Depth analysis for surveillance videos in the H.264 compressed domain

Knowledge about the distance of moving objects can be used to enhance the performance of object detection, tracking and classification schemes. However, such information is usually not known a priori. We present an unsupervised method to approximate basic scene geometry properties such as the camera pose in single-view video sequences. At present, the method is working in constraint environments such as traffic surveillance. The proposed approach is solely based on the motion information present in H.264 encoded, compressed video streams and does not rely on object tracking results. We start by constructing motion maps from compressed domain motion vectors. These maps are used to estimate the orientation angle of the camera, which allows to add a depth measure in the form of equidistant lines to the image plane.