A Hierarchical Dynamic Bayesian Network approach to visual tracking

Abstract

Usually a uniform observation strategy will result in frustrated tracking processes. To address this problem, we construct a flexible model with Hierarchical Dynamic Bayesian Network by introducing hidden variables to infer the intrinsic properties of the state and observation spaces. With this model, a dynamic-mapping is built between target state space and the observation space. Based on a decoupling based inference strategy, a tractable solution for this algorithm is proposed. Experiments of human face tracking under various poses and occlusions show promising results. © Springer-Verlag Berlin Heidelberg 2004.