Research Of Correlation Filter Tracking With The Integration Of Context Information

Visual object tracking is an attractive research topic in computer vision with plenty of applications,such as video semantic analysis,Human-Computer Interacting and etc.Although many trackers have been proposed in recent years,realizing robust tracking still remains challenging due to factors such as deformation,occlusion,rotation,and illumination variation.This thesis specifically focuses on correlation filter tracking.We propose two correlation filter based trackers by exploiting spatial and temporal context of the target,respectively.The contributions are as follows:First,the recent correlation filter based trackers exploit single attribute features and unable to estimate target scale.Therefore,such trackers tend to drift in cases of deformation and illumination variation.To tackle this problem,we exploit spatial context of the target and realize a correlation filter tracking algorithm by utilizing fusion feature and Bayesian model.We propose a multi-attribute fusion feature in this thesis.The feature is introduced into correlation filter to improve the robustness of the model.We further investigate the problem of scale estimation and establish a Bayesian probabilistic model to identify potentially target region.The scale prediction is determined by subsequently spatial analysis.Moreover,we also conduct reliability test to update the Bayesian model adaptively so that the model corruption is suppressed.Second,as the appearance information is lacked where the target undergoes occlusion,the learned correlation filter is corrupted.To solve this problem,we exploit temporal context of the target and realize a refining-by-detection correlation filter based tracker.The tracker conduct position and scale estimation based on translation filter and scale filter respectively.In addition,an explicit confidence correlation filter is employed to measure confidence of tracking results.We also introduce a detector as a detection and correction component.The detector is learned and updated online using the target appearance sampled from the reliable tracking predictions.The detector is employed to refine the position in cases of tracking unreliability caused by target appearance corrupted,thereby reducing the risk of model drift as well as alleviating the tracking failure.Finally,the two trackers are tested on the OTB-2013 datasets.Experimental results show that the spatial-based tracker achieves distance precision of 94.5% and success rate of 79.6% on the 25 test sequences,which is 7% and 4% higher than the second one.As for temporalbased tracker,all the 50 sequences are employed.Compared to other trackers,the distance precision of this tracker is improved by 9.5%,the location error is reduced by 7 pixels as well as the success rate is improved by 13%.For the results of qualitative experiments,the two trackers are robust in cases of deformation,rotation,scale variation and so on.These experimental results show the effectiveness of the proposed trackers.