Multi-Target Sound Source Localization And Tracking Based On Microphone Array In Near-Field Environment

The microphone array based sound source localization technology is becoming increasingly important due to the development of intelligent equipment. It is the major concern how the problem of natural noise, reverberation and randomness of sound source is solved for sound localization.In this paper, we conduct a deep study on the localization of sound sources with frequency during the Auditory range(20Hz-20KHz). Our research is based on the near-field acoustic propagation model and work on improving the effectiveness and practicability of sound source localization.The research is mainly in the following aspects:(1) In order to locate multiple sound sources in real time, choose the CSP-SRP algorithm for key studies. The noise robustness of traditional CSP method is bad and it can hardly locate the wide-narrow band hybrid sound sources at one time. To overcame these problems, a logarithmic peak-to-average ratio weighted CSP(LPARW-CSP) is proposed. LPARW-CSP is based on the peak-to-average ratio of multi microphone power spectrum summation. By introducing the weighting factor, it can inhibit the amplitude of noise frequency bins and enhance the amplitude of sound source frequency bins. To prevent the excessive enhancement of single frequency bin from masking other frequency bins, the peak-to-average ratio is transformed to the log domain. In addition, the SRP searching strategies are studied to save computation time. It is showed by theoretical study and tests that LPARW-CSP improves the estimation accuracy in the strong noise efficiently and can locate both narrowband tone and wideband speech. The computation cost of optimized near field SRP searching strategies is reduced greatly compared to global searching, which brings convenient for real-time processing.(2) Under the framework of Particle filter, the motion trajectory of sound source is being tracked. Based on the model of excitation-damping, we set up the state equations for describing the motion of sound sources, and the observation of sound source location is obtained by using LPARW-CSP combining with SRP searching, then use the Gauss distribution sampling density function and likelihood function to make importance sample on the posterior probability density of a real source. The distribution rule of multiple sound source is proposed to solve the problems in dynamic tracking such as the matching of observation results with tracking sound source, the birth of a new source and the vanish of an old source. Simulation experiments shows that the tracking trajectory can quickly converge to the true trajectory of sound source and the influence of reverberation and short silence on the estimation of sound source location is reduced.