Modeling of Motion Dynamics and its Influence on the Performance of a Particle Filter for Acoustic Speaker Tracking

Methods for acoustic speaker tracking attempt to localize and track the position of a sound source in a reverberant environment using the data received at an array of microphones. This problem has received significant attention over the last few years, with methods based on a particle filtering principle perhaps representing one of the most promising approaches. As a Bayesian filtering technique, a particle filter relies on the definition of two main concepts, namely the measurement process and the transition equation (target dynamics). Whereas a significant research effort has been devoted to the development of improved measurement processes, the influence of the dynamics formulation on the resulting tracking accuracy has received little attention so far. This paper provides an insight into the dynamics modeling aspect of particle filter design. Several types of motion models are considered, and the performance of the resulting particle filters is then assessed with extensive experimental simulations using real audio data recorded in a reverberant environment. This paper demonstrates that the ability to achieve a reduced tracking error relies on both the chosen model as well as the specific optimization of its parameters.