Using Segmented 3D Point Clouds for Accurate Likelihood Approximation in Human Pose Tracking

The observation likelihood approximation is a central problem in stochastic human pose tracking. In this article we present a new approach to quantify the correspondence between hypothetical and observed human poses in depth images. Our approach is based on segmented point clouds, enabling accurate approximations even under conditions of self-occlusion and in the absence of color or texture cues. The segmentation step extracts small regions of high saliency such as hands or arms and ensures that the information contained in these regions is not marginalized by larger, less salient regions such as the chest. To enable the rapid, parallel evaluation of many poses, a fast ellipsoid body model is used which handles occlusion and intersection detection in an integrated manner. The proposed approximation function is evaluated on both synthetic and real camera data. In addition, we compare our approximation function against the corresponding function used by a state-of-the-art pose tracker. The approach is suitable for parallelization on GPUs or multicore CPUs.