High performance GPGPU based system for matching people in a live video feed

One of the key problems of computer vision and automated surveillance is to determine if two snapshots of objects in a video feed correspond to the same real one. In this paper we propose an efficient GPGPU based system for short-term matching of people in a video feed. The main contributions of our approach consist of image enhancement techniques, data preprocessing methods based on statistical sampling combined with local algorithms for finding Voronoi diagrams and efficient similarity metric based on non crossing maximum matchings in weighted graphs. Our algorithms, thanks to their local nature, are easily parallelized. We propose an implementation on GPGPU that allows real time computation in reasonable circumstances. Achieved results show that described algorithms may be used in a variety of contexts.