A Target Tracking Robust Algorithm Based On CSK

In recent years,the target tracking technology for video frame images has become a hotspot in computer vision research.The method of tracking by detection is becoming a popular.Machine learning ideas are applied to model updates.The target tracking algorithm based on kernel correlation filtering theory has achieved outstanding results in tracking accuracy and real-time performance.However,uncontrollable factors from the complex real environment such as dramatic changes in light,target scale change,target occlusion and interframe large displacement are still the difficulties and challenges of target tracking research.For the above problems,this paper makes an in-depth study,and based on the Circulant Structure Kernels(CSK)tracking algorithm achieve the following results:(1)Through the research and analysis of image features,the multi-channel features are summed in the frequency domain by the kernel function to realize multi-channel expansion of CSK tracking algorithm.Finally,the algorithm can adapt Histogram of Oriented Gridients(HOG),Color Name(CN),Local Binary Pattern(LBP)and so on better image features.These enhance the ability of the algorithm to describe the target,and weaken the effect of optical changes and geometric changes on target tracking.(2)The image pyramid is constructed by transforming the source image,then the HOG features are extracted from every level to build the pyramid sample set.Finally,training Pyramid Kernel Correlation Filter(PKCF)to detect the target scale,and adjust the scale of the tracking rectangle and the sampling window according to the new scale of the target to reduce the error accumulation of the target model,improve the target tracking accuracy,and complete the CSK scale adaptive improvement.Of course,PKCF is also the target position detector.(3)The Kalman filter is introduced into the CSK tracking flow,which makes full use of the target movement state information to predict the location of the target in the next frame preliminarily.And then using PKCF for position calibration and scale detection to achieve target detection adaptive.Finally,improving the defects of CSK tracking algorithm:the current frame target detection area is fixed near the target center position of the previous frame.And solving the problem of target occlusion and interframe large displacement.(4)For Kalman filter and PKCF updates,combining offline updates with online updates to enable adaptive updates of target models and classifier parameters.First,using the target models and classifier parameters with good tracking effect to establish an alternative.When the tracking accuracy drops or the target is occluded,replacing the current online target model and classifier parameter with the alternative for offline update.The status input of Kalman filter for position prediction of the current frame is PKCF on the previous frame to obtain the correct location.That is to say,the current frame Kalman filter is updated with the previous frame PKCF output.(5)Combining scale adaptation,detection adaptation,and update adaptive ideas with occlusion mechanisms to propose the final algorithm of this paper:target tracking robust algorithm based on predictive-calibration-updating.Finally,selecting several groups of video with different challenges such as illumination change,scale change and target occlusion from the standard test set VOT and live shot of the video set to test this algorithm.The experimental results of this algorithm and CSK algorithm show that this paper successfully achieves the scale adaptive improvement,to a certain extent,solves the problem that the target is completely occluded and the large displacement between frames,in addition,tracking accuracy and success rate also increased significantly.This algorithm is compared with CSK,KCF,CN,MOSSE,TLD,Struck algorithm in overall performance,the results show that the this algorithm has the best performance in the center position error,tracking accuracy and success rate,but the performance of the track rate is not enough.