Compressed domain aided analysis of traffic surveillance videos

We present a novel system to perform efficient, compressed domain aided video analysis in the context of traffic surveillance applications. After camera installation, the system initializes by performing two short and fully automatic learning stages to gather information about the background and the principal moving directions in the scene. This knowledge is later used to assist the detection and tracking of vehicles. We combine processing in the pixel domain on decoded I-frames with motion based information from the H.264/SVC compressed domain in order to obtain a hybrid solution that delivers robust results at low computational complexity. Pan-tilt-zoom cameras are supported by the system, since global motion estimation is performed using the motion vectors that are present in the compressed stream.