2014 IEEE International Conference on Acoustic, Speech and Signal Processing (ICASSP)

A COMPUTATIONALLY EFFICIENT SOURCE LOCALIZATION METHOD FOR A

MIXTURE OF NEAR-FIELD AND FAR-FIELD NARROWBAND SIGNALS

Weiliang Zuo† , Jingmin Xin†∗ , Jiasong Wang‡ , Nanning Zheng† , and Akira Sano§
†
Institute of Artificial Intelligence and Robotics, Xi’an Jiaotong University, Xi’an 710049, China,
‡
State Key Laboratory of Astronautic Dynamics, Xi’an 710043, China
§
Department of System Design Engineering, Keio University, Yokohama 223-8522, Japan

ABSTRACT for constructing cumulant matrix. A new second-order statis-

In this paper, we consider the source localization for a mixed tics (SOS) based differencing method was developed in [12],
near-field (NF) and far-field (FF) narrowband signals im- where the crux is to eliminate the contribution of the FF sig-
pinging on a uniform linear array (ULA) with the symmet- nals and additive noise for localizing NF signals. Unfortu-
rical geometric configuration. A computationally efficient nately, the structure property of the covariance matrix of inci-
direction-of-arrivals (DOAs) and range estimation method dent signals is required, which is only valid for large number
for the mixed NF and FF signals is proposed, where the of snapshots. By utilizing the advantage of a symmetric uni-
DOAs of the NF and FF signals are estimated separately, form linear array (ULA), some SOS-based methods [13, 14]
and the computationally burdensome eigendecomposition is were proposed to localize the mixed NF and FF signals. These
avoided. Comparing to some existent methods, the proposed methods have low computational burden and use a subjective
method can separate the NF signals from the FF signals more criterion, which distinguishes the NF and FF signals by DOAs
efficiently, and consequently the estimation performance is or ranges estimation, to determine the type (i.e., NF or FF) of
improved. The effectiveness of the proposed method is veri- the incident signals. Additionally, the oblique projection [15],
fied though numerical examples. which projects the measurement onto a low-rank subspace a-
long a non-orthogonal subspace, was used to separate the in-
Index Terms— Source localization, far-field, near-field, cident signals [13]. Although it outperforms the differenc-
uniform linear array, direction-of-arrival. ing one, it possesses a “saturation behavior” for localizing
the NF signals. Moreover these SOS-based methods afore-
1. INTRODUCTION mentioned require a computationally intensive procedure of
The source localization is a fundamental problem in radar, eigendecomposition.
sonar, wireless communication, seismic exploration and so Therefore, we propose a more computationally efficient
on (e.g., [1] and references therein), and many algorithm- method for localizing the mixed NF and FF incident signals
s have been developed to deal with either the far-field (FF) impinging on the ULA with the symmetrical geometric con-
signals [2, 3] or the near-field (NF) signals [4, 5], respective- figuration. By taking the advantage of the oblique projector, a
ly. In many application scenarios of sources localization such new covariance matrix which only contains the information of
as speaker localization using microphone arrays [6,7], the NF the NF signals is formed, then the anti-diagonal elements of
and FF signals may coexist. Hence the source localization and this resulting matrix are used to estimate the DOA of the NF
classification of the mixed NF and FF signals have received signals. Although the oblique projector was used for separat-
considerable attention recently. ing the mixed signals in [13], in this paper, we present a more
By utilizing the properties of the higher-order statistics efficient way to calculate it. Furthermore in order to over-
(HOS) (i.e., cumulant), and some methods were develope- come the “saturation behavior” encountered by the differenc-
d to localize the mixed NF and FF of non-Gaussian signal- ing and the oblique projection based methods for localizing
s [8–11]. In these methods, the main step is to form a special the NF signals, an alternating iterative method is developed.
cumulant matrix, which only contains the direction-of-arrival Finally, the effectiveness of the proposed method is verified
(DOA) information of mixed signals, and from this result- through some numerical examples.
ing cumulant matrix, the DOA estimates are obtained firstly.
Then the range estimation is easily obtained, where an addi-
tional DOA association procedure is needed. However, the 2. DATA MODEL AND BASIC ASSUMPTIONS
HOS-based methods require high computational complexity As shown in Fig. 1, we consider K noncoherent narrow-
This work was supported in part by the National Natural Science Foun- band signals {sk (n)} impinging on the ULA consisting
dation of China under Grant 61172162. of 2M + 1 omnidirectional sensors with spacing d, and

