Research On Visual Tracking Based On Saliency Detection And Compressive Sensing

Visual tracking which is the fundamental research in the domain of computer vision has great practical application significance, but traditional tracking algorithms still have many shortcomings. Aiming at the technical problem, some improved methods are proposed in this paper. The major work is as follows:Firstly, the principle of different tracking algorithms and visual attention models are described systemically, and the advantages and disadvantages of it are analyzed. According to the cognitive mechanism of human vision build computer models, and detect visual saliency regions of the image. This paper focuses on the frequency domain algorithm. Above all works provide the theoretical basis for feature selection and stability optimization of tracking.Secondly, the state is predicted by the related information of target and proto-objects, and the proto-objects are used for describing the significant assumptions of real target. The proto-objects are detected by using the saliency value of the area near the center of the target. The joint posterior distribution of proto-objects and target are established according to the Gibbs sampling algorithm to approach this distribution. Finally, the MAP(maximum a posterior) algorithm is used to get the optimal estimation of the center of target.Thirdly, in order to solve the singular describing of candidate target in object tracking, an algorithm based on the adaptive feature fusion of visual significant feature is proposed. The algorithm which is based on the frequency domain filtering principle is used to extract saliency map. The visual significant feature and the color feature are combined to describe the target appearance, and the fusion weights adaptively adjusted according to the similarity coefficient. The experimental results demonstrate that the approach can effectively overcome the part occlusion and the interference of background, so as to realize the accurate tracking under the case of complex background.Finally, aiming at constant learning rate in Real-time Compressive Tracking, when great changes and fast moving have happened the target will drift or even be lost, the tracking algorithm adaptively adjust the learning rate of classifier is proposed. An adaptive weighted compressive sensing random sparse matrix is used to extract features. The learning rate of weak classifiers is adaptively determined based on symmetrical KL distance of samples’ conditional distribution. Use the na?ve Bayesian classifier to predict the target location in the next frame. Experimental results demonstrate that the proposed method can effectively track the target and has good robustness in the fast moving and occlusion cases.