Fast 2D DOA Estimation Algorithm by an Array Manifold Matching Method with Parallel Linear Arrays

In this paper, the problem of two-dimensional (2D) direction-of-arrival (DOA) estimation with parallel linear arrays is addressed. Two array manifold matching (AMM) approaches, in this work, are developed for the incoherent and coherent signals, respectively. The proposed AMM methods estimate the azimuth angle only with the assumption that the elevation angles are known or estimated. The proposed methods are time efficient since they do not require eigenvalue decomposition (EVD) or peak searching. In addition, the complexity analysis shows the proposed AMM approaches have lower computational complexity than many current state-of-the-art algorithms. The estimated azimuth angles produced by the AMM approaches are automatically paired with the elevation angles. More importantly, for estimating the azimuth angles of coherent signals, the aperture loss issue is avoided since a decorrelation procedure is not required for the proposed AMM method. Numerical studies demonstrate the effectiveness of the proposed approaches.


Introduction
Direction-of-arrival (DOA) estimation plays an important role in many fields such as wireless communication, multiple-input multiple-output (MIMO) radar, sonar, etc. [1][2][3]. Many DOA estimation algorithms have been proposed to address the DOA estimation problem, for example, the multiple signal classification (MUSIC) algorithm [4,5], estimation of signal parameters via rotational invariance techniques (ESPRIT) algorithm [6] and propagator method (PM) [7]. Particularly the root-MUSIC algorithm, proposed in [5], can estimate more signals than elements. Based on these three classical algorithms, a large number of two-dimensional (2D) DOA estimation algorithms [8][9][10][11][12][13][14][15][16] were developed as well. Compared with the one-dimensional (1D) DOA estimation algorithms [4][5][6][7], the corresponding 2D DOA estimation algorithms face two difficulties, namely angle matching and increased complexity. Based on the assumption that the elevation and azimuth angles are independently estimated, many effective pair-matching methods were proposed [8,17,18]. For those methods, computational complexity is high, since twice 1D DOA estimation algorithms are involved. An algorithm called joint singular value decomposition (JSVD) [9] was proposed to achieve automatic pairing. However, this algorithm needs SVD of a high-order block covariance matrix, which is computationally demanding. The PM is a low-complexity DOA estimation algorithm because EVD or SVD is not required. In [15], an improved PM algorithm was proposed to achieve automatic pairing of the 2D DOA estimation. Compared with the original PM [10], this algorithm showed improved performance in both complexity e e e C π β π β π β λ λ λ , and 2 ( 1) cos 2 c o s 1 ( ) 1, , , is the steering vector of the gth linear array to the kth signal, and

CRB
The CRB is the performance benchmark for the estimation algorithms. The CRB of 2D DOA estimation with G parallel linear arrays is now derived. In [27], the CRB of 1D DOA estimation was The observed vector at the gth linear array is z g ptq " " z g,1 ptq,¨¨¨, z g,M g ptq ı T P C M gˆ1 . With θ " rθ 1 ,¨¨¨, θ K s, β " rβ 1 ,¨¨¨, β K s, z g ptq now is z g ptq " A g pθqΦ g´1 pβqsptq`n g ptq g " 1,¨¨¨G; t " 1, 2¨¨¨T (1) where sptq " rs 1 ptq,¨¨¨, s K ptqs T P C Kˆ1 is the signal vector, A g pθqΦ g´1 pβq P C M gˆK is the manifold matrix of the gth linear array in which A g pθq " " a g pθ 1 q,¨¨¨, a g pθ K q ‰ P C M gˆK , gth linear array to the kth signal, and n g ptq " rn g,1 ptq, n g,2 ptq,¨¨¨, n g,M g ptqs T P C M gˆ1 is the noise vector in the gth linear array, which is assumed to be uncorrelated at different sensors.

