Mixing Graphics and Compute for Real-Time Multiview Human Body Tracking

This paper presents an effective algorithm for 3D modelbased human motion tracking using a GPU-accelerated particle swarm optimization. The tracking involves configuring the 3D human model in the pose described by each particle and then rasterizing it in each camera view. In order to accelerate the calculation of the fitness function, which is the most computationally demanding operation of the algorithm, the rendering of the 3D model has been realized using CUDA-OpenGL interoperability. Since CUDA and OpenGL both run on GPU and share data through common memory the CUDA-OpenGL interoperability is very fast. We demonstrate that thanks to GPU hardware rendering the time needed for calculation of the objective function is shorter. Owing to more precise rendering of the 3D model as well as better extraction of its edges the human motion tracing is more accurate.