A Multiple Hypothesis Gaussian Mixture Filter for Acoustic Source Localization and Tracking

In this work, we address the problem of tracking an acoustic source based on measured time difference of arrivals (TDOAs). The classical solution to this problem consists in using a detector, which estimates the TDOA for each microphone pair, and then applying a tracking algorithm, which integrates the "measured" TDOAs in time. Such a two-stage approach presumes (1) that TDOAs can be estimated reliably; and (2) that the errors in detection behave in a well-defined fashion. The presence of noise and reverberation, however, causes larger errors in the TDOA estimates and, thereby, ultimately lowers the tracking performance. We propose to counteract this effect by propagating the detection uncertainty. That is achieved by sampling from the GCCs and then integrating the resulting TDOAs in the framework of a Gaussian mixture filter. Experimental results show that the proposed filter has a significantly lower angular error than a multiple hypothesis particle filter.