A particle filtering framework with indirect measurements for visual tracking
Haihong Zhang, Weimin Huang, Zhiyong Huang, Bailing Zhang · 2005
Particle filtering is a stochastic approach to Bayesian recursive inference. In many computer vision applications with limited number of random samples, however, conventional particle filters may find it difficult to accurately represent the desired a posteriori distribution especially for target objects with narrow likelihood functions. This paper proposes a new particle filtering framework which, by incorporating a special indirect measurement model, can significantly improve the representation capability of the particle set, yielding an accurate estimation of a posteriori distribution for the purpose of tracking. In particular, an add-on resampling technique is proposed to incorporate the indirect measurement. In this way, we can alleviate the problem with large numbers of particles required in conventional particle filtering. Positive experimental results on both synthetic sequences and real world videos are obtained.