Multi-view Object Localization in H.264/AVC Compressed Domain

This paper presents a multi-view homography-based approach for object localization in H.264/AVC compressed video surveillance sequences. The proposed novel, low-complexity method is able to accurately localize moving objects on a ground plane using multiple camera data. Contrary to existing work that exploits motion vectors for object detection and tracking, our compressed domain multi-view object localization solely uses macroblock (MB) partition information. Foreground segmentation is performed on single view compressed video data using MB partition-based temporal differencing. Blob merging, convex hull fitting and noise removal are applied on the resulting foreground views to extract objects. Once relevant objects are found in single views, they are projected onto a ground plane by exploiting the homography constraint. Since projected foreground MB views of multiple cameras will only overlap on points where foreground intersects the ground plane, object locations can be extracted by detecting local maxima on the accumulated ground plane image.