978-1-4799-2893-4/14/$31.00 ©2014 IEEE 2276

 sk
without loss of generality, we assume the first K1 incident
rk
signals {sk (n)}K 1
k=1 are the FF ones with the locations of
K1 θk
{(∞, θk )}k=1 , while the other K2 signals {sk (n)}K k=K1 +1
are the NF ones with the locations of {(rk , θk )}K k=K1 +1 , d d
where K = K1 + K2 . By letting the center of the ULA be
the phase reference point, the received noisy array data can − M −1 0 1 M
be expressed in a vector-matrix form as Fig. 1. The ULA with the symmetrical geometric configura-
x(n) = Af sf (n) + An sn (n) + w(n) tion.
= As(n) + w(n) (1) 3.1. DOA Estimation of FF Signals
where x(n)  [ x−M (n), x−M +1 (n), · · · , xM −1 (n), Firstly, we can divide the array response matrix A in (1) into
xM (n)]T , w(n)  [w−M (n), w−M +1 (n), · · · , wM −1 (n), two submatrices as 
wM (n)]T , sf (n)  [s1 (n), s2 (n), · · · , sK1 (n)]T , sn (n)  
A1 } K
[sK1 +1 (n), sK1 +2 (n), · · · , sK (n)]T , and s(n)  [sf (n)T , A= (3)
A2 }2M +1−K .
sn (n)T ]T , while A is the array response matrix given by A  Under the basic assumptions, we can find that there exist a
[Af , An ], Af  [af (θ1 ), af (θ2 ), · · · , af (θK1 )], An  K ×(2M +1−K) linear operator P between A1 and A2 [3],
[an (rK1 +1 , θK1 +1 ), an (rK1 +2 , θK1 +2 ), · · · , an (rK , θK )], i.e.,
and the response vectors for the FF and NF signals are defined
as af (θk )  [e−jM ωk , · · · , e−jωk , 1, ejωk , · · · , ejM ωk ]T , A 2 = P H A1 (4)
and an (rk , θk )  [ejτ−M k , · · · , ejτ−k , 1, ejτk , · · · , ejτM k ]T , or
where ( · )T denotes transpose. Furthermore the phase de- [P H , −I 2M +1−K ]A = QH A = O (2M +1−K)×K (5)
lay of the FF signals ωk is defined as ωk  −2πd sin θk /λ, where O p×q denotes a p × q null matrix. Then from (5), we
where φk  πd2 cos2 θk /(λrk ), and λ is the wavelength of easily have
incident signals, while by using the so-called Fresnel approx-
QH an (r, θ) = 0(2M +1−K)×1 (6)
imation, the phase delay of the NF signals τmk is given by
τmk  ωk m + φk m2 for m = −M, · · · , −1, 0, 1, · · · , M , QH af (θ) = 0(2M +1−K)×1 . (7)
where φk  πd2 cos2 θk /(λrk ). Secondly, to get the noiseless covariance matrix R̄, we need
In this paper, we make the following assumptions: 1) The to estimate the noise variance. By considering the partition
array is calibrated and the array response matrix A has full K 2M +1−K
rank. 2) The incident signals {sk (n)} are zero-mean wide-  
R= R11 R12 K (8)
sense stationary random processes and are uncorrelated each R21 R22 2M +1−K
other. 3) The additive noises {wi (n)} are temporally and s-
patially complex white Gaussian random process with zero- then the noise variance σ 2 can be obtained by [16]
mean and variance σ 2 and are independent to the incident sig- tr{R22 Π}
σ2 = (9)
nals {sk (n)}. 4) The numbers of the incident NF and FF sig- tr{Π}
nals K1 and K2 are known, and the number of all incident where Π = I 2M +1−K − R21 R†21 , tr{·} denotes the trace
signals K satisfies the relation K < M + 1. operator, and ( · )† denotes the Moore-Penrose pseudoinverse.
Furthermore, we can also divide R̄ into two parts as R̄ =
3. NEW METHOD FOR SOURCE LOCALIZATION [G, H], in which G and H consist of the first K or the last
2M +1−K columns. Thus, when the number of snapshots is
Under the basic assumptions, from (1), we can obtain the ar- finite, the DOAs {θk }K 1
k=1 of the FF signals can be estimated
ray covariance matrix R of the received data as by minimizing the following cost function
R  E{x(n)xH (n)} gf (θ) = aHf (θ)ΠQ̂ af (θ) (10)
= Af Rsf AH H 2
f + An Rsn An + σ I 2M +1 H H H H
2
where ΠQ̂ = Q̂(Q̂ Q̂)−1 Q̂ = Q̂(I 2M +1−K −P̂ (P̂ P̂
= R̄ + σ I 2M +1 (2) H H H T
+I K )−1 P̂ )Q̂ , P̂ = (Ĝ Ĝ)−1 Ĝ Ĥ, and Q̂ = [P̂ ,
where R̄ is the noiseless array covariance matrix defined by
−I 2M +1−K ]T , while ΠQ̂ is calculated using the matrix in-
R̄  ARs AH = R − σ 2 I 2M +1 , Rsf and Rsn are the co-
version lemma implicitly [17].
variance matrix of the FF or NF signals defined by Rsf 
E{sf (n)sH H
f (n)} and Rsn  E{sn (n)sn (n)}, while Rs  3.2. DOA Estimation of NF Signals
E{s(n)sH (n)} = blkdiag(Rsf , Rsn ), E{w(n)wH (n)} =
By using the DOAs of the FF signals, from (2), we have a
σ 2 I 2M +1 , blkdiag( · ) and ( · )H denotes block diagonal ma-
(2M + 1) × (2M + 1) matrix as
trix operator and the Hermitian transpose, and I m is a m × m
identity matrix. Ro  R̄Π⊥ H ⊥
Af = An Rsn An ΠAf (11)

