Compressed Domain Motion Analysis for Video Semantic Events Detection

In this paper, a novel approach is proposed to estimate camera motion and segment moving objects from compressed video streams, aiming to detect semantic events in video clips. Simultaneously using the motion vectors and DC components of MPEG macroblocks (MB), the camera motion type and motion parameters of each frame are estimated with simplified models. Then the segmentation of moving objects is done at macroblock level. Exploring the variation of motion information in consecutive frame sequence, a composite feature can be formed to detect semantic events. The experiment results on TRECVID video corpus show that our approach is very effective and efficient.