Research On Near Field Sound Source Localization Algorithm

Sound source localization technology is widely applied to many scenes in our lives,such as classroom teaching,small venues,video conferences and so on.With the development of sound source localization technology,the study of near field sound source localization algorithms has become one of the research hotspots in the field of signal processing.There exist a number of defects in traditional algorithm,such as poor estimation accuracy,high computation complexity and necessity for distance and angle pairing in algorithm based on subspace method,etc.Based on the analysis of existing algorithms,the following three improvements have been made in this thesis.1.For the uniform linear array,the traditional algorithm of near field Multiple Signal Classification(MUSIC)is improved.The improvement of new algorithm is mainly reflected in the computation complexity.In the algorithm,a new covariance matrix is constructed which only contains the angle information.By angel searching based on the MUSIC algorithm,the direction of arrival can be estimated.Then by decomposition of the ordinary covariance matrix and distance searching in each arrived angle,the distance information can be obtained.The traditional near field MUSIC algorithm is a process of two-dimensional spectral peak searching.While the improved algorithm is a process of multiple one-dimensional spectral peak searching.Compared with the tranditional algorithm,the new algorithm can maintain similar performance and significantly reduce the computational complexity.2.For the uniform linear array,the Propagator Method(PM)algorithm is improved.A new cumulant-based near-field source distance and angle joint estimation method is proposed.The improved algorithm reflects the rotation factor by constructing different high-order cumulant matrices,and the rotation matrix are derived based on the propagator method.The moduli and phase angles of the rotation matrix's eigenvalues contain angles and distances information of the incident signals and the auto-paired angles and distances can be achieved.Compared with the traditional algorithm based on the propagator method,the improved algorithm only needs one eigenvalue calculation.So the computational complexity is reduced.The performance is similar to the performance of the existing algorithm based on the propagator method and the invariance of signal space rotation.What's more,traditional algorithms require pairing algorithms.3.For the time delay estimation method,a new frequency domain weighting function is proposed to improve the generalized cross correlation algorithm.Aiming at the frequency domain weighting function in the generalized cross-correlation algorithm,the improved algorithm combines the ROTH frequency domain weighting function with the PHAT frequency domain weighting function by introducing a weighting factor,and the weighting factor can be determined by the prior knowledge of the sound source.