Vision-Based 3D Articulated Pose Tracking Using Particle Filtering and Model Constraints

We describe a probabilistic approach for 3D upper body pose tracking by fusing depth, color and underlying body constraints. Existing tracking algorithms can be roughly divided into model-free and model-based methods. Probabilistic assembly of parts falls into model-free category. An important advantage of this technique is that pose can be estimated independently at each frame, allowing estimation for rapid movements, but most such approaches only get 2D tracking results. The use of an explicit model is the most widely investigated methodology, but often suffers from high computational costs. In this paper, we employ particle filtering to get candidate body parts with salient features, integrate probabilistic assembly of parts with model constraints to get the best pose configuration. Experimental results show that our approach can robustly track human motion even when hands move rapidly or self-occlusion exists, and can also automatically initialize and recover from tracking failure.