CRB
The CRB is the performance benchmark for the estimation algorithms. The CRB of 2D DOA estimation with G parallel linear arrays is now derived. In [27], the CRB of 1D DOA estimation was analyzed. In [15], the CRB of 2D DOA estimation with two parallel linear arrays was developed. By utilizing the similar approach in [15,27], the CRB of 2D DOA estimation with G parallel linear arrays is obtained in this section. The received signal in Equation (1) can be reexpressed in a matrix form as: where W " With the signal model in Equation (2), the CRB is expressed as: whereP " ff and σ 2 is the power of noise. In this paper, the two proposed AMM algorithms are based on the special parallel linear arrays with M 2 " M 3 "¨¨¨" M G and A 2 pθq " A 3 pθq "¨¨¨" A G pθq. In order to facilitate representation, we let M 1 " M, M 2 " M 3 "¨¨¨" M G " N, A 1 pθq " Apθq, a 1 pθq " apθq, A 2 pθq " A 3 pθq "¨¨¨" A G pθq " Bpθq and a 2 pθq " a 3 pθq "¨¨¨" a G pθq " bpθq in the following sections. Particularly, when M 2 " M 3 "¨¨¨" M G " 1, the parallel linear arrays can be seen as an L-shaped array.

Unilateral AMM Algorithm for Incoherent Signals
In this section, we present the unilateral AMM algorithm for incoherent signals. For this AMM algorithm, the number of linear arrays should not be smaller than 3, namely G ě 3. In Section 1, we mentioned that the AMM algorithm is based on the assumption that the elevation angles have been estimated. From Equation (1), we know that the vector z 1 ptq only contains the information of elevation angles. Therefore, existing 1D DOA estimation algorithms can be adopted to estimate the elevation angles with vector z 1 ptq. We now use the low-complexity PM [7] as an example to verify the availability of the unilateral AMM algorithm.

The Estimation of Elevation by PM Algorithm
The correlation matrix of the first linear array and other arrays is defined by Since the noises of different sensors are uncorrelated, the correlation matrix is given by: Specially, the correlation matrix of the first linear array and the second linear array is: where R s " E sptqs H ptq ( " diag tp 1 , p 2 ,¨¨¨, p K u. Partitioning the matrix A into two part yields: where A 1 P C KˆK is the submatrix containing the first K rows of A and A 2 P C pM´KqˆK is the submatrix containing the remaining M-K rows of A. It is easy to determine that A 1 is a nonsingular matrix, which means there must be a matrix P P C pM´KqˆK such that: Similarly, partitioning the matrix C 2 gives: where C 21 P C KˆN contains the first K rows of C 2 and C 22 P C pM´KqˆN contains the remaining M-K rows of C 2 . It is established that C 21 is a row full-rank matrix [7]. Utilizing Equation (7), the relationship of C 21 and C 22 is: Since C 21 is a row full-rank matrix, P is obtained as P " C 22 pC 21 q`. Denoting P 0 " can obtain P 0 A 1 " A. Let P 1 contain the first M-1 rows of P 0 , and P 2 contain the last M-1 rows of P 0 . Utilizing PM algorithm, we have: where Ωpθq " diag -.
The estimation of θ is now can be obtained by performing the EVD of pP 1 q`P 2 [15].

Remark 1.
From Equations (4)- (10), it is noted that this PM algorithm is based on the cross-covariance matrix of the received vectors from two different subarrays, which is different from Wu's PM [10] and Li's PM [15]. In order to achieve angle matching, Wu's PM and Li's PM used the covariance matrix of the received vector from whole array. That means the order of the covariance matrix is much higher than the cross-covariance matrix C 2 .

Unilateral AMM Algorithm for the Estimation of Azimuth Angle
According to Equations (4) and (5), a partitioned matrix C P C MˆpG´1qN is defined as: Suppose thatθ " rθ e1 ,θ e2 ,¨¨¨,θ eK s is the estimation of θ, where the arrangements of θ e1 ,θ e2 ,¨¨¨,θ eK are arbitrary. LetÂ be the manifold matrix, denoted byÂ " rapθ e1 q, apθ e2 q,¨¨¨, apθ eK qs, and it is easy to show thatÂ is a column full-rank matrix. Then, we have: Assume thatθ ek is the estimation of θ t , and we can derive: "Â`‰ k,: rAs :,j " According to Equation (13), we have: Since p t is constant, from Equation (14), we have: The estimation of β t now can be obtained using Equation (15) as: We knowβ t is matched withθ ek because it is obtained based on the assumption thatθ ek is the estimation of θ t . According to Equations (11)- (16), it is seen that the estimator Equation (16) is related to the estimated elevation anglesθ, but it does not require the method of obtaining the elevation angleŝ θ. This is the reason why the elevation angles can be estimated by any 1D DOA estimation algorithms. Hence, the proposed unilateral AMM algorithm can be combined with arbitrary 1D estimator such as [4][5][6][7]. Because of the similarity in the principle, we only take the PM as an example to avoid redundancy. In this section, we only consider the case of G = 3.

