Study On Visual Target Detection Based On SVM

Video target detect and tracking is widely used in military and civil application, it has become a hot research field of computer vision problems. Due to the video itself contains a large number of complex information, video target has many cues, the complex background of targets, the change of illumination and occlusion in the process of tracking and the targets of visual show complex appearance. There are many problems to solve that visual target tracking has the gap with the actual application. We need an accuracy and robustness target tracking algorithms. In order to improve the accuracy of tracking, we introduced the support vector machine into visual tracking in this paper, which apply the classifiable method to solve the tracking problems. Thus we achieve high accuracy visual target tracking. Mainly have a major study in the following aspects.In visual tracking, we take the target and the background of the two categories classification problems for improving the tracking accuracy. The support vector machine builds the classifiable interface of the foreground and background. In order to improve the tracking accuracy, which take describe the feature of each pixel as samples provided to classifier. This paper has a research on SVM online update and designs the framework of the online updates for adapting the tracking scenes.For solving the tracking problem caused by complex and varied the appearance of the target, in order to improve the tracking algorithm robustness to the complex scenes, this paper take the multi-cue integrated target representation method, which extracted the clolr, texture, edge information and weighted right fusion, then establish multi-cue integrated appearance model. Furthermore, this multi-cue integrated method applies to samples description of the classifier. Experimental results show that the multi-cue integrated method improved the tracking accuracy.The proposed visual tracking method, this paper introduced a large number of videos using clolr, texture, edge features methods compared with proposed in this paper respectively. In addition, the method proposed in this paper compared with popular particle filter tracking currently. For the above methods, this paper tested them in the video library of various betas. Experimental results show that the proposed tracking method has a good performance to the complex scenes and changing the appearance of the video. Its robustness and stability have a certain improvement.