Segmentation Algorithm For Mpeg-4 Video

Audio-video object-based coding and description are the distinct tbatures of the MPEG-4standard. A video sequence must be segmented into a set of video objects. Video segmentation is achallenging topic of research. In fact, no single algorithm has been reported in the literature that isgenerally applicable. Video object segmentation technique is not a normative part of the standard, butthe performance of the segmentation algorithm is crucial to the finaI MPEG-4 coding products.In this dissertation I concentrate my research work towards fast algorithms and real-timeimplementation of video object segmentation. l deem that the homogeneity of motion is essential forsegmentation. In order to approach a real-time application, we must make the compromise betweenhigh precision and fast speed of motion estimation algorithms. This article introduces a method basedon multiresolution scheme to estimate parametric motion. The motion information of a video objectcan be utilized to track the video object.There exist three contributions in this dissertation as follows.1. Under the framework of the multiresolution estimation of parametric motion, images aredecomposed by the discrete wavelet transform (DWT) at the first step. The estimation of motionparameters is carried out from Iower resoIution to higher one. In each resolution, the steepest descentmethod is used to refresh the motion parameters. Under the Iowest resolution the initial estimation ofthe motion parameters is achieved by the optical flow method. The initial motion estimation of higherlevels can be obtained by projection. To reduce the effect of mismatch of the model, a robustestimation method is demanded. In this dissertation the M-estimator is embedded in themultiresoIution estimation algorithm. As a result, the estimation precision is improved. Because themotion estimation is a time-consuming task, a hardware implementation is desired for the real timeapplications. An architecture of the hardware implementation is proposed in the article.2. Unlike the other video object segmentation algorithms, the method in the articIe mainly makesuse of the motion homogeneity to segment the video objects. The motion models are estimated one byone. There are associated support regions for the motion modeIs. Each support region impIicates avideo object. A comparison is made between our aIgorithm with a typical one by D. Wang. The resuItshows that the computational cost of our algorithm is lower.3. The first step of video object tracking is to project every video object of the previous frameinto the current frame. The result is the initiaI segmentation of the current frame. The finalsegmentation is gotten through a refresh process where the edge pixels are reassigned in terms of themotion homogeneity. To improve the efficiency of the algorithm, a raster and anti-raster scan methodis introduced. The results of simulation experiments confirm the validity of the algorithm.It is generally realized that all the reported MPEG-4-oriented video segmentation algorithmshave its restrictions. There are stil1 many problems unsolved. A challenging and unavoidable questionis how to impIement the algorithms by hardware. These are the subjects of author's future study.