The Selection of M, N
From Section 3.1, we know that the estimation accuracy of elevation angles is affected by the value of M. From Section 3.2, we also know that the estimation accuracy of azimuth angles is affected by the value of N. In addition, it should be noticed that the azimuth angles are obtained by estimated elevation angles. Hence, the accuracy of elevation angles also affects the accuracy of azimuth angles. It is difficult to determine the exact relationship between M and N, but after intensive experiments, the three-parallel linear array with M > N is chosen. To produce satisfactory performance, N should not be too small. In Section 6, the results of the first simulation experiment can roughly demonstrate the validity of this selection. Although we are unable to obtain the exact values of M and N, an approximate range is that N should be close to M/2.

Complexity Analysis
In Section 1, we have introduced many PM algorithms [10][11][12][13][14][15], where the algorithms [11][12][13][14] are based on L-shaped array and the algorithms [10,15] are based on two parallel arrays. Hence, we only compare the proposed PM-AMM to Wu's PM [10] and Li's PM [15] in this subsection. With T " M, K, and the complexity of proposed PM-AMM is O{2NMT}. Suppose the number of elements for Li's PM [15] and Wu's PM [10] is 2L + 1, the complexity of Li's PM is O{[2L+1] 2 T} and the complexity of Wu's PM is O{(3L) 2 T}. To guarantee the same number of elements, if L is odd number, we let M = L +1 and N = L/2, and if L is even number, we let M = L and N = (L + 1)/2. Therefore, the complexity of proposed PM-AMM is O{(L+1)LT}. The complexity comparison versus different L and T is provided in Figure 2. It is observed that the complexity of proposed PM-AMM is far lower than that of Li's PM and Wu's PM. It is in the agreement with the theoretical analysis. can roughly demonstrate the validity of this selection. Although we are unable to obtain the exact values of M and N, an approximate range is that N should be close to M/2.

Complexity Analysis
In Section 1, we have introduced many PM algorithms [10][11][12][13][14][15], where the algorithms [11][12][13][14] are based on L-shaped array and the algorithms [10,15] are based on two parallel arrays. Hence, we only compare the proposed PM-AMM to Wu's PM [10] and Li's PM [15] in this subsection. With , T M K  , and the complexity of proposed PM-AMM is O{2NMT}. Suppose the number of elements for Li's PM [15] and Wu's PM [10] is 2L + 1, the complexity of Li's PM is O{[2L+1] 2 T} and the complexity of Wu's PM is O{(3L) 2 T}. To guarantee the same number of elements, if L is odd number, we let M = L +1 and N = L/2, and if L is even number, we let M = L and N = (L + 1)/2. Therefore, the complexity of proposed PM-AMM is O{(L+1)LT}. The complexity comparison versus different L and T is provided in Figure 2. It is observed that the complexity of proposed PM-AMM is far lower than that of Li's PM and Wu's PM. It is in the agreement with the theoretical analysis.

Bilateral AMM Algorithm for Correlated Signals
In this section, we develop the bilateral AMM algorithm for correlated signals using parallel linear arrays. For this AMM algorithm, the number of linear arrays should not be smaller than 2, namely 2 G ≥ . The principle of bilateral AMM algorithm is similar to the unilateral AMM algorithm proposed in Section 3. We also need to use an existing method to estimate the elevation angles of the correlated signals. Here, we adopt the BCM-based ESPRIT-like [24] to estimate the elevation angles. Then we develop the bilateral AMM algorithm to estimate the azimuth angles of correlated signals.

