GPU-Accelerated Real-Time Tracking of Full-Body Motion With Multi-Layer Search

Compared to monocular pose tracking, 3D articulated body pose tracking from multiple cameras can better deal with self-occlusions and meet less ambiguities. Though considerable advances have been made, pose tracking from multiple images has not been extensively studied: very seldom existing work can produce a solution comparable to that of a marker-based system which generally can recover accurate 3D full-body motion in real-time. In this paper, we present a multi-view approach to 3D body pose tracking. We propose a pose search method by introducing a new generative sampling algorithm with a refinement step of local optimization. This multi-layer search method does not rely on strong motion priors and generalizes well to general human motions. Physical constraints are incorporated in a novel way and 3D distance transform is employed for speedup. A voxel subject-specific 3D body model is created automatically at the initial frame to fit the subject to be tracked. We design and develop the optimized parallel implementations of time-consuming algorithms on GPU (Graphics Processing Unit) using CUDA (Compute Unified Device Architecture), which significantly accelerates the pose tracking process, making our method capable of tracking full body movements with a maximum speed of 9 fps. Experiments on various 8-camera datasets and benchmark datasets (HumanEva-II) captured by 4 cameras demonstrate the robustness and accuracy of our method.