Research Of Video Object Segmentation Based On Spatio-temporal Technology

Motion Picture Experts Group (MPEG) published the secondary generationvideo coding standard: MPEG-4. Compared with frame-based techniques adopted bythe traditional video coding standard, MPEG-4is object-based coding. So the videoobject segmentation is one of the most important techniques for MPEG-4codingsystem. And it has many important applications in content-based index, objectrecognition and tracking, video conference and other multimedia technologies. So theresearch on video object segmentation techniques has turned into the most activeresearch field, until now there have been many segmentation algorithms approved, butnone of them can satisfy the MPEG-4’s requirements. Generally, there are twoproblems in the current segmentation methods: one is no universal algorithm suitedfor all video sequences because of the complexity and diversity of the real world; theother is most of the current segmentation algorithms have complex calculation processwhich are hard to meet the real-time performance.In this thesis, the novel methods are explored and investigated for video objectsegmentation. Major work of this thesis is as follow:(1) This thesis proposes a fast video object segmentation algorithm focusing onthe complex calculation problem and the over-segmentation problem in the watershedtransformation. Firstly, the first level wavelet transformation is performed on theoriginal frame image, and then the watershed transformation is performed on thelow-resolution image from wavelet transformation. The low-resolution image keepsthe edge information of the original object, and reduces the calculation of watershedtransformation. Secondly, the multi-frame difference is adopted to get the changedetection template, and then for receiving the temporal template the Gaussianclustering and morphology operator are used to remove the noisy pixels. Finally, asimple map mechanism is used to combine the temporal segmentation result and thespatial segmentation result. Experimental results demonstrate that it can not onlyreceive the complete and accurate video object, but also improve the processing speedgreatly to meet the real-time requirements.(2) This thesis proposes a novel video object segmentation algorithm based onweighted visual feature clustering focusing on video sequences with lots of noise andcomplex background. According to analyze the human visual system, researchers findhuman eyes are sensitive to image edge and motion information. Firstly, thisalgorithm extracts the visual features, such as RGB color information, edgeinformation, frame difference and optical flow information. Secondly, different weightcoefficients are added to these visual features, and the weight coefficients of edgeinformation and motion information are lager than others. Finally, this algorithm adopts the Kohonen’s SOFM algorithm to cluster pixels in the frame image, andadopts threshold method and morphological filter to receive the accurate video object.Experiment results show that convergence of the SOFM network is faster and thealgorithm can segment the video object mote correctly. Even for the video sequencewith complex background and lots of noise, the experiment results are good.(3) This thesis proposes an improved SVM video object segmentation algorithmfocusing on robustness of video object segmentation. Firstly, this algorithm adopts theadaptive change detection method to get the original video object template, whosepixels constitute the sample set for SVM training, and extracts these sample pixels’feature. Secondly, this algorithm improves the SVM using the idea of active learningwhich thinks that samples reducing the error rate mostly have the most information,so the improved SVM only studies the foreground samples. Finally, the foregroundpixels are recognized and extracted. Experiment results show that this algorithmovercomes the disadvantage of supervision learning, and it can satisfy the real-timerequirements.In summary, this thesis researches on the methodology and techniques for videoobject segmentation under the framework of MPEG-4. Focusing on the disadvantagesof current video object segmentation algorithms, the thesis proposes three videoobject segmentation algorithms and validates their effectiveness through experiments.