vector is the steering
Pakistan (e-mail: [email protected], [email protected])

DOI: 10.7763/IJFCC.2013.V2.246 654

 International Journal of Future Computer and Communication, Vol. 2, No. 6, December 2013

vector. For different angles, the output power is measured [3] 1

CBF
ˆ (φ )
P (φ ) = a (φ )Ra
H
(4) 0.9 MVDR

0.8

where the steering vector 0.7

Normalized Spectrum
0.6
T
a(φ ) = g (φ ) ⎡⎣1 e − jkd sin φ e − jk 2 d sin φ ...... e − jk ( N −1) d sin φ ⎤⎦ (5) 0.5

0.4
And g (φ ) = 1 , for omnidirectional sensors. Following plot
0.3
shows the CBF output power when 2 sources are present
0.2
at φ1 = −30o and φ1 = 20o
0.1
Conventional Beamformer
1 0
-100 -80 -60 -40 -20 0 20 40 60 80 100
0.9 DOA(deg)

0.8 Fig. 5. N=4, Sources at 0o and 15o

0.7
Normalized Spectrum

0.6
From the above figure, it can be seen that the resolution of
0.5
MVDR is much better than that of CBF.
0.4
Now the high resolution subspace based methods are
0.3
discussed which are based on the decomposition of
correlation matrix into signal and noise subspaces.
0.2

0.1 R = APA H + σ 2 I = U s Λ s U sH + U n Λ n U nH (7)

 
 

0
-100 -80 -60 -40 -20 0 20 40 60 80 100
signal subspace noise subspace
DOA(deg)

Fig. 3. N=4 Sensors, SNR=10dB and K=200 samples=200 C. Multiple Signal Classification (MUSIC)
This algorithm uses the fact that all noise eigenvectors are
A limitation of the CBF is that it cannot resolve 2 targets
orthogonal to the signal steering vectors [5]
within the beamwidth. Consider 2 sources at 0o and 20o
Conventional Beamformer
U nH a(φ ) = 0, φ ∈{φ1 , φ2 ,...., φM } (8)
1

0.9
The output spectrum for MUSIC beamformer is
0.8 1 1
P (φ ) = = (9)
0.7
U a(φ )
H 2
a (φ )U n U nH a(φ )
H
Normalized Spectrum

n
0.6

0.5 In the output spectrum, M largest peaks correspond to

0.4
DOAs
MUSIC Beamformer
0.3 1

0.2 0.9

0.1
0.8

0
-100 -80 -60 -40 -20 0 20 40 60 80 100 0.7
DOA(deg)
Normalized Spectrum

0.6
Fig. 4. Can’t resolve targets inside beamwidth of 4 sensors i.e. 30o
0.5

One way to increase the resolution is to increase the 0.4

number of sensors, which reduces the beamwidth. But it also 0.3
increases the cost of the beamformer.
0.2
B. Minimum Variance Distortionless Response
0.1
Beamformer (MVDR)
0
To reduce the limitations of conventional beamformer, -100 -80 -60 -40 -20 0 20 40 60 80 100
DOA(deg)
such as to increase the resolving power of two sources spaced
closer than a beamwidth, a method was proposed by Capon Fig. 6. N=4, 2 sources separated by 5o
[4]. The power spectrum is
The above plot shows that using 4 sensors, even two
1 closely spaced targets can be resolved using MUSIC
P (φ ) = H (6)
ˆ −1a(φ )
a (φ )R beamformer. When the number of snapshots (available data
is small), a modified version of MUSIC known as
Following plot shows the comparison of resolutions of Root-MUSIC proves to be useful. It is based on polynomial
CBF and MVDR rooting.

