Research On Target Tracking Based On Improved Correlation Filtering

As a hot research direction in recent years,target tracking has attracted the attention of many researchers.It has a very wide range of applications in many aspects,such as auto driving,security monitoring system,human-computer interaction,military field.However,there are many challenges in the actual video tracking scene,such as the fast motion of moving objects,the occlusion of objects by other objects,similar background effects and so on.It is precisely because of so many difficult scenes that it is very difficult for us to obtain a reliable and stable tracker.At the same time,it also attracts the interest of researchers.In these numerous research directions,the tracking method based on correlation filtering has achieved good results in dealing with various challenging scenes,but it also has many deficiencies to be studied and improved in the face of complex real scenes.In this paper,through a detailed study of the advantages and disadvantages of the correlation filtering algorithm,a multi-scale feature fusion spatiotemporal regularization correlation filtering tracking method and a multi-tracker correlation filtering tracking method are proposed as follows:(1)Aiming at the problems of scale change and boundary effect in target tracking,a multi-scale feature fusion spatio-temporal regularization correlation filter tracking is proposed.In the actual tracking scene,we often encounter the change of the target’s own scale caused by the target’s moving far and near,which makes the tracking performance decline.Therefore,we propose a scale adaptive target tracking algorithm.In the aspect of the target’s apparent features,we have carried on the fusion processing of various features on the basis of the correlation filtering algorithm to improve the robustness of the tracking algorithm in various scenes.To solve the boundary effect problem in the correlation filter tracking algorithm,we adopt the regularization method in time and space.By reducing the influence of the target background in the actual scene,the trainer can obtain more real samples for training.In order to solve the objective function of correlation filter quickly,ADMM is used to improve the operation speed and the tracking performance of the algorithm.Finally,for the model updating of the whole algorithm,an interval updating module is added,which can reduce the drift phenomenon of the tracking target to a certain extent,and effectively reduce the computational load pressure caused by each frame updating of the model.(2)To solve the problem that a single tracker can not meet the tracking requirements of various complex scenes,a multiple tracker correlation filter tracking is proposed.By flexibly switching different trackers to deal with different tracking scenes,the problem of single tracker instability is effectively solved.This decision level fusion method can effectively enhance the robustness of the tracking algorithm.The multiple tracker tracking algorithm keeps the tracking clues of each expert,and selects reliable experts to track through self-evaluation and mutual evaluation mechanism.At the same time,multi-scale change strategy and interval update model method are applied to improve the tracking effect of the algorithm.