Research On The Mapping Of TDOA To DOA For Sound Source DOA Estimation

As one of the key technologies in microphone array signal processing, the sound sources direction of arrival(DOA) estimation has been widely used in many fields, such as video conference system, fault detection, medical diagnosis, and military. The technique based on time differences of arrival(TDOAs) with multiple channels is an important method for the sound sources DOA estimation. While researchers focus on the acquisition of TDOAs, rather than the mapping of TDOAs to DOA. The mapping approach based on least squares support vector regression(LS-SVR) has shown its good performance, where its research still less of comprehensiveness. This paper focuses on the mapping of TDOAs to DOA based on LS-SVR, studies the choice of kernel functions, the construction of multi kernel LS-SVR and the sparsification analysis on support vectors. Moreover, we have proposed a tuning parameter-free mapping approach for TDOA-based sound source DOA estimation via sparse representation. The main jobs of this paper are:1) For the performance of different kernel functions are various, this paper focuses on the mapping construction of LS-SVR with radial basis kernel, polynomial kernel and linear kernel function, which influence the sound sources DOA estimation in reverberant and noise environment, and makes a comparison with least squares method. The research results show that the radial basis kernel has better estimation performance.2) Aiming at the problem that the outliers of TDOAs appear in reverberant environment, this paper proposes a TDOA processing approach based on median filtering, according to the characteristic of TDOAs to DOA mapping, to eliminate outliers. The research results shows that the sound source DOA mapping performance has been promoted effectively in reverberant environment, after using this approach.3) To further improve the sound source DOA mapping performance, this paper combines the theory of multi kernel learning and K-means clustering method, proposing a multi kernel LS-SVR mapping approach based on K-means clustering idea. The research results shows that the proposed mapping approach has better performance than single kernel LS-SVR and least squares method. In general, the more kernels the multi kernel LS-SVR owns, the better the performance it has, and the advantage of its performance shown more obviously following the increase of reverberant time.4) Aiming at the problem that the training set of LS-SVR mapping approach has some redundancy, this paper applies the sparse approximation based on pruning the minimum support values using LS-SVR to sound sources DOA estimation, and analyzes the sparsification of single kernel and multi kernel LS-SVR mapping approaches. The research results show that comparing with the basic LS-SVR approach, the sparse LS-SVR not only keeps good performance of sound sources DOA estimation, but also reduces the calculation amount of test effectively.5) This paper has proposed a tuning parameter-free mapping approach for TDOA-based sound sources DOA estimation via sparse representation. To further reduce the amount of calculation, this paper applies a two-step grid searching approach to match the TDOAs with data dictionary. The research results show that the proposed approach has some advantages over traditional tuning parameter-free mapping approach.