Sequential non-rigid point registration for 3D human pose tracking

We proposes a new generative framework for 3D human pose tracking based on point cloud sequences captured by a single depth sensor. The problem is formulated as a sequential non-rigid registration process where a subject-specific shape/skeleton model is involved and updated during pose tracking. The proposed framework is featured by four steps, visible point extraction, non-rigid registration for body segment labeling, segment volume validation and articulated registration for pose estimation. The experimental results on the benchmark depth dataset demonstrate that the new method can achieve competitive results compared with the states-of-the-art ones.