Potential Problems in Emerging Applications of Sparse Arrays 2021View this Special Issue
Computationally Efficient Unitary ESPRIT Algorithm in Bistatic MIMO Radar
A low complexity unitary estimating signal parameter via rotational invariance techniques (ESPRIT) algorithm is presented for angle estimation in bistatic multiple-input-multiple-output (MIMO) radar. The devised algorithm only requires calculating two submatrices covariance matrix, which reduces the computation cost in comparison with subspace methods. Moreover, the signal subspace can be efficiently acquired by exploiting the NystrÖm method, which only needs flops. Thus, the presented algorithm has an essentially diminished computational effort, especially useful when , while it can achieve efficient angle estimation accuracy as well as the existing algorithms. Several theoretical analysis and simulation results are provided to demonstrate the usefulness of the proposed scheme.
Target estimation has been a significant problem in radar systems, which has been applied in widespread in sonar, guidance systems, speech processing, communication, medical signal processing, and other fields [1–3]. In recent years, considerable research interests have been drawn to MIMO radar [4–13], which exploits multiple antennas to emit diverse waveforms and utilizes multiple antennas to receive the echo signals . This leads to its more underlying benefits over phased-array radar [15–17] (e.g., enhancing the spatial resolution, fading effect overcoming, and enhancing the parameter identifiability). MIMO radar can be regarded as an expansion of the phased-array radar, where the exploited waveforms are effectively independent . Generally, MIMO radars can be divided into two types, the collocated MIMO radar and the statistical MIMO radar, based on the different array antenna configurations . Furthermore, the collocated MIMO radars are categorised into two types, namely, the monostatic MIMO radar and the multistatic MIMO radar. Due to the fact that the emitting and receiving antennas are not in the identical location, the DOD and DOA estimation has become a considerable research matter [4, 18–21]. In our work, we mainly focus on the DOD and DOA estimation issue in the bistatic MIMO radar.
According to the recent researches, several algorithms [4–10] have been proposed for estimation angle in the bistatic MIMO radar. In , the reduced-dimension multiple signal classification (MUSIC) algorithm that uses one-dimensional search is presented to angle estimation, which achieves high angle estimation accuracy in comparison with Capon algorithm . In , the Capon algorithm is extended to DOD and DOA estimation, which has heavy computational complexity for requiring two-dimensional angle search. Moreover, the technique is subjected to some performance degradations for the proximate receiving steering vector. Besides, the estimation of angles needs peak searching with computational intensive. The root-MUSIC algorithm  without peak searching is presented by utilizing polynomial rooting technique to reduce the computational cost. In , the ESPRIT technique that uses the invariance technique of both the transmitting array and the receiving array is presented to estimate angle in the bistatic MIMO radar. However, the algorithm requires the pairing operation. In , to address the problem of automatic pairing, a combination ESPRIT-MUSIC algorithm is developed, which provides beneficial angle accuracy. In , a unitary ESRPIT technique that exploits the real-valued processing is devised for estimating angle in the bistatic MIMO radar, which has high estimation precision. In , the maximum likelihood algorithm is presented for direction finding estimation in MIMO radar. In , the novel joint angle estimation method is proposed by using tensor decomposition in the nested bistatic MIMO radar. Moreover, various methods are introduced for bistatic MIMO radar in [23–25]. However, the abovementioned algorithms have a large amount of computation since they require the calculation of sample covariance matrix (SCM) and its eigenvalue decomposition (EVD) to obtain the noise subspace or signal noise, especially for large MIMO radar array and a great deal of snapshots scenarios. In order to tackle this serious problem, a computationally efficient algorithm is devised for direction estimation in this work. Unlike the existing algorithms [4–10], the presented algorithm only requires to compute two submatrices of the SCM, which avoids calculating of SCM and its EVD by exploiting the NystrÖm technique. The proposed method can be also applied in the nonuniform linear array, L-shape array, and uniform circular array for angle estimation. The NystrÖm method has been extensively applied in speed up methods [26, 27] and is first utilized by Williams and Seeger  for sparsifying kernel matrices. By exploiting the NystrÖm method [28, 29], we extend the previous work  and develop a low complexity unitary ESPRIT algorithm which not only has high angle estimation precision but also obtains light computational cost, especially in large MIMO radar array scenario. In this paper, we derive a new powerful unitary ESPRIT approach, which exhibits many benefits as follows: (a) it has much lower computational cost than that of the ESRPIT and unitary ESPRIT methods; (b) it enjoys higher angle estimation precision than the ESPRIT algorithm; (c) it is suitable for direction finding estimation of large MIMO radar array. The benefits of the presented algorithm are shown by some simulation experiments.