655
 International Journal of Future Computer and Communication, Vol. 2, No. 6, December 2013

D. Root-MUSIC IV. COMPARISON OF BEAMFORMING ALGORITHMS

The Root ‐MUSIC method converts the MUSIC spectrum Following figure shows the resolving power of different
into a polynomial whose solution results directly in numeric beamformers discussed
values for the estimated directions [6]. From the above plot, it can be seen that the ESPRIT is a
E. Estimation of Signal Parameters via Rotational very high resolution algorithm.
Invariance Technique (ESPRIT) In the following plot, the accuracy of beamformers for
This algorithm is based on translational invariance different SNR values are compared. The RMSE is given by
structure (e.g.ULA) of sensors [7]. ESPRIT is
computationally more efficient as it doesn’t require an E (φ1 − φˆ1 ) 2 + E (φ2 − φˆ2 ) 2
exhaustive search through all possible steering vectors for RMSE = (11)
DOA estimation. Following figure shows the pair of 2
subarrays used in ESPRIT
1.4
MVDR
MUSIC
1.2 RMUSIC
ESPRIT

1
Δ

Normalized RMSE
0.8

0.6
Fig. 7. Two identical subarrays

0.4
The DOA estimates are given by
0.2
⎡ arg(Φ k ) ⎤
φk = sin −1 ⎢ ⎥ (10)
⎣ βd ⎦ 0
0 5 10 15 20 25
SNR
where φk is the estimated DOA, Φ k is the kth eigen value of
Fig. 10. N=4, separation=10o, 100 trials and 200 samples
subspace rotational operator [7], β = 2π λ ( wavelength) and
d is the spacing between the sensors. The following plot It can be observed that ESPRIT is the most accurate for
shows the output of ESPRIT beamformer when sources were low SNR conditions.
present at -2o, 1o and 4o.
ESPRIT TABLE I: SUMMARY OF BEAMFORMING ALGORITHMS
1

0.9 Algorithm Resolution Complexity General

0.8
Remarks
CBF Poor Simple Resolution
Normalized Spectrum

0.7

0.6
Implementation, depends on main
0.5
1-D search lobe
0.4 MVDR Good Inverse of R, 1-D Poor
0.3 search performance in
0.2
low SNR
0.1 MUSIC Very Good Eigenvalue Also estimates
0
Decomposition, number of
-10 -8 -6 -4 -2 0 2 4 6 8 10
DOA(deg) 1-D search sources [5]
Fig. 8. High resolution ESPRIT beamformer ESPRIT Excellent Eigenvalue Array needs
Decomposition, doublets
1 Calculating ψ
MUSIC
MUSIC
0.9 Capon
MVDR
CBF
Conventional
0.8
ESPRIT
0.7 V. COMPUTATIONAL COMPLEXITIES
Normalized Spectrum

0.6
The following tables show the number of multiplications
0.5 and additions required for each of the algorithm. The
0.4 following symbols are used (Table I-Table VI)
0.3

0.2 N = number of sensors

0.1 M = number of signals
0
-20 -10 0 10 20 30 40 K = number of samples
DOA(deg)
L = number of angles to scan
Fig. 9. Resolution comparison of BEAMFORMERS

656
 International Journal of Future Computer and Communication, Vol. 2, No. 6, December 2013

TABLE I: COMPUTATION OF CORRELATION MATRIX

Operation Multiplications Additions Divisions
1 K
⎛N N⎞ 2
( K − 1) N 2 N2
R N×N =
K
∑ x( n) N ×1 x H (n)1× N K⎜ + ⎟
n =1
⎝ 2 2⎠

TABLE II: COMPUTATIONS OF CBF

Operation Multiplications Additions Divisions
P (φ )1×1 = a (φ )1× N R N × N a(φ ) N ×1
H 2
L( N + N ) 2
L( N − 1) −

TABLE III: COMPUTATIONS OF MVDR

Operation Multiplications Additions Divisions
−1 3 2 3 2
R N×N (Gauss Jordan Inversion) [8] N N N N N N N2 N
+ + + + +
3 2 6 3 2 6 2 2
1 L( N 2 + N ) L( N 2 − 1) L
P(φ )1×1 =
a H (φ ) ˆ −1 a(φ )
R
1× N N×N N ×1

