A new approach for acceleration of motion tracking in video using a combination of foveation and CUDA technology is described. The use of GPU-accelerated foveation allows motion segmentation to be performed at high frame rates on high-resolution video sequences. To illustrate the technique, the implementation of an optical flow algorithm and its application to motion-segmented video for the real-time visual position-servo of a robotic manipulator is provided. Mapping of the foveated motion segmentation algorithm to a 240 processor GPU is illustrated and the performance of the algorithm is characterized with examples of both synthetic and real data. The non-foveated segmentation algorithm is shown to have a significant performance increase over a single-threaded CPU application and the foveated-based segmentation is found to give an additional performance gain of up to 27x over non-foveated optical flow.

Paper available at ACM.