2. Data Model
In this paper, we think about a bistatic MIMO radar system (Figure 1) constituted of -transmitting antenna array and -receiving antenna array, both of which are half-wavelength spaced uniform linear arrays [7–9]. Assume that there exist noncoherent targets located in the same range bin. The DOD and DOA of the target relative to the transmitting array normal and the receiving array normal are denoted by and , respectively. Thus, the signal model can be given as [7, 8]where denotes an matrix consisting of the steering vectors and illustrates the Kronecker product of the receiving array steering vector and the transmitting array steering vector for the source. and are respectively rewritten asand , where and represent the DOA and DOD, respectively. is a column vector, in which denotes the envelope of the reflected signal with being the amplitude containing the reflection coefficients and path losses and so on [7–9]. denotes an complex Gaussian white noise vector with zero mean and covariance matrix .
3. NystrÖm Method-Based Unitary ESPRIT for Angle Estimation
3.1. Real-Valued Processing
In order to reduce computational complexity, we have to transform the complex data to real data by matrix method since the array, the received data, is complex data. Let be represented as the data matrix consisting of snapshots , . The augmented data matrix is defined as , where represents the exchange matrix including ones on its antidiagonal and zeros elsewhere. Then, the real-valued matrix is expressed as [9, 11]where signifies sparse unitary matrix, expressed as
3.2. Signal Subspace Estimation
To use the NystrÖm technique [27, 28] for estimating angle, we disintegrate the matrix as follows :where and are the real-valued submatrices received by the first antenna and the rest of the antennas, respectively. We define
Moreover, we must ensure that denotes full rank matrix where satisfies , . It is noted that has not been required to ascend substantially with . For instance, when grows from to , a relatively little , such as , is sufficient to insure estimating precision, which also reduces the computational complexity.
Suppose that the EVD of is , where denotes the eigenvector matrix and represents the diagonal matrix. Defining , we can constitute a new matrix as follows :
Then, according to the results from the remark, we can obtain the signal subspace without the computation of SCM and its EVD.
Remark 1 (see ). Suppose that the EVD of is and , where denotes the eigenvalue matrix with and represents the corresponding eigenvector matrix with being the eigenvector. Then, the signal subspace is constructed by the first column vectors of as follows:where .
3.3. Angle Estimation
Then, according to the unitary ESPRIT algorithm [9, 11], the real-valued invariance relation is described as follows:where and denote real-valued matrix, respectively, and is defined in [9, 11]. denotes a steering vector that is real-valued. Thus, represents the real-valued invariance equation for the transmitter array where and signifies a real-valued diagonal matrix whose diagonal elements include information of estimating DOD [9, 11]. In the receiving array, similarly, the real-valued invariance equation is constructed bywhere and , and , , , and stands for a nonsingular matrix. represents a real-valued diagonal matrix whose diagonal elements include information of estimating DOA. Then, is described as [9, 11]
Then, the DODs and DOAs can be estimated by
4. Computational Complexity and Cramér-Rao Bound (CRB)
The presented technique does not need utilizing the whole SCM. Instead, it requires calculating and which need and flops, respectively. Meanwhile, the signal subspace is constructed by exploiting the NystrÖm approach, where the computational complexity is . Thus, the presented method requires . However, the classical unitary ESPRIT and ESPRIT algorithms need and flops, respectively, which are much higher than flops on condition that . Furthermore, referring to , we use CRB in simulation as follows:where
5. Simulation Results
In this installment, a vast number of computer simulations are demonstrated to prove the effectiveness of the proposed technique. We compare performance of the estimating angle of the presented method with the ESPRIT algorithms  and unitary ESPRIT  and present their computational complexity analysis. In the following simulation experiments, 200 Monte-Carlo iterations are adopted for the bistatic MIMO radar in the experiments. We suppose that there exist three noncoherent targets and their location is at angles , , and , respectively. The root mean squared error (RMSE) of over angle  is exploited in the simulation experiments.
Figures 2 and 3 describe the angle estimation paired results of the presented scheme with SNR = 10 dB and SNR = 10 dB, respectively. It can be shown that the transmit angles (DODs) and receive angles (DOAs) can be clearly seen. Figure 3 also implies that the presented scheme can efficiently estimate angle of the targets in low SNR scenario.
