Non-Rigid Articulated Point Set Registration for Human Pose Estimation

Abstract

We propose a new non-rigid articulated point set registration framework for human pose estimation that aims at improving two recent registration techniques and filling the gap between the two. One is Coherent Point Drift (CPD) that is a powerful Gaussian Mixture Model (GMM)-based non-rigid registration method, but may not be suitable for articulated deformations due to the violation of motion coherence assumption. The other is articulated ICP (AICP) that is effective for human pose estimation but prone to be trapped in local minima without good correspondence initialization. To bridge the gap of the two, a new non-rigid registration method, called Global-Local Topology Preservation (GLTP), is proposed by integrating a Local Linear Embedding (LLE) -based topology constraint with CPD in a GMM-based formulation, which accommodates articulated non-rigid deformations and provides reliable correspondence estimation for AICP initialization. The experiments on both 3D scan data and depth images demonstrate the effectiveness of the proposed framework.