Kernelized Correlation Filter Tracking Method Based On Robust Matching

Target tracking,which is one of the research objects in the field of computer vision,has been concerned owing to its great value in the society production.With developing of target tracking,many excellent tracking algorithms emerge,which improve the accuracy and success rate of tracking.However,it is still a great challenge to design a target tracking algorithm with high robust matching as a result of the uncertainty of tracking background and target itself.Aimed at the problems of occlusion,deformation,motion blur and scale change in the process of target motion,this paper introduces the template matching technology into the Kernelized Correlation Filter(KCF)tracking framework.The main jobs are as follows:(1)Aimed at the problem of low tracking accuracy of KCF algorithm in complex environment,Best Buddies Similarity(BBS)-guided KCF tracking algorithm was proposed.Because Edge boxes algorithm could extract the target edge information in the image with high accuracy,a large number of target candidate boxes could be generated by using this algorithm.Then,the candidate boxes with low similarity to the target template could be eliminated by designing the candidate box screening mechanism,and the BBS score of the target candidate boxes was obtained by matching the reserved target candidate boxes with the target template.Finally,the BBS score and KCF maximum position response score were combined to estimate the position of the target.The tracking results show that the designed tracker can locate the position of the target when the target rotates,occludes or deforms.(2)Aimed at the problem of target blur in the tracking process,a KCF tracking method combining fuzzy feature detection was proposed.In this algorithm,two tracking schemes were designed according to the image definition.If the current image was fuzzy,the reduced Scale Invariant Feature Transform(SIFT)descriptor was combined with Local Binary Pattern(LBP)algorithm to construct the fuzzy feature detector to extract the target feature points in the fuzzy image,and match them with the target template to predict the position of the target.If the current image was clear,the position of the target was estimated on the basis of the output response of the traditional KCF algorithm position filter.The tracking results show that the designed tracker is suitable for target tracking in fuzzy images.(3)Aimed at the problems of using a fixed size tracking box in the traditional KCF algorithm and the poor real-time performance of the fuzzy feature detector designed in work 2,an adaptive scale KCF tracking algorithm with anti-fuzzy properties was proposed.The Speed-Up Robust Features(SUFR)algorithm and Binary Robust Invariant Scalable Keypoints(BRISK)descriptor were used to detect and describe the feature points in the target region,and the target position in the fuzzy image was obtained by the second matching of the feature points.In addition,the scaling factor was introduced,and the optimal scale of the target was obtained by comparing the peak response of the target under different scales.The tracking results show that the designed tracker can solve the problem of target blur and scale change well,and it can be tracked in real time.