Video Object Planes SegmentationBased onMoving Object Statistical Model

• Main Authors: Chao-Ping Lin, 林照彬
• Other Authors: none
• Format: Others
• Language: zh-TW
• Published: 2003
• Online Access: http://ndltd.ncl.edu.tw/handle/85937628561822282526

碩士 === 逢甲大學 === 電機工程所 === 91 === In MPEG-4 standard, each frame of the video sequences would be segment into several video object planes (VOP), then each VOP be coded using suitable tools such as shape, color, texture and motion, etc. Therefore, video object segmentation is a crucial problem in MPEG-4 coding standard. The existing approaches utilize a statistical background model and significant test to distinguish the moving object from background. An intrinsic problem is the difficult to find a consistent statistical model for the displaced frame different of the moving object. Thus, it is unable to evaluate the error probability for which an object point is misclassified as a background. After a close observation, we find that this error probability occurs especially in the homogeneous regions. Since, it requires the gradient of the frame in order to identify a moving object by the cues of the frame different. In this research, we derive a statistical model of displaced frame difference (DFD) for both the foreground and the background to detect the predictable points in the frame. With the statistical models, an error probability of the false alarm can be calculated. If the probability is too large, we declare it as an unpredictable pixel. The unpredictable pixel means that there is no information in DFD for segmenting video object plane (VOP). Based on the statistical model for DFD, we can obtain a very stable and accurate segmentation of VOP for predictable points. Therefore, we segment VOP using the temporally displaced frame different, and then a modified watershed algorithm and video region merge is utilized to determine the unpredictable points.