Bayesian filtering is an elegant probabilistic framework which is often applied to the recovery of human pose and motion from video sequences. Existing methods typically use weak priors that fail to account for the complex dynamics and strong constraints associated with human motion. More powerful priors can facilitate efficient inference over pose in the presence of noisy image observations, occlusions and inherent ambiguities. In Chapter 4 we showed that the Conditional Restricted Boltzmann Machine (CRBM) could synthesize and denoise human motion using binary latent variables. By defining a joint probability distribution over pose and binary latent features given a past history of pose observations, they fit naturally into the Bayesian filtering framework. This chapter explores CRBMs as motion priors in the context of Bayesian filtering.
The player will show in this paragraph unless you do not have flash player installed
The player will show in this paragraph unless you do not have flash player installed