Correlation Filter Tracking Based On Robust Template Matching

Visual target tracking technology is one of the most important research branches in the computer field,which has been widely applied to many fields,such as unmanned aerial vehicle,automatic driving,intelligent transportation and medical diagnosis.In recent years,a great deal of excellent visual target tracking algorithms are proposed based on correlation filtering algorithms.However,target tracking algorithms still faces many challenges in the actual tracking environment,such as scale variation,rotation and occlusion.How to solve the robust target tracking problem in complex scenes is still a difficult problem.In this paper,aiming at the problem that the Kernelized correlation filters(KCF)algorithm is easy to track failure in the case of the target scale variation,occlusion,fast motion and background noise,the correlation filter tracking algorithm based on robust template matching is proposed in this paper.The main contents and the result of the paper are as follows:(1)Aiming at the problem of tracking failure in KCF tracking algorithm under the condition of target scale variation,occlusion and fast motion,the kernelized correlation filter tracking algorithm via Best-buddies similarity is proposed.Firstly,the target location estimation module and scale estimation module are designed respectively.In the location estimation module,background information and target information around the target is used to train the classifier,and uses multiple features to construct the target appearance model,which can improve the classification ability of the classifier.In the scale estimation module,the algorithm only uses target information to train the classifier,and the optimal scale is selected by the classifier.Secondly,according to the target response Peak-to-Sidelobe(PSR)and the confidence of the target appearance model,the algorithm discriminates whether the target is occluded.When the target is occluded,the target re-detection is carried out by the Best-buddies similarity matching algorithm,and the problem of the target re-location is solved.Finally,the adaptive template updating strategy is used to solve the problem of the target template drift under occlusion.(2)Aiming at the boundary effect existing in the KCF tracking algorithm,the spatial regularization correlation filtering tracking via deformable diversity similarity is proposed.Firstly,the spatial regularization weight is introduced on the basis of the KCF algorithm,and regularization weight is set according to the spatial location information of training samples,which can weaken the influence of boundary effect and background interference.Secondly,using Gauss-Seidel iterative optimization methods to solve the classifier parameters,and the complex optimization problem in ridge regression is converted to the real number optimization problem,which reduces the computational complexity of the algorithm.Thirdly,the target re-detection module is constructed by using deformable diversity similarity matching algorithm,the nearest neighbor search problem in the matching algorithm is solved by the Tree CANN algorithm.Finally,the continuous detection in the time domain is realized by the triangular polynomial interpolation,and the maximum response position of the target is obtained by the Newton's iteration method.(3)100 video sequences and evaluation standard in OTB-100 dataset are used to test the performance of the algorithms.The algorithms are qualitatively and quantitatively analyzed respectively,and the performances of the proposed algorithms and other popular algorithms in11 video attributes are tested.Experimental results show that the precision score and success score of the kernelized correlation filtering tracking algorithm via Best-Buddies similarity are respectively 0.775 and 0.570,which is respectively improved by 11.4% and 19.5% compared with the KCF algorithm.The precision score and success score of the spatial regularization correlation filtering tracking via deformable diversity similarity are respectively 0.825 and0.625,which is respectively improved by 18.5% and 31.0% compared with the KCF algorithm.Compared with the KCF algorithm,two improved algorithms in this paper can better solve the tracking problem of the target scale variation,occlusion,fast motion and background noisy,which has a wide range of application prospects and practical value.