GLAICP: A global-local optimization algorithm for robust human pose tracking from depth data

Due to its high efficiency, the Iterative Closest Point (ICP) algorithm has become a popular choice in computer vision and robotics for the registration of point cloud data sets when the point-to-point correspondences are unknown. Its generalization for articulated structures, although possible through a joint optimization of all pose parameters, is challenging as it is necessary to solve a non-closed form. It also suffers heavily from the local minima problem. A number of proposed Articulated ICP (AICP) algorithms circumvent the problem of the non-closed form solution and offer an efficient alternative. However, they still exhibit an increased tendency, caused by the local minima, to converge to an incorrect pose. Typically, the above problem manifests itself after a transient disturbance in the convergence, such as an occlusion which causes an increase in the point-to-point association distances between the model and the data. In this paper, we propose an extension to the AICP algorithm that benefits from the efficiency of ICP as well as avoids its problems by using global pose optimization elements to guide the convergence process to the correct pose. The proposed approach is to merge adaptively the joint adjustments computed by AICP with the adjustments needed for a number of key points to reach their respective target positions, identified by a local feature descriptor search. Experiments show that the proposed Global-Local Articulated ICP algorithm exhibits improved robustness to transient disturbances, like occlusions, in comparison with the AICP algorithm.