TABLE IV: COMPUTATIONS OF MUSIC ALGORITHM

Operation Multiplications Additions Divisions
R N ×N = U N ×N Λ N ×N U H N ×N 16 3 44 3 −
N N
[9], [10] 5 5

Q N × N = U n[ N ×( N − M )] U nH [( N − M )× N ] (N − M )N 2 N 2 −
(N − M )N 2 (N − M )N
+ − +
2 2 2 2
(N − M )N N
−
2 2

1 L( N 2 + N ) L( N 2 − 1) L
P (φ )1×1 =
a (φ )1× N Q N × N a(φ ) N ×1
H

TABLE V: COMPUTATIONS OF ESPRIT ALGORITHM

Operation Multiplications Additions
R N ×N = U N ×N Λ N ×N U H N ×N 16 3 44 3
N N
5 5
A M × M = U sH1[ M ×( N −1)] U s1[( N −1)× M ] ( N − M )( N − 1) 2 ( N − M )( N − 1) 2 ( N − 1) 2
+ − +
2 2 2
( N − M )( N − 1) ( N − M )( N − 1) ( N − 1)
−
2 2 2
B M × M = A −M1× M (Gauss Jordan 3 3 2 Divisions
N N2 N N N
+ + +
Inversion) [8] 3 2 6 3 2 N2 N
+
N 2 2
+
6
CM × M = U sH1[ M ×( N −1)] U s 2[( N −1)× M ] ( N − M ) 2 ( N − 1) ( N − M ) 2 ( N − 1) ( N − M ) 2
− +
2 2 2
( N − M )( N − 1) ( N − M )( N − 1) ( N − M )
+ −
2 2 2
ψ M ×M = B M ×M CM ×M ( N − M )3 ( N − M )3 − ( N − M ) 2

The array patterns are compared below where 4 sensors are

VI. ENHANCEMENT IN RESOLUTION USING DIRECTIONAL used
SENSORS
TABLE VI: FOR FIXED N, M, K AND L
When directional sensors are used instead of Computational complexity decreases downward
omnidirectional sensors, the half power beamwidth of the MVDR
array’s response reduces which increases the resolution of CBF
the beamformer. MUSIC
ESPRIT
Consider an array of 4 sensors each having a linear
aperture of length D (D<d)

657
 International Journal of Future Computer and Communication, Vol. 2, No. 6, December 2013

shown below
MUSIC
1
Directional
0.9 Omnidirectional
0.8

0.7

Normalized Spectrum
Array pattern 0.6
7
Directional 0.5
Omnidirectional
6 0.4

0.3
5
0.2
4
Magnitude

0.1

3 0
-100 -80 -60 -40 -20 0 20 40 60 80 100
DOA(deg)
2
Fig. 14. MUSIC beamformer

1
VII. CONCLUSIONS
0
-100 -80 -60 -40 -20 0 20 40 60 80 100 In this paper different algorithms for estimation of
Degrees
direction of arrival are discussed. The main focus is to
Fig. 11. Array patterns of directional and omnidirectional sensor array increase the resolution so that closely spaced targets in space
1
CBF
can be separated. The conventional beamformer has a
Omni
Directional
resolution limitation due to beamwidth. It has been shown
0.9
that beamwidth can be reduced by increasing the number of
0.8
sensors but it also inceases the cost of beamformer. Adaptive
0.7 beamforming algorithms have the advantage of much better
0.6 resolution. The minimum variance distortionless response
beamformer has a relatively higher resolution due to the
magnitude

