Tracking a varying number of sound sources using particle filtering

Tracking a varying number of moving acoustic sources is a challenging problem with many potential applications. Particle filters provide a possible solution, however as the number of sources varies, spurious estimates arise leading to highly inaccurate tracking results. In this paper we propose an algorithm for tracking a varying number of acoustic sources based on the particle filtering algorithm proposed in [1]. Firstly we propose the inclusion of a frequency selection step which ensures that only frequency subbands containing signal components are included in the particle filter. This substantially reduces the number of spurious estimates which in turn allows the tracks of the separate sources to be maintained. We also propose that when no signal is detected at a frequency subband the noise covariance for that subband can then be estimated. The improvement in the resulting noise model produces significantly better tracking results and as the assumption of white background noise can now be relaxed, the proposed algorithm is applicable to practical tracking situations. The improvement in the accuracy of the particle filtering algorithm due to the proposed modifications is demonstrated using experimental recordings.