A robust rigid skeleton extraction method from noisy visual hull model

Abstract

The existing skeleton extraction algorithms from a coarse and noisy model cannot achieve a satisfactory skeleton, let alone the joints' central position in a markerless motion capture system (e.g., a rigid skeleton). To solve this problem, we propose a rigid skeleton extraction algorithm from a noisy visual hull model with phantom volumes. Firstly, we reconstruct the subject visual hull and the corresponding volumetric model from a multiple view synchronized video sequence. Secondly, the curve skeleton of the volume model is computed based on the theory of repulsive force fields. Thirdly, we propose a criterion for linking a curved skeleton to link the different skeleton limbs using a back-tracking method. At the same time, we obtain the distance and angle threshold values adaptively using a binary search algorithm. Finally, after achieving a smooth curve skeleton, we determine the joints' central positions in the skeleton using a priori information of the human body to form a rigid skeleton. Experimental results show that the proposed algorithm can obtain a desirable rigid skeleton with good robustness, less sensitivity to noise, and using an automatic procedure.