A 2.5D Thinning Algorithm for Human Skeleton Extraction from a Single Depth Image

This paper presents a novel unsupervised approach to human skeleton extraction from a single depth image. In this approach, a 2. 5D thinning line detecting algorithm is proposed to extract thinning lines of human body in the depth data. Based on the thinning lines, we detect the outer and inner layer benchmark joints of skeletons following the anthropotomy theory. We extract skeletons in two steps, the hypothesis generation and the hypothesis validation. In hypothesis generation step, we construct a human kinematic model with 16 skeleton joints, and initialize the model in accordance with human body proportions. Next, in hypothesis validation step, the final skeleton can be achieved by fitting a standard skeleton model to benchmark joints. We evaluated the proposed algorithm on the Stanford database, and compared our results with the skeletons captured by Microsoft Kinect SDK. The experiment results prove that the proposed 2. 5D thinning algorithm can solve self-occlusion problems. Meanwhile, as the proposed approach needs no training and tracking, it is able to extract skeletons more conveniently and faster than other motion information based approaches.