Research Of Video Object Extraction Based On Support Vector Machine

Video object extraction technology is one of the important research topics in the fieldsof digital video processing and computer vision. Also, it is the critical technique in manyapplications such as intelligent video surveillance, video communication, long-rangemedical treatment and so on. Therefore, video object extraction technology has become theresearch hotspot, and it also gains more and more attentions of the researchers. Although alot of algorithms on the video object extraction have been developed, no single algorithmcan be fit for all kinds of video scenes. On the other hand, there is a certain distancebetween them and the MPEG-4standard. As a newly emerging technology, Support VectorMachine (SVM) has the characteristics of small samples and non-linearity. Not only can itsimplify the computational complexity, but also it has high robustness. When used to dealwith the classification problem, it acquires better performance.In this thesis, SVM is used for extracting video object. We combine it with the BinaryTree (BT) decision-making model and realize the adaptive classification. The main workand results are as follows:(1) Make a detailed study on the video object extraction technology and the basictheory of SVM. Analyze and implement some classic video object extraction algorithmsthat involved in this thesis including the threshold segmentation method, edge detectionmethod and the change detection method.(2) Improve the traditional change detection method, and combine it with the edgedetection to reach the rough extraction of the foreground objects. Firstly, use the improvedframe difference method to detect the target moving region; secondly, carry out the edgedetection to the moving region and the current frame, and then operate the matchingprocess on the two edges to obtain the accurate edges of the moving target; finally,post-process the obtained edges. It is shown that the algorithm can realize rough extractionof the foreground object, and the extraction results are satisfactory. (3) Extract the features of the foreground object including the motion feature, edgefeature, neighborhood gray features and local transform domain features. The motionfeature is computed by the optical flow motion vector; the edge feature is calculated by thegradient information; the neighborhood gray features are the local entropies which aremainly calculated from the block units, and we take the local entropies of the current blockand its four neighborhood blocks as their feature vectors. The local transform domainfeatures are calculated by the Discrete Cosine Transform (DCT) of the current block whichcenters on the target pixel, and the local feature vectors are calculated by the transformedparameters.(4) At the stage of SVM training and recognition, the method of SVM combining withBT decision-making model is applied in the process of video object extraction. SVMclassifier calculates the accuracy at the end of each classification and automaticallydetermines whether the further training and classification are carried out or not. If thefurther classification is needed, SVM classifier will carry out them on the foregroundpixels which have been separated in the previous time. This is the basic idea of BT-SVM.In this thesis, Akiyo, Grandma and Mother and daughter sequences are used in thesimulation experiments. At last, we compared our experimental results with thesegmentation results of COST AM. It is shown that the spatial accuracy (SA) and thetemporal coherency (TC) of the BT-SVM algorithm are improved by5–27%and3–9%compared with that of COST AM. From this it can be seen that BT-SVM algorithm forvideo object extraction is feasible and also has a high stability.