Efficient body part tracking using ridge data and data pruning

This paper proposes a model-based human pose estimation from a sequence of monocular depth images using ridge data and data pruning. The proposed method uses the ridge data that is defined as the local maxima in the distance map because it estimates the human pose robustly and fast due to its selective representation of body skeletons. The proposed method performs four functional subtasks sequentially: (1) it segments human depth silhouettes from depth images by executing floor removal, object segmentation, human detection and human identification, (2) it extracts ridge data from each segmented human depth silhouette by finding the local maxima over the distance map, (3) it generates initial human model parameters such as the lengths between two neighboring joints, and (4) it estimates the human pose by tracking the body joints in a hierarchical order of head, torso, and limbs and pruning illegal ridge data based on the joint length constraints. In pose estimation experiments on the benchmark dataset, SMMC-10, the proposed method achieved 0.9671 mean Average Precision (mAP) and 280 frames per second (fps). The experimental results over the SMMC-10 dataset show that the proposed method estimates the human pose fast and tracks the body joints accurately under various self-occlusion and fast moving condition.