Figures 4 and 5 demonstrate performance comparison of the estimating angle with , and , , respectively. We compare the presented technique with the ESPRIT and the unitary ESPRIT methods. Figures 4 and 5 demonstrate that the proposed algorithm has much better estimation precision than the ESPRIT method and enjoys high estimation precision that is almost the same as the unitary ESPRIT scheme at high SNR range. However, the presented algorithm is somewhat inferior to the unitary ESPRIT scheme at the low SNR scenario.
Figures 6–9 show performance comparison of estimation of the presented technique with and for different , respectively. From Figures 6–9, we can find that the angle estimation precision of the presented scheme is significantly enhanced with the number of transmitting array elements/receiving array elements increasing. Multiple receiving/transmitting array elements enhance estimation precision owing to diversity gain.
Figures 10 and 11 illustrate estimation precision comparison of the presented technique with and for different values of , respectively. As shown in Figures 10 and 11, the estimation precision of the presented technique is boosted with increasing. Meanwhile, Figure 10 also indicates that the presented method has beneficial estimation precision at small number of snapshots scenario.
Figures 12–14 illustrate the complexity comparison with , where we can find the proposed algorithm has much less computational cost in comparison with the ESPRIT and the unitary ESPRIT schemes, particularly when becomes larger.
Figure 15 and Table 1 describe the runtime of the three ESPRIT schemes. They depict the average CPU time required to calculate each ESPRIT approach on the personal computer with Intel(R) core(TM) 2 Duo CPU T3700 processor. We can clearly observe that the presented approach is much more computationally effective than the existing schemes, especially when becomes bigger.
In this paper, we have developed a low complexity unitary ESPRIT method for estimating angle in the bistatic MIMO radar. Compared with the existing unitary ESPRIT and ESPRIT algorithms which require and flops, respectively, our approach only needs flops, thereby being much more computationally effective, especially for the case of a large MIMO radar array. Moreover, extensive simulation results demonstrate that the estimation precision of the presented scheme is much higher by comparison with the ESPRIT method and very similar to the unitary ESPRIT algorithm. In the future research, the presented technique can be extended to a different application such as estimating angle in the monostatic MIMO radar.
The relevant data used to support the findings of this study are within the manuscript.
Conflicts of Interest
The authors declare that there are no conflicts of interest in this work.
The authors wish to express their sincere thanks to the Natural Science Research Plan in Shaanxi Province of China (General Program, no.2020JM-574), the Ph.D Science Foundation of Polytechnic University (no.107020313), the National Natural Science Foundation of China (no.62001352), and the Technology Innovation Guidance Project of Shaanxi Province of China (no.2020CGXNG-012) for their support.
M. L. Bencheikn, Y. D. Wang, and H. Y. He, “Polynomial root finding technique for joint DOD and DOD estimation in bistatic MIMO radar,” Signal Processing, vol. 90, pp. 2723–2730, 2010.View at: Google Scholar
D. F. Chen, B. X. Chen, and G. D. Qian, “Angle estimation using ESPRIT in MIMO radar,” Electronics Letters, vol. 44, no. 12, pp. 770-771, 2008.View at: Google Scholar
E. Fishler, A. Haimovich, R. Blum et al., “An idea whose time has come,” in Proceedings of the 2004 International Radar Conference, pp. 71–78, Delft, Netherlands, April 2004.View at: Google Scholar
A. Nicholas and J. W. Patrick, “Estimating principal components of large covariance matrices using the NystrÖm method,” in Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), pp. 3784–3787, Prague, Czech Republic, May 2011.View at: Google Scholar
P. Drineas and W. M. Micheal, “On the NystrÖm method for approximating a gram matrix for improved kernel-based learning,” Journal of Machine Learning Research, vol. 6, pp. 2153–2175, 2005.View at: Google Scholar
C. K. I. Williams and M. Seeger, “Using the NystrÖm method to speed up kernel machines,” Advances in Neural Information Processing Systems 2000, MIT Press, Cambridge, MA, USA, 2001.View at: Google Scholar
J. P. Shi, F. Q. Wen, and T. P. Liu, “Nested MIMO radar: coarrays, tensor modeling and angle estimation,” IEEE Transactions on Aerospace and Electronic Systems, vol. 57, no. 1, pp. 574–585, 2021.View at: Google Scholar