2277
where Π⊥ H
Af  I 2M +1 − Af (Af Af )
−1 H
Af . By employing ˆ (Q̄
where ΠQ̄ˆ = Q̄ ˆ )−1 Q̄
ˆ H Q̄ ˆ H = Q̄
ˆ (I ˆH ˆ ˆH
M +1−K2 −P̄ (P̄ P̄
the QR decomposition of Ro as Ro = Q̃R̃, we can obtain +I )−1 P̄
K2
ˆ )Q̄
ˆ H , P̄
ˆ = (Ḡ ˆ )−1 Ḡ
ˆ H Ḡ ˆ H H̄
ˆ , and Q̄
ˆ = [P̄
ˆ T,
the oblique projector E Af |An in the following manner −I M +1−K2 ]T .
H
E Af |An = Af (Q̃2 Q̃2 Af )† (12)
3.3. Range Estimation of NF Signals
where Q̃2 is the last 2M + 1 − K2 columns of the Q̃ [18]. Once we get the DOA estimates of the NF signals, the ranges
Then from (2) and (12), a array covariances matrix Rn corre- can be found from (6). In the case of the finite number of
sponding to the NF signals can be obtained snapshots, {rk }K k=K1 +1 can be estimated in the Fresnel re-
Rn  An Rsn AH n gion (i.e., rf ⊂ (0.62(D3 /λ)1/2 , 2D2 /λ)) by minimizing the
= (I 2M +1 − E Af |An )R̄(I 2M +1 − E Af |An )H (13) following cost function
where it’s anti-diagonal element rn (i) is given by f¯n (r) = aH
n (r, θ̂)ΠQ̂ an (r, θ̂). (22)
K where D is the aperture of the array [5], and the estimated
rn (i)  (Rn )i,2M +2−i = rs2k e−j2(M +1−i)ωk (14) range {r̂k }K K
k=K1 +1 and the estimated DOAs {θ̂k }k=K1 +1 are
k=K1 +1 automatically paired without any additional procedure.
where i = 1, 2, · · · , 2M + 1, and rs2k is the power of the
3.4. Alternating Iterative Scheme for NF Signals
kth NF signal defined by rs2k  E{sk (n)s∗k (n)}, where ( · )∗ In the DOA estimation of the NF signals, we need to calcu-
denotes the complex conjugate. The element rn (i) only con- late the oblique projector E Af |An , where the block diagonal
tains the DOA information of the NF signals. Then we can structure of signal covariance matrix Rs and the anti-diagonal
construct a new (M + 1) × (M + 1) covariance matrix like elements of Rn are exploited. Unfortunately, when the num-
the FF Toepliz covariance matrix [19] as follows ber of snapshots is finite, Rs may be not block diagonal, and
⎡ ⎤
rn (M + 1), . . . , rn (2), rn (1) the anti-diagonal elements of Rn contain not only the infor-
⎢ rn (M + 2), . . . , rn (3), rn (2) ⎥
⎢ ⎥ mation of the NF signals but also that of the FF signals. Then
R̄n  ⎢ .. . . .. ⎥ (15)
⎣ . . . .
. . ⎦ from (11), we can see that
rn (2M + 1), . . . , rn (M + 2), rn (M + 1) R̂o = An R̂sn AH ⊥ H ⊥
n ΠAf + Af R̂sf n An ΠAf (23)
N H
