The Research Of Video Object Segmentation Based On CRF

With the continuous development of information, the ways people get information have beenchanged from traditional ways of texts to multimedia ways, such as video, image and sound. Whilenew ways bring people great convenience, they also bring a lot of unnecessary information. Theaim of video object segmentation is to pick out the necessary information in an efficient andaccurate manner from a large number of information. As an integral part of computer vision, videoobject segmentation has been widely applied in many areas, such as video surveillance, intelligenttransportation, human-computer interaction and search engine, which has already become a hot anddifficult point in computer science.In this dissertation, we do some research on video object segmentation algorithm in a gradualapproach. First of all, we achieve the object segmentation of the video at pixel level. And then byconsidering the spatial relationships between the neighborhood pixels to achieve video objectsegmentation. At last, we proposed a video object segmentation algorithm which is baesd onmulti-information fusion. We obtain the results by doing comparison experiments with otheralgorithms. The specific work as follows:(1) Analyze and summarize the existing video object segmentation algorithms to find out themainly technical difficulties and then classify the algorithm. For each category of video objectsegmentation algorithm, we propose some measures for improvement.(2) Do some research on object segmentation of the video at pixel level. A video obiectsegmentation algorithm which is based on GMM is presented.Then do a comparison experimentbetween the average background model and GMM with the same video sequence.(3) The misclassification of the video object segmentation at pixel level is serious, so we try toimprove the algorithm. A video obiect segmentation algorithm which is based on is conditionalrandom field (CRF) is presented.By using CRF, the color、texture、motion characteristics andneighborhood relations of objects are modeled to construct the unary energy functions and thepairwise energy functions. The model is trained with annotated samples by using SAMME classifier.Then do a comparison experiment between the CRF model which just contains unary energyfunctions and the CRF which contains unary energy functions and the pairwise energy functions.(4) The video object segmentation algorithm based on CRF above only applies to single picture.When we use it in video sequence, we need to improve the energy function.The energy function isamended by adding a constraint factor which is used to indicate the interaction between two adjacent images in the video sequence. Then do a experiment in the new algorithm with a videosequence which is shooted by myself. At last do a comparison experiment between the CRF andGMM with the same video sequence.Experimental results show that the video object segmentation algorithm baesd on GMM has alow complexity. But this algorithm is based on pixel level and the phenomenon of misclassificationis serious. The video object segmentation algorithm baesd on CRF can solve this shortcoming.Experimental results show that the algorithm can achieve high performance for multi-class objectsegmentation in video under a complex environment. It can also get good recognition resultswhen dealing with multi-view and serious shelter.