The Estimation of Elevation by BCM-Based ESPRIT-Like Algorithm
In this case, the correlation matrix is denoted by Section 3. Similarly, we have: We assume that the number of signals K and the number of coherent group q are known, also we assume signals in the same group are coherent, but uncorrelated with the signals in other groups. Without loss of generality, assume the largest coherent group contains Lmax coherent signals, and then we use the C2 to reconstruct a partitioned matrix as [24]:

Bilateral AMM Algorithm for Correlated Signals
In this section, we develop the bilateral AMM algorithm for correlated signals using parallel linear arrays. For this AMM algorithm, the number of linear arrays should not be smaller than 2, namely G ě 2. The principle of bilateral AMM algorithm is similar to the unilateral AMM algorithm proposed in Section 3. We also need to use an existing method to estimate the elevation angles of the correlated signals. Here, we adopt the BCM-based ESPRIT-like [24] to estimate the elevation angles. Then we develop the bilateral AMM algorithm to estimate the azimuth angles of correlated signals.

The Estimation of Elevation by BCM-Based ESPRIT-Like Algorithm
In this case, the correlation matrix is denoted by 3,¨¨¨, G as in Section 3. Similarly, we have: We assume that the number of signals K and the number of coherent group q are known, also we assume signals in the same group are coherent, but uncorrelated with the signals in other groups. Without loss of generality, assume the largest coherent group contains L max coherent signals, and then we use the C 2 to reconstruct a partitioned matrix C 2 P C pM`1´L max qˆNL max as [24]: where C 2l P C pM`1´L max qˆN , l " 1, 2,¨¨¨, L max is: It is easy to determine rankpC 2 q " K. Using the ESPRIT-like algorithm in [24], estimations of elevation angles are produced by SVD of the C 2 .
We should point out that the BCM method is similar to forward/backward spatial smoothing (SS) technique. However, compared with the SS method, it showed improved performance in the case of low SNR.

Bilateral AMM Method for the Estimation of Azimuth Angle
From Equation (17), the diagonal elements of matrix R s are the powers of the K signals. In the general case, the R s is expressed as: where p kk , k " 1, 2,¨¨¨K denotes the power of the kth signal, and it is a real number.
With Equation (20), C g now is expressed as: where "*" stands for the unknown element. Similarly as in Section 3.2, with the notations ofθ " "θ e1 ,θ e2 ,¨¨¨,θ eK ‰ , A " " apθ e1 q, apθ e2 q,¨¨¨, apθ eK q ‰ ,B " " bpθ e1 q, bpθ e2 q,¨¨¨, bpθ eK q ‰ and assume thatθ e k is the estimation of θ t , and we have: From Equations (22) and (23), for any g " 2, 3,¨¨¨, G, we have: Utilizing w 1 " 1 and w g " "Â`‰ k,: C g´"B`‰ k,:¯H , g " 2, 3,¨¨¨, G, and based on Equation (24), we have: Since p tt is a real number, utilizing Equation (25) produces estimation of β t , given by: , . - The estimate ofβ t is matched withθ ek since it is obtained based on the assumption thatθ ek is the estimation of θ t . From Equations (20)- (26), it is seen that the estimator Equation (26) also is related to the estimated elevation anglesθ, but it does not need to know how to obtain the elevation anglesθ. Hence, this AMM algorithm also can be applied to any 1D DOA estimation algorithms of correlated signals. In this section, we only consider G = 2, 3 in the following sections.

Remark 2.
For many DOA estimation algorithms of coherent signals using decorrelation approach, loss of aperture is a highlighted weakness. However, the estimator Equation (26) will not cause aperture loss. From Equations (20)-(26), we also find the bilateral AMM method is suitable for incoherent signals.