After some simple algebraic manipulations, it can be rewrit- where R̂sf n = (1/N ) n=1 sf (n)sn (n). It is obvious that
ten in a more compact form as the range space of R̂o is not strictly equal to the range space
H
R̄n = Ān (θ)Rsn Ān (θ) (16) of An due to the existence of R̂sf n . As a result, we can-
T
where Ān (θ)  [ān (θ1 ), · · · , ān (θK1 )] , and ān (θk )  not get the exactly estimation of E Af |An by using (12), and
[1, ejωk , · · · , ej2M ωk ]T . In a similar way to the above, we R̂sf n is not affected by the additive noise, which means that
can also divide Ān (θ) into two submatrices as even the the signal-to-noise ratio (SNR) tends to infinity, the
  influence of R̂sf n still exists. Hence in the localization of the
Ān1 (θ) }K2
Ān (θ) = (17) NF signals, the estimation error doesn’t decrease with the in-
Ān2 (θ) }M +1−K2
creasing SNR, which is so-called “saturation behavior” [18].
where Ān1 (θ) is of full rank, and the rows of Ān2 (θ) can To cope with this problem, we propose an alternating iterative
be expressed as a linear combination of the rows of Ān1 (θ). scheme for localizing the NF signals.
Hence there exist a K2 × (M + 1 − K2 ) linear operator P̄ When the DOAs of the FF and NF signals and the ranges
between Ān1 (θ) and Ān2 (θ) [3] of the NF signals are estimated from (10), (21), and (22), we
H
Ān2 (θ) = P̄ Ān1 (θ) (18) can recalculate the oblique projector [15] as
or E Af |An = Af (AH ⊥
f ΠAn Af )
−1 H ⊥
Af ΠAn (24)
H H
[P̄ , −I M +1−K2 ]Ān (θ) = Q̄ Ān (θ) = O (M +1−K2 )×K2 then we can refine the estimation of the DOAs and ranges of
(19) the NF signals with (13), (15), (21) and (22). By repeating
Then we have these steps serval times, the saturation problem is solved.
H
Q̄ Ān (θ) = 0(M +1−K2 )×1 . (20)
3.5. Implementation of Proposed Method
Similarly, we can divide R̄n into two parts as R̄n = [Ḡ, H̄], When the N snapshots of array data are available, the im-
in which Ḡ and H̄ are two submatrices consisting of its first plementation of the proposed method can be summarized as
K2 or the last M +1−K2 columns. Thus when the number of follows:
snapshots is finite, the DOAs {θk }Kk=K1 +1 of the NF signals
can be estimated by minimizing the following cost function 1) Calculate the estimates of covariance matrices R in (2) as
N
fn (θ) = āH ˆ ān (θ)
n (θ)ΠQ̄ (21) 1 
R̂ = x(n)xH (n) (25)
N n=1

