Switching particle filters for efficient real-time visual tracking

Particle filtering is an approach to Bayesian estimation of intractable posterior distributions from time series signals distributed by non-Gaussian noise. A couple of variant particle filters have been proposed to approximate Bayesian computation with finite particles. However, the performance of such algorithms has not been fully evaluated under circumstances specific to real-time vision systems. In this article, we focus on two filters: condensation and auxiliary particle filter (APF). We show their contrasting characteristics in terms of accuracy and robustness. We then propose a novel filtering scheme that switches these filters, according to a simple criterion, for realizing more robust and accurate real-time visual tracking. The effectiveness of our scheme is demonstrated by real visual tracking experiments. We also show that our simple switching method significantly helps online learning of the target dynamics, which greatly improves tracking accuracy.