Correlation Filter Tracking Based On Adaptive Scale And Feature

Visual target detection and tracking is a high-profile computer vision field emerging research direction,and also is the foundation of intelligent monitoring,human-computer interaction and robot vision navigation.Researchers of computer vision also timely put forward some new visual target detection and tracking technology,and obtained a lot of outstanding achievements.In spite of this,there are still many worthy of study direction in the field of target tracking,such as shade problem,lighting effect and deformation of the target,in order to improve the robustness of target tracking,this paper makes a deep research on the problem of the occlusion problem and adaptive scale transformation,the main results are as follows:(1)We put forward a method to deal with adaptive scale transform problem.We process characteristics of samples in advance.Then gain the maximum response characteristics of the scale matrix through the classifier training and finally know the most suitable scale value.can be the most appropriate target dimension value.This method can better realize adaptive changing of the tracking box and improve the accuracy of target tracking at the same time.(2)An adaptive feature selection method is proposed.We extract HOG feature and color feature of targets respectively.And then training the related classifier to get their response value,To adaptive weighted fusion,the final result for target location information automatically assigned its own weight.This method can better deal with the occasion situation.(3)A tracking template update mechanism improvement is proposed.Our algorithm adopts the threshold seted update mechanism.Through to the tracking result credibility decision,making the decision whether to update the existing track model.This can improve the real-time performance in tracking.(4)The above three kinds of algorithm together to form the final algorithm in this paper.An adaptive scale correlation filter tracker with feature fusion is presented.We also improve the template updating mechanism.Extensive experimental results on a large-scale benchmark dataset show that the proposed algorithm performs favorably against state-of-the art methods.