Markerless articulated human body tracking from multi-view video with GPU-PSO

In this paper, we describe the GPU implementation of a markerless fullbody articulated human motion tracking system from multi-view video sequences acquired in a studio environment. The tracking is formulated as a multi-dimensional nonlinear optimisation problem solved using particle swarm optimisation (PSO). We model the human body pose with a skeleton-driven subdivision-surface human body model. The optimisation looks for the best match between the silhouettes generated by the projection of the model in a candidate pose and the silhouettes extracted from the original video sequence. In formulating the solution, we exploit the inherent parallel nature of PSO to formulate a GPU-PSO, implemented within the nVIDIA™ CUDA™ architecture. Results demonstrate that the GPU-PSO implementation recovers the articulated body pose from 10-viewpoint video sequences with significant computational savings when compared to the sequential implementation, thereby increasing the practical potential of our markerless pose estimation approach.