Human body pose recognition from a single-view depth camera

We propose a model-based approach for human body pose recognition from a single-view depth camera. The proposed algorithm applies an articulated cylinder model to detect human pose and track them based on a particle filter without numerous training data or heuristic detectors. To reduce high degrees of freedom, we adopt a hierarchical method that detects torso and limbs successively. Moreover, we take the advantage of a particle filter to track complex human motion and the results show that the proposed system is robust in human motion tracking. The qualitative evaluation shows that our method can deal with self-occlusion problem and ambiguous human motion effectively, and the quantitative evaluation shows that the average tracking error is 0.06 meters with a standard deviation of 0.04 meters. The proposed method tracks human poses successfully at the speed of 18 frames per second on a laptop with Intel Core i3-2100 CPU and without graphic processing unit.