Moving Object Detection And Tracking In MPEG-Compressed Domain

In the multimedia communication and storage applications, most videos are compressed before transferred and stored. Information processing in the compressed domain is of significant importance.This thesis focuses on moving object detection and tracking in MPEG-compressed videos. Based on the video code theory and MPEG standard, the structure of MPEG-4 video data is analyzed and the motion vectors (MV) containing the motion information are extracted. For single object detection problem, the MV filed is refined using the information of MV distribution, the intra-frame MV filed consistency of the moving object, the inter-frame MV filed correlation of the macro-blocks and the characteristic of background occluded or reappears. Then the MV filed is denosied and densed. The background is eliminated in the local region to make the MV filed more consistent with the real motion. Finally the moving object is detected by thresholding the refined MV filed. For the multi-object detection problem, the MV filed is denoised firstly. Then the amplitude and angle difference between a motion vector and its eight neighbor motion vectors are computed. The individual object is assembled and labeled using the consistency of object motion. Finally each object is tracked by nearest centroid rule and its trajectory is obtained. The proposed algorithms are tested by several hand ground-truthing video sequences. Experimental results show their validity and precision.