A new hierarchical particle filtering for markerless human motion capture

Abstract

Particle filtering (also known as the condensation algorithm) has been widely applied to model-based human motion capture. However, the number of particles required for the algorithm to work increases exponentially with the dimensionality of the model. In order to alleviate this computational explosion, we propose a two-level hierarchical framework. At the coarse level, the configuration space is discretized into large partitions and a suboptimal estimation is calculated. At the fine level, new particles in the vicinity of the suboptimal estimation are created using a more likely and narrow configuration space, allowing the original coarse estimate to be refined more efficiently. Our preliminary results demonstrates that this hierarchical framework achieves accurate estimation of the human posture with significantly reduction in the number of particles.