Complexity Analysis
In Section 1, we have discussed several DOA algorithms [20][21][22][23][24][25] for coherent signals, where the algorithms [20][21][22][23] are based on rectangular array, the algorithm [24] is based on two L-shaped arrays and the algorithm [25] is based on two parallel arrays. Hence, in this work, we compare the complexity of proposed BCM-AMM algorithm with TMR [25]. For a fair comparison, both algorithms use two (2L + 1)-element parallel linear arrays, where T " 2L`1. The main complexity of . The main complexity of the proposed BCM-AMM algorithm ) . The complexity comparison versus different L and T with l max " 3 is provided in Figure 3. It shows that the complexity of proposed BCM-AMM algorithm is far lower than that of the TMR, which agrees with our theoretical analysis.
Since ptt is a real number, utilizing Equation (25) produces estimation of t β , given by: The estimate of ˆt β is matched with ˆe k θ since it is obtained based on the assumption that ˆe k θ is the estimation of t θ . From Equations (20)- (26), it is seen that the estimator Equation (26) also is related to the estimated elevation angles θ , but it does not need to know how to obtain the elevation angles θ . Hence, this AMM algorithm also can be applied to any 1D DOA estimation algorithms of correlated signals. In this section, we only consider G = 2, 3 in the following sections.

Remark 2.
For many DOA estimation algorithms of coherent signals using decorrelation approach, loss of aperture is a highlighted weakness. However, the estimator Equation (26) will not cause aperture loss. From Equations (20)- (26), we also find the bilateral AMM method is suitable for incoherent signals.

Complexity Analysis
In Section 1, we have discussed several DOA algorithms [20][21][22][23][24][25] for coherent signals, where the algorithms [20][21][22][23] are based on rectangular array, the algorithm [24] is based on two L-shaped arrays and the algorithm [25] is based on two parallel arrays. Hence, in this work, we compare the complexity of proposed BCM-AMM algorithm with TMR [25]. For a fair comparison, both algorithms use two (2L + 1)-element parallel linear arrays, where 2 1 T L+  . The main complexity of TMR [25]   The complexity comparison versus different L and T with max 3 l = is provided in Figure 3. It shows that the complexity of proposed BCM-AMM algorithm is far lower than that of the TMR, which agrees with our theoretical analysis.

AMM Algorithm for L-Shaped Array
From Section 2, we can know that the parallel linear array can be seen as an L-shaped array with M 2 " M 3 "¨¨¨" M G " 1. In order to make the proposed AMM algorithm more convincing, we combine the AMM with JSVD and ESPRIT algorithm and analyse the improved performance in complexity.
We use JSVD [9] algorithm to estimate elevation angles and use proposed AMM algorithm to estimate azimuth angles (we call this algorithm as JSVD-AMM). We use ESPRIT [6] algorithm to estimate elevation angles and use proposed AMM algorithm to estimate azimuth angles (we call this algorithm as ESPRIT-AMM). Consider an L-shaped array consisting of two linear arrays, namely, M 1 " L and M 2 " M 3 "¨¨¨" M L`1 " 1. This array configuration is the same as the array used in JSVD [9] and CCM-ESPRIT [17]. The main complexity of JSVD is O , where η is the number of scanning. The main complexity of the JSVD-AMM algorithm is O ) .

The main complexity of CCM-ESPRIT is
. Obviously, JSVD-AMM is more effective than JSVD and ESPRIT-AMM is more effective than CCM-ESPRIT.

Simulation Results
In this section, totally seven experiments are presented to demonstrate the effectiveness of proposed algorithms. The root-mean-square error (RMSE) of DOA estimation as the performance measure is given by: RMSE "

AMM Algorithm for L-Shaped Array
From Section 2, we can know that the parallel linear array can be seen as an L-shaped array with . In order to make the proposed AMM algorithm more convincing, we combine the AMM with JSVD and ESPRIT algorithm and analyse the improved performance in complexity. We use JSVD [9] algorithm to estimate elevation angles and use proposed AMM algorithm to estimate azimuth angles (we call this algorithm as JSVD-AMM). We use ESPRIT [6] algorithm to estimate elevation angles and use proposed AMM algorithm to estimate azimuth angles (we call this algorithm as ESPRIT-AMM). Consider an L-shaped array consisting of two linear arrays, namely, . This array configuration is the same as the array used in JSVD [9] and CCM-ESPRIT [17]. The main complexity of JSVD is

