3D Human Upper Body Pose Tracking Based on Multi-Template Register

In three-dimensional (3D) human motion tracking from monocular video sequences, twodimensional (2D) body template cannot register accurately due to occlusion, which lead to track fail. To overcome this problem, a 3D texture model is used to track human motion. First, the template of the texture model and the local coordinates of each pixel under straight pose are initialized in the first frame. Then, the human motion is optimized by the pixel intensity consistency constraint, and a new 2D template is obtained by the tracked result. To contain the latest body information, the oldest template in the collection is dumped while a new template arrives, and the finally tracking result is determined by the output of template with the highest similarity. The experimental results demonstrate that the proposed algorithm is more effective in solving occlusion caused by segment self-rotation than the single 2D template matching method.