Research On Object Tracking Algorithm Based On Correlation Filter

Object tracking aims to based on the bounding box given in the first frame obtain target positions in subsequent frames,and furthermore generate trajectory of targets in the entire video sequence.With the advancement of artificial intelli-gence and increase of human needs,object tracking has achieved wide applications in the fields of video surveillance,human-machine interaction,unmanned driv-ing,robotics and so on.Correlation filter-based tracking methods have attracted much attention,owing to their fast speed and high accuracy.However,limited by such factors as appearance changes,motion blur,similar background interfer-ence,occlusion,and fast motion in complex tracking scenes,object tracking still faces many problems in actual applications,and its accuracy needs to be further improved in the case of interferences.Under the correlation filter framework,this paper focuses on the problems of model updating,feature fusion,multi-scale search,and combination with deep learning methods.The main research contents are as follows:? A robust object tracking algorithm with background awareness is proposed to solve boundary effect problem caused by cyclic shift of samples while improving tracking accuracy.In the proposed tracking algorithm,the samples are generated in real background.This sampling method not only alleviates boundary effect problem,but also enlarges object search area and enhances discriminative ability of the model.The complementary histogram of oriented gradient feature and color name feature are extracted to fully express visual information of tracking objects and significantly enhance accuracy and success rates of proposed tracking algorithm.Meanwhile,a multi-scale search strategy based on scale pyramid is designed to solve the problem of scale changes during the tracking process.The effectiveness of proposed tracking algorithm is exhibited by comparing with other advanced algorithms on public datasets.?An improved object tracking algorithm based on hierarchical convolu-tional features is proposed to solve the problems of feature extraction,model updating and target re-detection while realizing accurate and fast tracking.In the improved tracking algorithm,the different convolutional layers are exploit-ed to extract features of targets.In this way,the improved tracking algorithm can not only achieve accurate target positioning,but also separate targets from noisy background.Besides,in order to improve frame-by-frame model updating method,a reasonable response threshold is set to evaluate the degree of occlusion for targets.When model response values are high,tracking results are enough reliable and a more convenient and effective model updating method is adopted;when model response values are low,tracking results are unreliable and the target re-detection mechanism needs to be activated.The comparative experiments on public datasets show the effectiveness of proposed tracking algorithm.The paper concludes with a summary for the research work and makes expec-tations for the future direction of object tracking.