Efficient tracking of cyclic human motion by component motion

A set of techniques are presented for Bayesian tracking of cyclic human motion based on decomposing a complex cyclic motion into component motions. Phases of the component motions are defined and two different mechanisms for coupling the phases are described: importance sampling and an observation model. The intensity of coupling is adaptively adjusted during tracking such that strong coupling is triggered during self-occlusion. Tracking of a walking human using motion decomposition and phase coupling is performed with an improved particle filter called the approximate kernel particle filter. We show that our approach handles foreign object occlusion and self-occlusion with improved accuracy and efficiency compared with conventional tracking without decomposition.