Simulation Results
In this section, totally seven experiments are presented to demonstrate the effectiveness of proposed algorithms. The root-mean-square error (RMSE) of DOA estimation as the performance measure is given by:   In the second experiment, the pairing effectiveness and resolution of the PM-AMM algorithm are demonstrated. We use a three-parallel array with M = 11, N = 5, and consider three uncorrelated sources located at the angles of rθ 1 , θ 2 , θ 3 s " r85˝, 95˝, 100˝s, rβ 1 , β 2 , β 3 s " r45˝, 65˝, 55˝s. Figure 5a depicts the 2D DOA estimation results of the proposed PM-AMM with T = 200 and SNR = 10 dB. Now we keep the same elevation angles and change the azimuth angles to rβ 1 , β 2 , β 3 s " r45˝, 55˝, 55˝s. Figure 5b depicts the 2D DOA estimation results of the PM-AMM with T = 200 and SNR = 10 dB under the new azimuth angles. From both the figures, the elevation and azimuth angles can be clearly observed and correctly matched. Particularly, the proposed PM-AMM algorithm is able to separate the signals with the same azimuth angles. . Figure 5b depicts the 2D DOA estimation results of the PM-AMM with T = 200 and SNR = 10 dB under the new azimuth angles. From both the figures, the elevation and azimuth angles can be clearly observed and correctly matched. Particularly, the proposed PM-AMM algorithm is able to separate the signals with the same azimuth angles.
In the third experiment, we compare the proposed PM-AMM algorithm with Wu's PM [10], Li's PM [15] and CRB. A three-parallel array with M = 11, N = 5 is used, and two uncorrelated sources are located at the angles of 1  . For the Wu's PM [10] and Li's PM [15], we use an 11-element linear array and a 10-element linear array, respectively. Figure 6 shows the RMSEs of the proposed PM-AMM, Wu's PM and Li's PM versus SNR with T = 200. Figure 7 shows the RMSEs of proposed PM-AMM, Wu's PM and Li's PM versus snapshots with SNR = 5 dB. Inspecting both figures shows that the estimation precision of the proposed PM-AMM is close to that of Li's PM and Wu's PM. Keep in mind that from the complexity analysis, the complexity of the proposed PM-AMM is far lower than that of Li's PM and Wu's PM. Therefore, the PM-AMM is an attractive option to practical uses.   In the third experiment, we compare the proposed PM-AMM algorithm with Wu's PM [10], Li's PM [15] and CRB. A three-parallel array with M = 11, N = 5 is used, and two uncorrelated sources are located at the angles of rθ 1 , θ 2 s " r50˝, 60˝s, rβ 1 , β 2 s " r20˝, 30˝s. For the Wu's PM [10] and Li's PM [15], we use an 11-element linear array and a 10-element linear array, respectively. Figure 6 shows the RMSEs of the proposed PM-AMM, Wu's PM and Li's PM versus SNR with T = 200. Figure 7 shows the RMSEs of proposed PM-AMM, Wu's PM and Li's PM versus snapshots with SNR = 5 dB. Inspecting both figures shows that the estimation precision of the proposed PM-AMM is close to that of Li's PM and Wu's PM. Keep in mind that from the complexity analysis, the complexity of the proposed PM-AMM is far lower than that of Li's PM and Wu's PM. Therefore, the PM-AMM is an attractive option to practical uses.      Figure 10 shows the RMSEs of the proposed BCM-AMM algorithm and TMR algorithm versus snapshots with SNR = 10 dB. The two figures show that the estimation precision of the proposed BCM-AMM algorithms is higher than that of the TMR. Figures 9 and 10 also show that the precision of BCM-AMM with two-parallel array is better than the BCM-AMM with three-parallel array for coherent signals. In addition, from Figure 9, we can find that BCM shows better performance than SS in the case of low SNR.
