| Passive sources localization is one of the main topics of array signal processing and plays an important role. Based on the range between the source signal and the sensor localization, the traditional sources localization techniques mainly contain the far-field sources localization and the near-field sources localization. However, in some practical applications, such as speaker localization using microphone array and guidance system, each speaker may be in the Fraunhofer region or in the Fresnel region of the sensor position. That is, the received signals may be the mixture of far-field and near-field sources. In fact, the far-field signal model and the near-field signal model can be seen as the special forms of the mixed far-field and near-field one. Compared to these two models, the mixed far-field and near-field signal model is more common and useful. When the traditional far-field sources localization methods are directly used to the mixed sources situation, it is impossible to obtain the range estimations of near-field sources. If the previous near-field sources localization solutions are exteeded to the mixed sources localization, they may suffer from the high computational burden, the difficulty of mixed sources classification and the estimation failture problem. Therefore, it is important to research the novel algorithms to locate the mixed far-field and near-field sources, which make the passive sources localization become a much more complete theoretical system, and is required for the solutions of the practical applications, such as the speaker localization.The present mixed sources localization algorithms based on the eigen subspace can be summarized as the following two kinds: 1) Estimate the direction-of-arrivals(DOAs) of far-field and near-field sources simultaneously, then obtain the range estimations of near-field sources by substituting the angles into the two-dimensional(2-D) spectral search. The most important method for this is the two-stage multiple signal classification(MUSIC) one; 2) Estimate the DOAs of far-field sources, then classify the near-field sources from the far-field ones, and obtain the DOAs and range of near-field sources. The most famous method for this is the oblique projection based one. According to these two main ideas aforementioned, we firstly analyze the properties of the mixed far-field and near-field signal model and eigenstructure differences between the far-field and near-field covariance matrices. Then, we explore the third-order cyclic moments to reduce the computational burden, and develop the signal subspace difference and covatiance matrix difference to classify the mixed sources, and utilize the two-stage matrix difference to avoid the spurious peak problem. These works can provide the effective versions to cope with the mixed far-field and ndar-field sources localization.The main contributions and innovative points of this dissertation mainly contain the following four aspects:1. Considering the high computational complexity of the first kind eigen subspace based methods, we propose a third-order cyclic moment based algorithm and its improved version for mixed sources localization. Based on the symmetric property of the uniform linear array, this algorithm firstly constructhets two special third-order cyclic moment matrices, in which the direction matrix only consists of DOAs, and the rotational factor is the function of DOAs and range. By jointly implementing the MUSIC spectral search the Estimation of signal parameters via rotational invariance technique(ESPRIT) solution, the DOAs and range for mixed far-field and near-field sources are estimated. The proposed algorithm is computationally more efficient than the two-stage MUSIC method, and it avoids pairing parameters. In addimtion, we exploit the cyclic correlation(second-order moment) and provide a mixed-order cyclic moments based algorithm, which reduce the mean squre root error(RMSE) of range estimation by one order.2. Considering the unsatisfactory accuracy of near-field sources of the second kind eigen subspace based methods, we propose a subspace difference algorithm and its improved version for mixed sources localization. After estimating the DOAs and powers of far-field signals, the related far-field components can be eliminated from the signal subspace. Then, based on the symmetric property of the uniform linear array geometry, a near-field estimator without two-dimensional(2-D) spectral search or parameter-pairing is performed. Compared with the previous oblique projection based method, the resultant algorithm can provide the improved estimation accuracy, as well as realize a more reasonable classification of the signal types. Furthermore, we exploit the polynomial rooting to replace the MUSIC spectral search, which reduce the computationally burden of the subspace difference algorithm by one order.3. Considering the reasonable classification related to far-field DOAs estimation performance of the second kind eigen subspace based methods, we propose a spatial differencing algorithm and its modified version for mixed sources localization. For far-field sources, the array covariance matrix is an Hermitian Toeplitz one, whereas for near-field sources, the array covariance matrix only holds the Hermitian Structure. We exploit this property and firstly involve in the covariance matrix difference to cope with the mixed far-field and near-field sources localization. The MUSIC spectral search is performed to estimate DOAs of far-field sources. Based on the eigenstructure differences of the far-field and near-field covariace matrices, the spatial differencing technique is exploited to eliminate the far-field components and the sensor noise, which can realize a more reasonable classification of the signals types. Then, a modified ESPRIT-like version is conducted to locate near-field sources, which can provide the improved estimation accuracy. Besides, we exploit the symmetric Toeplite structure of the colored noise and propose an efficient application of covariance difference based version, which improves the robustness of the spatial differncing algorithm.4. Considering the spurious peak problem of the second kind eigen subspace methods, we propose a two-stage matrix differencing algorithm for mixed sources localization. By exploiting the property of the Toeplitz structure associated with the far-field covariance matrix, the covariance differencing technique is firstly carried out to eliminate the far-field components. That is, the pure near-field components can be obtained. Based on a symmetric uniform linear array, an ESPRIT-like solution can be implemented, and the DOAs and range estimations for near-field sources are performed. After estimating the powers of near-field signals, the related near-field components can be eliminated from the signal subspace, and the DOAs for far-field sources are determined via the MUSIC spectral search. The resultant algorithm can achieves a more reasonable classification of the signals types, avoid the spurious problem as well as provide the improved near-field estimation accuracy,This paper follows the two main ideas of the present mixed far-field and near-field sources localization methods, and has done some deep research on mixed sources classification and localization. The novel algorithm proposed in this paper have the advantages in reducing the computational complexity, improve the estimation accuracy of near-field sources, classify the mixed far-field and near-field sources as reasonable as possible, as well as avold the spurious peak problem. The research results of this paper will provide reference for further study on the mixed far-field and near-field sources localization based on eigen subspace.
|
PDF Full Text Request