Research On Target Tracking Method Based On Kernel Correlation Filter

In recent years,with the development of science and technology and social progress,the application of computer vision technology has been becoming more and more important in people's lives,which has most of the applications are based on the application of video surveillance,schools,hospitals,railway stations and other public places have a large number of monitoring cameras,providing a guarantee for the people of the society,in addition,life safety,traffic monitoring has become more and more intelligent,intelligent probe of highway and railway has been deployed to solve many issues and the key events can be automatically record,the application of intelligent monitoring technology mostly rely on the theory of computer vision,computer vision and one branch of target tracking has become a research hotspot in recent years,but in the real scene,object tracking faces many difficulties,with multi-scale target occlusion,illumination changes and the hindering factors change.The purpose of the work is to make some improvements based on previous works,mainly to solve two problems of occlusion and scale change in target tracking process,the specific work is as follows:(1)A target size estimation based on kernel correlation filter KCF(Kernelized Correlation Filters)and a tracking method to deal with occlusion are presented in this paper(SOD-KCF,Scale Estimation and Occlusion Detection-KCF).The target scale estimation model is multi-scale to the current target.a sequence is constructed by calculating the maximum of the classifier response at every time scale,and the maximum value is selected,and the corresponding scale is used as the target tracking scale.In this paper,we establish occlusion processing model according to the distribution characteristics of the maximum response of the classifier and the threshold method for occlusion detection,the target is searched by the block region helix search method after the target is occluded,and the response of the sliding box is calculated in the target search process to determine whether the target is found.The algorithm is tested on OTB(Object Tracking Benchmark)test sequence set and compared with 4 tracking algorithms.The tracking accuracy and tracking success rate are increased by 6.1% and 1.5% respectively compared with the suboptimal method.(2)In order to further solve the scale changes and partial occlusion problem,we proposed a method using a special filter and multi-block processing technique based on kernelized correlation filter.In addition,in order to improve the robustness of the algorithm,this paper build a appearance update model,the model can approximate the change state of the target occlusion and deformation.In addition,this paper proposed an adaptive update learning rate model instead of kernelized correlation filter algorithm originally fixed learning rate to strengthen the robustness of the tracking algorithm for each process.Experimental results show that our algorithm outperforms other comparison algorithms in test sequence set.Especially,compared with kernelized correlation filter tracking algorithm,there are nearly 8% and 18% performance improvement in test sequences with occlusion and scale change.Therefore,this algorithm is more accurate and effective for the improvement of the original algorithm.