Annealed particle filter based on particle swarm optimization for articulated three-dimensional human motion tracking

Abstract

Three-dimensional articulated human motion tracking is challenging due to the high-dimensional parameter space and poor image observations. When particle swarm optimization (PSO) is used for human motion tracking, due to unreliable image likelihood, particles may be misled and be unable to find the most plausible pose space. This paper proposes a new PSO-based algorithm for human motion tracking, annealed PSO-based particle filter (APSOPF). The sampling covariance and annealing factor are incorporated into the velocity-updating equation of PSO; they are initialized with appropriate values at the beginning of the PSO iteration, and decreased (annealed) in reasonable steps. Through the sampling covariance, the motion prior is introduced into APSOPF, constraining particles to the most likely region of pose space and reducing the generation of invalid particles. By adopting decreasing coefficients in the updating principle, the leading effects of the local and global best on particles decrease with generations, making particles preserve their own divergence and self-exploration capabilities before convergence. Hence the problem of insufficiently reliable image likelihood can be mitigated in some degree. We compare APSOPF quantitatively with an annealed particle filter and a standard particle filter on the challenging HumanEvaI data set. Experimental results show that the proposed algorithm achieves lower estimation error in tracking real-world 3-D human motion.