Research On Single Target Tracking Algorithm Based On Correlation Filter

With the development of computer hardware and software,camera,cloud storage and artificial intelligence,Target tracking technology is becoming more and more widely studied.Since the its emergence in the last century,a large number of target tracking algorithms have developed,which have achieved good results to a certain extent.However,due to the complexity of tracking background,obtaining an efficient and robust tracking algorithm remains an important and challenging task.The performance of the target tracking algorithm is usually determined by the accuracy of the tracking and the tracking rate.Most of the algorithms can not achieve the desired effect between them at the same time,especially the tracking rate the reduction when a certain tracking accuracy is satisfied.However,with the increasing application of target tracking technology in life,tracking algorithms with real-time performance are beginning to be paid attention to.The traditional Kernelized Correlation Filter is characterized by fast tracking;this thesis mainly analyzes and the algorithm existing problems and optimizes it.The main work has the following aspects:(1)Focusing on the issue that using a single image feature tracking target has poor tracking accuracy and robustness in the complex environment,a kernel correlation filter target tracking algorithm based on features fusion is proposed in dense sampling framework.In the process of tracking,the target appearance features are described by using the histogram of oriented gradient and the color name attribute.The filter coefficients and target feature templates obtained by the ridge regression classifier are respectively used to detect the candidate samples obtained by cyclic shift.The classifier response values are linearly weighted by fusion,and the position of the current frame target is determined by the maximum response value.(2)Focusing on the issue that the correlation filtering algorithm is difficult to adapt to the change of the target scale accurately in real time,a target tracking method with adaptive variable scaling is proposed.On the basis of the cyclic structure of the traditional correlation filter,estimating the target scale is added.Firstly,the correlation filter is used to estimate the target position,and then a scale estimation method is used to train a one-dimensional scale filter.Based on the size of the previous frame,multiple scale samples are obtained at the determined target position,the scale of the current frame target is determined by the maximum response value to achieve adaptive variable scale target tracking.(3)Focusing on the issue that the real-time model updating does not consider the credibility of the tracking result,an occlusion detection updating method is introduced into the traditional correlation filtering algorithm.When the target is severely obstructed,if the tracking algorithm still introduces the appearance information of the current frame target into the updated model,the tracking error caused by the interference information will continue to accumulate.By analyzing the two-dimensional distribution of the detected values before and after the occlusion,the distribution of the detected values in a certain area threshold is used to judge the target occlusion.According to the difference of target occlusion,the difference between target information and interference information in the tracking frame obtained by the feature fusion algorithm and the scale adaptive algorithm,different occlusion detection and update methods are adopted.