2278
 Proposed Method−FF 1
Proposed Method−FF
Proposed Method−NF 10 Proposed Method−NF
Proposed Method without Iterative−NF SOS−Based−FF
1 SOS−Based−FF SOS−Based−NF
10 CRB−FF
SOS−Based−NF
CRB−FF CRB−NF
CRB−NF

RMSE(deg)
RMSE(deg)

0
10

0
10
−1
10

−2
10

1 2 3
−10 0 10 20 30 40 10 10 10
SNR (dB) Number of Snapshots

Fig. 2. RMSEs of the DOA estimates versus SNR. Fig. 3. RMSEs of the DOA estimates versus snapshots.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32(2M + 1)2 N flops 4. NUMERICAL EXAMPLES

2) Estimate the DOAs of the FF signals by using (10). In this section, we evaluate the performance of the proposed
method in localizing the mixed NF and FF signals by using
. . 8(2M + 1)3 + 8K(2M + 1 − K)2 + 8K 2 (2M + 1) a ULA consisting of N = 7 sensors with element spacing
+8(2M + 1)(2M + 1 − K)2 + 16K 2 (2M + 1 − K) d = λ/4. Two signals with equal power arrive from the loca-
+2(2M + 1 − K)2 + 2K 2 + O(K 3 ) tions (∞, −5◦ ) and (1.7λ, 25◦ ), where the first one is the FF
+4M 2 + O(8M 3 ) flops signal and the second one is the NF signal. Meanwhile, the
3) Calculate the matrix R̄ in (2) with (8) and (9). behavior of the SOS–based algorithm [13] and the Cramer-
. . . 8(2M + 1 − K)3 + 16(2M + 1 − K)K 2 + O(K 3 ) Rao lower bound (CRB) [13] are also presented. The results
+8(2M + 1 − K)2 K + 2(2M + 1 − K)2 in each of the examples below are obtained from 500 inde-
+4(2M + 1 − K) + 2(2M + 1)2 flops pendent Monte Carlo trails, where SNR is defined as the ratio
of the signal power to the noise variance at each sensor.
4) Calculate the oblique projector E Af |An in (12).
Example 1–Performance versus SNR: The number of s-
. . . 22(2M + 1)3 + 32K12 (2M + 1) + 16(2M + 1)2 K1 napshots is N = 200. In Fig. 2, we can see that the DOA
+16(2M + 1)2 (2M + 1 − K2 ) + 22(2M + 1) estimates of the mixed signals are obtained separately. The
+32K1 (2M + 1)(2M + 1 − K2 ) − 5 performance of the proposed FF estimator is almost same as
−4(2M + 1)2 + O(K13 ) flops the SOS-based algorithm, while the proposed NF estimator is
5) Estimate the new covariance matrix R̄n in (15). super better than it, especially at the high SNR. Furthermore,
the saturation problem is efficiently solved by alternating it-
. . . . . . . . . . . . . . . . . 16(2M + 1)3 + 4(2M + 1)2 flops
erative procedure.
6) Estimate the DOAs of the NF signals with (21). Example 2–Performance versus Number of Snapshots:
8(M + 1)3 + 8(M + 1)(M + 1 − K2 )2 + 2K22 + O(K23 ) The SNR is fixed at 0dB, and the number of snapshots varies
+8K2 (M + 1 − K2 )2 + 2(M + 1 − K2 )2 from 10 to 1000. The results are shown in Fig. 3. It can be
+8K22 (M + 1) + 16K22 (M + 1 − K2 ) observed that for both the FF and NF signals, the proposed
+M 2 + O(M 3 ) flops method behaves better than the SOS-based algorithm.
7) Estimate the ranges of the NF signals with (22).
5. CONCLUSION
. . . 8K2 (M + 1)(2M + 1)2 + 8K2 (M + 1)2 (2M + 1)
+K2 M 2 flops Based on the SOS and the oblique projector technique, this
paper proposes a new method for the mixed FF and N-
In the above implementation, the computational complexity F sources localization problem without multidimensional
of each step is roughly indicated in terms of the number of search, HOS and eigendecomposition. The examples show
MATLAB flops, and the computational complexity of the pro- that the estimates of the parameters of both FF and NF sig-
posed method is nearly 32(2M + 1)2 N + 84(2M + 1)3 + nals are reasonably good. Meanwhile, the proposed method
16(M + 1)3 flops, when 2M + 1  K, which occurs often in is computationally efficient, and achieves a classification of
application of source localization. the source types without extra procedure.