In the sixth experiment, we consider two incoherent sources located at the angles of three-parallel array are provided. Since the two signals are incoherent, the ESPRIT-like algorithm is utilized to estimate elevation angles. Therefore, the BCM-AMM algorithm should be changed to ESPRIT-AMM algorithm. Figure 11 shows the RMSEs of proposed ESPRIT-AMM algorithms versus SNR with T = 200. From Figure 11, it is observed that the precision of ESPRIT-AMM with three-parallel array is better than that of the ESPRIT-AMM with two-parallel array for incoherent signals. In the fourth experiment, the pairing effectiveness and resolution of the BCM-AMM algorithm are demonstrated for coherent signals. A two-parallel array with M = 15, N = 15 is used, and three sources are located at the angles of rθ 1 , θ 2 , θ 3 s " r80˝, 85˝, 90˝s, rβ 1 , β 2 , β 3 s " r47.5˝, 45˝, 50˝s, where the first and third signals are coherent and they are uncorrelated with the second signal. Figure 8a depicts the 2D DOA estimation results of proposed BCM-AMM algorithm with T = 200, SNR = 5 dB. We now keep the same elevation angles and change the azimuth angles to rβ 1 , β 2 , β 3 s " r45˝, 45˝, 50˝s. Figure    In the fifth experiment, we compare the BCM-AMM algorithm with TMR algorithm [25], SS-AMM and CRB. We called the algorithm that SS technology combines with AMM as SS-AMM. For the BCM-AMM algorithm, we use a two-parallel array with M = 15, N = 15 and a three-parallel array with M = 14, N = 8, respectively. For the TMR algorithm [25], we use two 15-element linear arrays. Two coherent sources are located at the angles of rθ 1 , θ 2 s " r60˝, 70˝s, rβ 1 , β 2 s " r50˝, 60˝s. Figure 9 shows the RMSEs of the proposed BCM-AMM algorithm and TMR algorithm versus SNR with T = 200. Figure 10 shows the RMSEs of the proposed BCM-AMM algorithm and TMR algorithm versus snapshots with SNR = 10 dB. The two figures show that the estimation precision of the proposed BCM-AMM algorithms is higher than that of the TMR. Figures 9 and 10 also show that the precision of BCM-AMM with two-parallel array is better than the BCM-AMM with three-parallel array for coherent signals. In addition, from Figure 9, we can find that BCM shows better performance than SS in the case of low SNR.        In the sixth experiment, we consider two incoherent sources located at the angles of rθ 1 , θ 2 s " r60˝, 70˝s, rβ 1 , β 2 s " r50˝, 60˝s. The RMSEs of BCM-AMM for two-parallel array and three-parallel array are provided. Since the two signals are incoherent, the ESPRIT-like algorithm is utilized to estimate elevation angles. Therefore, the BCM-AMM algorithm should be changed to ESPRIT-AMM algorithm. Figure 11 shows the RMSEs of proposed ESPRIT-AMM algorithms versus SNR with T = 200. From Figure 11, it is observed that the precision of ESPRIT-AMM with three-parallel array is better than that of the ESPRIT-AMM with two-parallel array for incoherent signals.
In the last experiment, we test the performance of JSVD-AMM and ESPRIT-AMM for L-shaped array. We consider an L-shaped array consisting of two 10-element linear arrays, and three incoherent sources located at the angles of rθ 1 , θ 2 , θ 3 s " r40˝, 50˝, 60˝s, rβ 1 , β 2 , β 3 s " r20˝, 30˝, 40˝s. Figure 12 shows the RMSEs of JSVD-AMM, ESPRIT-AMM, JSVD [9] and CCM-ESPRIT [17] versus SNR with T = 500. Figure 13 shows the RMSEs of JSVD-AMM, ESPRIT-AMM, JSVD [9] and CCM-ESPRIT [17] versus snapshots with SNR = 10 dB. From Figure 12, it is observed that the performance of JSVD-AMM is better than that of the JSVD with SNR > 2.5 dB and the performance of ESPRIT-AMM is better than CCM-ESPRIT with SNR < 15 dB. From Figure 13, it is observed that the performance of JSVD-AMM is better than that of the JSVD with snapshots >200 and the performance of ESPRIT-AMM is better than CCM-ESPRIT for different snapshots. But we should not neglect that JSVD-AMM and ESPRIT-AMM have obvious advantages in reducing complexity, which shows in Section 