Markerless Motion Capture of Human Body Using PSO with Single Depth Camera

This paper presents a novel approach to model human body, recover and track its 3D position, orientation and articulated-skeleton pose parameters from a single depth camera video sequence observed by Kinect sensor. In our work, human body is modeled as assembled 3D geometric primitives whose dimensions are estimated automatically. Motion parameters are recovered by projecting hypothesized body model pose to camera imaging space and seeking for optimal solution that best matches camera observation as well as physical constraints. An objective function is designed to quantify the discrepancy between the predicted and the actual, observed features and penalize implausible or unnatural pose. We exploit body skeleton's tree structure and propose a self-adaptive version of Particle Swarm Optimization (PSO) to solve the optimization problem. In order to avoid swarm collapse and accelerate convergence, motion temporal continuity over frame sequence is exploited as initial pose using from-coarse-to-fine strategy. The overall system does not require any markers, special capture environment or complex image acquisition setup, and is ready-to-use for users.