Research And Implementation Of Single Obiect Tracking Based On Multi-layer Visual Cue Fusion

In computer vision,visual object tracking has attracted a lot of attention from academia and industry because of its wide application in many fields,such as intelligent monitoring,scene understanding and robot navigation.However,the deformation,rotation,motion blur,scale variation for target,illumination variation,occlusion,background clutters,camera shake and other factors make the task quite challenging.Therefore,it is still difficult to develop a more robust and accurate tracking algorithm.In this paper,aiming at solving the problems existing in the correlation tracking,multi-layer visual representation cue fusion is proposed to adapt to the challenge in complex scenarios.The main results are as follows:In order to effectively reduce the in:fluence of background noise within the target bounding box on the target modeling,and to solve the insufficient robustness in the common correlation tracking caused by the overall filter template,a superpixel constrained correlation tracking algorithm is proposed,which is constructed the model by the combination of multi-layer visual cues.In the middle-level,a novel pixel-level confidence feature is proposed to characterize the superpixels,which improves the accuracy of target appearance modeling.Moreover,the sample purification and the superpixel regression have obtained more accurate and effective foreground segmentation,which alleviates the problems caused by the complex conditions such as deformation,occlusion and background noise that often exist in the tracking process.Experimental results show that the proposed algorithm performs well on many datasets.In order to alleviate the problem of "predicting target drifts to non-object background",which often exists in the correlation tracking algorithm,this paper proposes a double-layer based model to respectively perceive the appearance and object characteristics of the target.In the first layer,joint modeling of time continuity regression and reliability regression is performed to achieve effective information acquisition and reliability judgment.In the second layer,a simple strategy involviiing spatio-temporal constraints,is presented to obtain target-related proposals,which effectively improves the sample construction.Moreover,based on the structured output support vector machine and the effective training samples,a reliable objectness-level classifier is constructed,which can select more discriminative support vectors to better distinguish the target object from other object proposals.The experimental results show that compared with other methods,the proposed algorithm can achieve better tracking accuracy and improve tracking effectiveness.