Clustering and identification of body extremities for pose recognition through a network of calibrated depth sensors

This thesis presents a framework of a marker-less human pose recognition system by identifying key body extremity parts through a network of calibrated depth sensors. The depth sensors can overcome challenges related to low illuminations which usually compromises the information from the RGB cameras. The thesis proposed a novel approach for calibrating multiple depth sensors using retro-reflective (RR) marked spheres. The calibrated parameters are then used to align the point cloud data of the human body associated with multiple depth sensors with respect to a common coordinate frame. This fusion of point clouds facilitates in overcoming the self-occlusion problems from body parts without incurring disjointedness in the fused point cloud data. The second part of the thesis introduces a novel algorithm for the identification of key body extremities such as head, hands, and feet of a human subject. A geodesic mapping is applied on the fused point cloud to produce a set of distinct topological clusters of 3D points. From these clusters, a hierarchical skeleton tree graph is generated and used for key extremities classification which finally leads to pose recognition. The thesis presents the assessment of each proposed part and its comparison with other available techniques in a succession of experimental configurations.