Real-Time 3D Articulated Pose Tracking using Particle Filtering and Belief Propagation on Factor Graphs

Abstract

This article proposes a new statistical model for fast 3D articulated body tracking, similar to the loose-limbed model, but using the factor graph representation. Belief propagation is used to estimate the current marginal for each limb. All belief propagation messages are represented as sums of weighted samples. The resulting algorithm corresponds to a set of particle filters, one for each limb, where an extra step recomputes the weight of each sample by taking into account the interactions between limbs. To take into account fast moving limbs, proposal maps are used to steer samples to regions of high likelihood. Applied to upper-body tracking with disparity and colour images, the resulting algorithm estimates the body pose in quasi real-time (10Hz). 1