Sound target tracking by sensor fusion using finite random set state space model

Abstract

We propose a new method for sound target tracking by sensor fusion using Finite Random Set (FRS) state space model. We use omnidirection camera and two microphones for the sound target tracking where observed signal from each sensor involve missing and false detection, and the scene involves appearance and occlu-sion. FRS is suitable for representing the multiple sensors situation with variable number of targets and observa-tions. To estimate the state of the model, we use Sequential Monte Carlo (SMC) implementation of Probability Hypothesis Density (PHD) filter, which gives approximated solution of the PHD filter by weighted particles. We pro-pose a method to decompose multiple modes of estimated PHD into each mode by clustering the weighted particles sequentially. We have been conducted experiments to demonstrate the performance of proposed method for sound targets tracking experiments.