0.5
output power minimization subject to the constraint.
0.4
Subspace methods for estimation of DOA are based on the
0.3 signal and noise subspaces. MUSIC algorithm which shows
0.2 high peaks for angles corresponding to DOAs, has a much
0.1
higher resolution. It has also been shown that using
directional sensors instead of omnidirectional sensors gives
0
-100 -80 -60 -40 -20 0 20 40 60 80 100 the advantages of a relatively reduced beamwidth and higher
DOA(deg)
gain. The tables of computational complexities show that
Fig. 12. CBF for directional array is better ESPRIT is computationally much efficient as it does not
1
MVDR Beamformer
require a scan through all possible angles.
Omni
0.9 Directional
REFERENCES
0.8
[1] D. E. Dudgeon and D. H. Johnson, Array Signal Processing: Concepts
0.7 and Techniques, Prentice Hall, Englewood Cliffs, NJ, 1993.
[2] B. D. van Veen and K. M. Buckley, “Beamforming: A versatile
Normalized Spectrum

0.6
approach to spatial filtering,” IEEE Signal Process. Mag., vol. 5, no. 2,
0.5
pp. 4–24, Apr. 1988.
[3] H. Karim and M. Viberg, “Two decades of array signal processing
0.4 research,” IEEE Signal Process. Mag., pp. 67-94, July 1996.
[4] J. Capon, “High-resolution frequency-wavenumber spectrum
0.3
analysis,” IEEE, vol. 57, pp. 1408-1418, 1969.
0.2 [5] R. O. Schmidt, “Multiple emitter location and signal parameter
estimation,” IEEE Trans. antennas Propagat., vol. AP-34, pp. 276-280,
0.1 1986.
[6] A. Barabell, “Improving the resolution of eigenstructured based
0
-100 -80 -60 -40 -20 0 20 40 60 80 100 ndirection finding algorithms,” in Proc. ICASSP, pp. 336-339, 1983.
DOA(deg)
[7] R. Roy and T. Kailath, “ESPRIT - estimation of signal parameters via
Fig. 13. Response of 4 directional sensors rotational invariance techniques,” IEEE Trans. Acoust., Speech, Signal
Processing, vol. ASSP-37, pp. 984- 995, 1989.
[8] M. B. Malik and M. Salman, “State-space least mean square,” Digital
The output of CBF is shown below. The output power of Signal Processing, vol. 18, 2008, pp. 334–345.
each of the beamformers with directional and [9] S. B. Himane and D. Zikic, “Singular value decomposition,” Basic
Mathematical Tools for Imaging and Visualization (WT 2006).
omnidirectional sensors is first normalized and then
[10] M. Lee and S. K. Oh, “A Per-User Successive MMSE Precoding
combined Technique in Multiuser MIMO Systems,” in Proc. IEEE 65th
Similarly for MVDR and MUSIC beamformer, the Vehicular Technology Conference, 2007. VTC2007-Spring, pp.
increase in resolution obtained by using directional sensors is 2374-2378.

658
 International Journal of Future Computer and Communication, Vol. 2, No. 6, December 2013

Nauman Anwar Baig received his B.S. degree in electrical engineering Mohammad Bilal Malik received his B.S. degree in electrical engineering
from University of Engineering & Technology, Taxila, Pakistan, in 2008. He from College of Electrical and Mechanical Engineering (E&ME),
received his M.S. degree in electrical engineering from College of Electrical Rawalpindi, Pakistan, in 1991. He received his M.S. degree in Electrical
and Mechanical Engineering (E&ME), Rawalpindi, Pakistan, in 2010. Engineering from Michigan State University (MSU), Michigan, USA, in
Currently he is doing Ph.D. in Electrical Engineering from College of 2001. In 2004, he received his Ph.D. degree in Electrical Engineering from
Electrical and Mechanical Engineering (E&ME), National University of National University of Sciences and Technology (NUST), Rawalpindi,
Sciences and Technology (NUST), Rawalpindi, Pakistan. His research Pakistan. He has been teaching at College of E&ME, National University of
focuses on signal processing and communications. Sciences and Technology (NUST) since 1991. His research focuses on signal
processing and adaptive filtering for communications and control systems.

659