Research On Improved Algorithm Of Visual Target Tracking

As an important topic in the field of computer vision,Visual tracking has a wide range of application in military,visual surveillance,car navigation and medical diagnose,etc.With the rapid growth of the computer techniques,the function of computer has been promoted continuously.In order to enable the computer to imitate the motion sensibility of human vision,visual target tracking technology emerges as the times require.In recent years,many researchers have researched on visual target tracking deeply,and propose a large number of novel theory and related tracking algorithms.However,due to the existence of clutter background,local occlusion,illumination changes,motion blur and other factors of interference,visual target tracking technology still has many problems.So the development of a highly robust,high accuracy and practical visual target tracking algorithm is still facing challenges and has a broad research prospects and practical value.On the basis of traditional visual tracking methods,combining a data mining theories,three improved algorithms are proposed to improve the performance of tracking algorithm.(1)In order to improve the effectiveness of video tracking and reduce the drift phenomenon in tracking process,a tracking algorithm based on black hole theory and Camshift algorithm,in which using the black hole principle,the weight of the color components of the visual sequence is selected and the optimal weights are obtained,is proposed.Firstly,the proposed algorithm reconstruct the images according to the optimal weight;then transfer the reconstructed image to the HSV space,the probability map of the color is obtained by computing the H component;and lastly the Mean-shift algorithm was employed to achieve the target tracking.(2)In order to deal with the appearance changes of target object effectively in the tracking process,alleviate the tracking drift problem,and improve the robustness of target tracking,an improved sparse appearance model tracking algorithm is proposed.The proposed algorithm reduced the number of the target templates and improved the real-time by searching cluster centers from the target templates via cluster theories.Firstly,the learning classification was used to select features,and then the template dictionary,which represented the object,was obtained by using the principle of cluster.Lastly the target samples were got by Gaussian function,and the best target location was obtained by using observation model under the Bayesian filter framework.(3)In order to improve the accuracy of target tracking in the occlusion,alleviate the tracking drift problem,an improved principal component analysis tracking algorithm is proposed.The proposed algorithm updated the target templates and improved the real-time by searching cluster centers from the target templates.Firstly,the principal component analysis was used to obtained target templates;and then the template dictionary,which represented the object,was obtained by using the principle of black hole.Lastly the target samples were got by Gaussian function,and the best target location was obtained by using observation model under the Bayesian filter framework.