Object Tracking Based On Kernel Correlation Filter

In the era of intelligence,computer vision technology solves the problem of machines "seeing" the world.As a beam of light that computers communicate with the world,target tracking technology is one of the core technologies built into the spindle.At present,target tracking technology has important applications in the fields of visual navigation systems,intelligent video surveillance systems,medical diagnostics,mobile robots,and military.Although the current tracking methods can meet the requirements of machine vision to a certain extent,in the tracking scenarios where there are challenging factors such as occlusion,motion blur,and fast motion,the existing tracking methods still have difficulty in successfully tracking the target robustly.In order to improve the comprehensive performance and anti-occlusion ability of the tracking method,this paper conducts research based on the current mainstream method of kernel correlation filter tracking algorithm.The main research work is as follows:This paper first analyzes the research background,related applications and research significance of target tracking,introduces common difficult problems in tracking scenarios,sorts out the current research status of target tracking,and elaborates the theoretical basis of kernel correlation filter methods.Discuss the limitations of the kernel correlation filter tracking algorithm,analyze the tracking confidence problem,and propose a joint confidence to observe the tracking state,and use the joint confidence as a switch for anti-disturbance execution strategies such as occlusion.When the joint confidence reflects that disturbance such as occlusion occurs in the tracking scene,a robust tracking strategy is further designed.Aiming at the problem of lack of background information learning in the kernel correlation filter tracking algorithm,a context-aware weighted learning method is proposed,and the context background during tracking is divided into pure background areas and target similar background areas by joint confidence.For different context regions,construct different objective functions and solve their closed solutions.When the joint confidence reflects that there is no disturbance in the tracking scene,the context background is divided into pure background areas,and the background is zeroregressed to train the filter.Otherwise,the context background is divided into background areas similar to the target,and the context background is weighted regression to train the filter.The proposed method is applied to the correlation filter benchmark algorithm and analyzed on the OTB100 dataset.The tracker applying the method in this paper has higher tracking accuracy.Aiming to the challenge of occlusion during tracking,a context-aware antiocclusion target tracking method is proposed.Based on the kernel correlation filter tracking principle,the dimension of the feature is extended to construct a robust target appearance model,and the size of target is estimated during the tracking process.The joint confidence is used to judge whether occlusion occurs,and a corresponding antidisturbance tracking strategy is designed.If the occlusion does not occur during the tracking process,the kernel correlation filter tracking is performed.Otherwise,redetection is introduced to locate the target position,and the region corresponding to the re-detection is added to the regulation term of the kernel correlation filter for context learning.The fusion of the filter template before occlusion and the context model learned during occlusion is used to locate the target and to update the model.Experimental evaluations on the datasets OTB2013,OTB100,TC128 demonstrate that compared with the state-of-the-art algorithms,the proposed algorithm has stronger robustness and higher tracking accuracy when occlusion occurs.