2279
 6. REFERENCES [11] B. Wang, Y. Zhao, and J. Liu, “Mixed-order music algo-
rithm for localization of far-field and near-field sources,”
[1] H. Krim and M. Viberg, “Two decades of array signal IEEE Signal Process. Lett., vol. 20, no. 4, pp. 311–314,
processing research: The parametric approach,” IEEE Apr. 2013.
Signal Process. Mag., vol. 13, no. 4, pp. 67–94, Jul.
1996. [12] G. Liu and X. Sun, “Efficient method of passive local-
ization for mixed far-field and near-field sources,” IEEE
[2] R. O. Schmidt, “Multiple emitter location and signal Antennas and Wireless Propag. Lett., vol. 12, pp. 902–
parameters estimation,” IEEE Trans. Antennas Propag., 905, Jul. 2013.
vol. 34, no. 3, pp. 267 –280, Mar. 1986.
[13] J. He, M. N. S. Swamy, and M. O. Ahmad, “Efficien-
[3] S. Marcos, A. Marsal, and M. Benider, “The propa- t application of music algorithm under the coexistence
gator method for sources bearing estimation,” Signal of far-field and near-field sources,” IEEE Trans. Signal
Process., vol. 42, no. 2, pp. 121–138, Mar. 1995. Process., vol. 60, no. 4, pp. 2066–2070, Apr. 2012.
[4] N. Yuenand B. Friedlander, “Performance analysis [14] J. Jiang, F. Duan, J. Chen, Y. Li, and X. Hua, “Mixed
of higher order esprit for localization of near-field near-field and far-field sources localization using the u-
sources,” IEEE Trans. Signal Process., vol. 46, no. 3, niform linear sensor array,” IEEE Sensors Journal, vol.
pp. 709–719, Mar. 1998. 13, no. 8, pp. 3136–3143, Aug. 2013.
[5] W. Zhi and M. Y. W. Chia, “Near-field source local- [15] R. T. Behrens and L. L. Scharf, “Signal processing ap-
ization via symmetric subarrays,” IEEE Signal Process. plications of oblique projection operators,” IEEE Tran-
Lett., vol. 14, no. 6, pp. 409–412, Jun. 2007. s. Signal Process., vol. 42, no. 6, pp. 1413–1424, Jun.
1994.
[6] D. B. Ward and G. W. Elko, “Mixed near field/far field
beamforming: A new technique for speech acquisition [16] J. Sanchez-Araujo and S. Marcos, “Statistical analy-
in a reverberant environment,” in Proc. IEEE ASSP sis of the propagator method for doa estimation without
Workshop on Applications of Signal Processing to Audio eigendecomposition,” in Proc. IEEE 8th Workshop on
and Acoustics. New Paltz, NY, Oct. 1997, pp. 19–22. Statistical Signal and Array Processing. Corfu, Greece,
Jun. 1996, pp. 570–573.
[7] P. Tichavsky, K. T. Wong, and M. D. Zoltowski, “Near-
field/far-field azimuth elevation angle estimation using [17] J. Xin and A. Sano, “Computationally efficien-
a single vector-hydrophone,” IEEE Trans. Signal Pro- t subspace-based method for direction-of-arrival estima-
cess., vol. 49, no. 11, pp. 2498–2510, Nov. 2001. tion without eigendecomposition,” IEEE Trans. Signal
Process., vol. 52, no. 4, pp. 876–893, Apr. 2004.
[8] J. Liang and D. Liu, “Passive localization of mixed near-
field and far-field sources using two-stage music algo- [18] H. Tao, J. Xin, J. Wang, N. Zheng, and A. Sano, “Two-
rithm,” IEEE Trans. Signal Process., vol. 58, no. 1, pp. dimensional direction of arrival estimation method for
108–120, Jan. 2010. a mixture of noncoherent and coherent narrowband sig-
nals,” in Proc. IEEE 14th Workshop on Signal Process-
[9] B. Wang, J. Liu, and X. Sun, “Mixed sources localiza-
ing Advances in Wireless Communications. Darmstadt,
tion based on sparse signal reconstruction,” IEEE Signal
Germany, Jun. 2013, pp. 440–444.
Process. Lett., vol. 19, no. 8, pp. 487–490, Aug. 2012.
[19] H. Liu and W. Zhang, “A novel near-field localiza-
[10] G. Liu, X. Sun, Y. Liu, and Y. Qin, “Low-complexity
tion method based on second order statistics,” in Proc.
estimation of signal parameters via rotational invariance
Congress on Image and Signal Processing. Sanya, Chi-
techniques algorithm for mixed far-field and near-field
na, May 2008, vol. 5, pp. 29–33.
cyclostationary sources localisation,” IET Signal Pro-
cess., vol. 7, no. 5, pp. 382–388, Jul. 2013.

2280