Model-based Human Pose Estimation Using Labelled Voxels by ICP

We present a system for markerless motion capture by using labelled voxels which can recover human posture of the subsequent frames robustly and precisely on temporal coherence. The system uses 3D voxel data reconstructed from multiple synchronized video streams as input, and initialize the model posture by segmented silhouette, and then labeling the voxel data of next frame by fitting the human body model to the configuration of previous frame. After then, we track the human body by matching the model to the voxel data using Iterative Closest Point (ICP) algorithm. In order to gain the robust results, we make the intrinsic voxel data to motion tracking, and the surface voxels are used for global optimization. The experiment results indicate that this method is valid and robust.