5. In the last experiment, we test the performance of JSVD-AMM and ESPRIT-AMM for L-shaped array. We consider an L-shaped array consisting of two 10-element linear arrays, and three incoherent sources located at the angles of 1 2 3 = 40 ,50 ,60 ] , , ] , , ]= 20 ,30 ,40 ] β β β [ [°°°. Figure 12 shows the RMSEs of JSVD-AMM, ESPRIT-AMM, JSVD [9] and CCM-ESPRIT [17] versus SNR with T = 500. Figure 13 shows the RMSEs of JSVD-AMM, ESPRIT-AMM, JSVD [9] and CCM-ESPRIT [17] versus snapshots with SNR = 10 dB. From Figure 12, it is observed that the performance of JSVD-AMM is better than that of the JSVD with SNR > 2.5 dB and the performance of ESPRIT-AMM is better than CCM-ESPRIT with SNR < 15 dB. From Figure 13, it is observed that the performance of JSVD-AMM is better than that of the JSVD with snapshots >200 and the performance of ESPRIT-AMM is better than CCM-ESPRIT for different snapshots. But we should not neglect that JSVD-AMM and ESPRIT-AMM have obvious advantages in reducing complexity, which shows in Section 5.   In the last experiment, we test the performance of JSVD-AMM and ESPRIT-AMM for L-shaped array. We consider an L-shaped array consisting of two 10-element linear arrays, and three incoherent sources located at the angles of 1 , , ]= 20 ,30 ,40 ] β β β [ [°°°. Figure 12 shows the RMSEs of JSVD-AMM, ESPRIT-AMM, JSVD [9] and CCM-ESPRIT [17] versus SNR with T = 500. Figure 13 shows the RMSEs of JSVD-AMM, ESPRIT-AMM, JSVD [9] and CCM-ESPRIT [17] versus snapshots with SNR = 10 dB. From Figure 12, it is observed that the performance of JSVD-AMM is better than that of the JSVD with SNR > 2.5 dB and the performance of ESPRIT-AMM is better than CCM-ESPRIT with SNR < 15 dB. From Figure 13, it is observed that the performance of JSVD-AMM is better than that of the JSVD with snapshots >200 and the performance of ESPRIT-AMM is better than CCM-ESPRIT for different snapshots. But we should not neglect that JSVD-AMM and ESPRIT-AMM have obvious advantages in reducing complexity, which shows in Section 5.    In the last experiment, we test the performance of JSVD-AMM and ESPRIT-AMM for L-shaped array. We consider an L-shaped array consisting of two 10-element linear arrays, and three incoherent sources located at the angles of , , ]= 20 ,30 ,40 ] β β β [ [°°°. Figure 12 shows the RMSEs of JSVD-AMM, ESPRIT-AMM, JSVD [9] and CCM-ESPRIT [17] versus SNR with T = 500. Figure 13 shows the RMSEs of JSVD-AMM, ESPRIT-AMM, JSVD [9] and CCM-ESPRIT [17] versus snapshots with SNR = 10 dB. From Figure 12, it is observed that the performance of JSVD-AMM is better than that of the JSVD with SNR > 2.5 dB and the performance of ESPRIT-AMM is better than CCM-ESPRIT with SNR < 15 dB. From Figure 13, it is observed that the performance of JSVD-AMM is better than that of the JSVD with snapshots >200 and the performance of ESPRIT-AMM is better than CCM-ESPRIT for different snapshots. But we should not neglect that JSVD-AMM and ESPRIT-AMM have obvious advantages in reducing complexity, which shows in Section 5.

Conclusions
In this paper, AMM methods are proposed for 2D DOA estimation for parallel linear arrays. Under the assumption that elevation angles are known a priori or estimated, the azimuth angles are estimated without EVD or peak search. Moreover, the azimuth angles are matched with the estimated elevation angles automatically. Compared with existing 2D DOA estimation algorithms, the advantages of the AMM methods are threefold. First, they can be used in conjunction with any existing 1D DOA estimation algorithms and the complexity is close to the used 1D DOA estimation algorithm. Second, they can achieve automatically paired 2D angles. Third, in the process of estimating azimuth angles, aperture loss is avoided for the coherent signal for the bilateral AMM algorithm.