Real-time tracking of unconstrained full-body motion using Niching Swarm Filtering combined with local optimization

Zheng Zhang, Hock Soon Seah

Abstract: We address the problem of 3D articulated full-body pose tracking from 3D volumetric data, and present an approach for tracking accurate unconstrained human motions at 4-9 fps while without using strong prior information of the dynamics. We propose a hybrid search method that combines a novel particle filter based algorithm, named Niching Swarm Filtering (NSF), with a refinement step of local optimization. In NSF, a non-parameter niching method - ring topology based Bare Bones Particle Swarm Optimization algorithm is naturally integrated with the particle filter framework. Benefiting from the niching search process, NSF can robustly and efficiently find multiple significant modes, both global and local peaks, of the configuration distribution. After the search of NSF, more accurate results are obtained through a refinement process using local optimization. With GPU implementation of NSF, the tracking can be performed in near real-time.