Research On Target Tracking Algorithm Based On Kernelized Correlation Filters

Now,Moving target tracking is one of the important research topics in the field of computer vision.The digital video image collected by the camera is processed frame by frame to realize the positioning and tracking of one or some regions of interest in the video.In recent years,with the cooperation of a large number of researchers,more and more excellent visual target tracking algorithms have emerged.This technology has also been widely used in many scenarios.However,the complex tracking environments such as lighting and occlusions,and the deformation of the target itself still bring many challenges.Therefore,researching algorithms with strong robustness and high accuracy is still the current key task.Correlation filtering tracking algorithms stand out from many visual target tracking algorithms due to their fast tracking rate,high accuracy,and strong robustness.They have received extensive attention and research in recent years.In view of the common problems in the tracking process,this paper starts from the traditional correlation filtering algorithm,analyzes its advantages and disadvantages,and proposes corresponding improvement strategies for the deficiencies and performs comparative simulation experiments to verify its effectiveness.The main work and innovations of this paper are as follows:Firstly,aiming at the problems that the target is easy to be lost when it encounters occlusion and the error of the target model will continue to accumulate over time,a tracking algorithm with anti-occlusion and high stability is proposed.The current state of the target is determined by calculating the PSR value of the response graph firstly.If the occlusion area is large,the template parameter update will be stopped,and simultaneously starts up the occlusion detector to relocate the target position using the characteristics of bidirectional optical flow.The results show that although the speed of this paper is reduced,it can still meet the real-time performance,and at the same time,it is found to be efficient in the corresponding environment of occlusion and lighting.Secondly,in order to solve the problem of scale estimation caused by the change of the target size encountered during tracking,an adaptive scale tracking algorithm based on contextaware framework is proposed.The algorithm decomposes the target tracking task into two large blocks,and trains and learns the optimal target position and scale through separate regularized least squares classifiers.In solving the position problem,the target and background are sent to the filter as positive and negative samples to learn to enhance the discriminative power.In the scale filter,coarse scale pool interpolation are used to speed up the calculation.From the experimental results,the algorithm in this paper has better robustness under fast-moving,target deformation and other environments compared with the relevant filtering algorithms that solve the scale.In this case,the real-time performance of the algorithm can also be satisfied.Thirdly,in order to solve the problem the HOG feature description of the target from a single angle,and when a variety of environments are mixed,it may fail,a weighted feature fusion tracking algorithm is proposed.Using Bayes' theorem to calculate color features,the obtained probability model is used to distinguish foreground objects and background regions.Finally,the confidence maps of the two features are calculated and assigned weights by APEC,and the two tracking results are fused at the decision layer to form a complementary learning method to enhance the adaptability of the filter to appearance.From the experimental results,the algorithm in this paper performs well in terms of accuracy and real-time performance when dealing with similar targets and deformation.