In this paper, we propose a novel block-based approach for global motion estimation (GME) and panorama image construction in encoded MPEG-1 video streams. Direct extraction of motion vectors (MV) from MPEG stream greatly improves efficiency of our method against pixel-based methods. However, some MVs in MPEG stream do not indicate real object or camera motion in the scene. Therefore, we introduce a reliability measure to discriminate real MVs from noisy or outlier MVs. In addition, an iterative reweighting process is applied to increase accuracy of GME. Finally, panorama image of several sequences is constructed by using estimated camera motion. Experiments show that the proposed method has high accuracy and can produce high quality panorama images for compressed videos faster than realtime.

Keywords