Evaluation of a hierarchical partitioned particle filter with action primitives
Zsolt L. Husz, Andrew Michael Wallace, Patrick Richey Green · 2007
We introduce the Hierarchical Partitioned Particle Filter (HPPF) designed specifically for articulated human tracking. The HPPF is motivated by the hierarchical dependency between the human body parameters and the partial independence between certain of those parameters. The tracking is model based and follows the analysis by synthesis principle. The limited information of the video sequence is balanced by prior knowledge of the structure of the human body and a motion model. This is based on action primitives, a sequence of consecutive poses, each predicting in a stochastic manner the next pose. Tracker performance is evaluated on the HumanEva-II dataset for different motion model parameters, number of camera views and particles. 1.