3D registration of multi-view depth data for hand-arm pose estimation

Human motion analysis has been applied in a wide range of applications. Specifically, hand-arm motion plays an important role in the tele-operation of a robot and in pattern analyses of human motions. However, estimations of the full pose of a human hand with the arm are challenging due to the limited recognition range and the computational cost for real-time processing. In this paper, we propose a fast and efficient solution to articulate hand-arm pose estimations for different modal depth sensors. A marker-less motion capture system is built using multiple depth sensors with different recognition areas. In the 3D registration process, a new and fast outlier rejection method is proposed. It uses the concept of skeletal consistency, which the computation time of the 3D registration process without a loss of robustness. The performance is demonstrated to assess the accuracy, the robustness and the computational cost by of the proposed system in comparison with other methods.