Spatio-temporal Regularization Correlation Filter For Target Tracking

Target tracking is an important research field in computer vision,which is widely used in many fields such as robots,intelligent security,visual navigation,and precision guidance.Although many researchers have made extraordinary achievements in recent years,it's still difficult to design a tracker that can maintain good accuracy and robustness when the target appearance changes significantly.The reason is that complex tracker has better robustness but low efficiency,and the simple tracker has better efficiency but poor robustness.The correlation filter is a simple tracker,so it's necessary to improve accuracy and robustness.An improved correlation filter tracking framework is proposed,which focuses on improving the accuracy and robustness of the correlation filter and uses an engineering method to optimize the calculation efficiency,which specifically researches on the feature extraction,boundary effect problem,and model drift problem.The main contributions of this article are as follows:Firstly,for feature extraction,a tracking target characterization method is proposed,which completes a powerful characterization of the tracking target by extracting manual feature and deep feature at the same time.Among them,manual features are standard FHOG features;deep feature is extracted from a residual neural network that maintains translation invariance.Meanwhile,this network integrates an attention mechanism for the importance of information to discriminate the target and background,it also integrates a pyramid mechanism for the scale to enhance the robustness of the image to scale changes.In this paper,the whole tracking progress is divided into two parts: scale filter and position filter.The scale filter is calculated by efficient manual features,and the position filter is calculated by high-precision fusion features,making the balance of efficient and precise.The improved algorithm has a success rate of 83.7%,70.0%,60.9% and 31.9% on the OTB2015,TC128,UAV123 and LaSOT datasets,respectively,which is higher than the baseline algorithm(Learning Background-Aware Correlation Filters for Visual Tracking,BACF)in the success rate score,which were 7.8%,8.7%,4.2% and 4.2% higher.Secondly,for the boundary effect problem,a new adaptive estimation method of target contour based on spatial regularization is proposed,which alleviates the problem by adding spatial regularization terms.In this paper,the traditional rectangular mask and the spatial regularization have achieved fusion in the optimization algorithm,the iterative spatial regularization term is used as an additional term to allow the filter to adaptively adjust the spatial penalty.When occluded,the tracker gives a larger penalty to the occluded part,and which encourages the acquisitive correlation filter to focus on the reliable unobstructed area.The ablation study based on the OTB2015 dataset shows that after adding the spatial regularization term,the success rate of the algorithm is increased by 2.0%,which is a 9.8% increase compared with the benchmark algorithm BACF.Thirdly,for the model drift problem,a new objective function optimization method is proposed,which alleviates the problem by adding temporal regularization terms.In this paper,temporal regular terms are added to the standard correlation filtering objective function,it allows historical high-confidence samples to constrain the update direction of the correlation filter,which avoiding the mutation of the filter between adjacent frames,and alleviating the influence when the appearance of the target changed significantly.In addition,an update strategy is proposed,which based on the peak sidelobe correlation energy coefficient to complete the update of the temporal regular terms.The peak sidelobe correlation energy coefficient is proposed based on the peak sidelobe ratio,and a complete verification mechanism is designed based on this.The filters that pass the verification mechanism will be recorded and updated through the weighted fusion method to complete the temporal regularization terms.The ablation study based on the OTB2015 dataset shows that after adding the temporal regularization term,the success rate of the algorithm is increased by 0.9%,which is 8.7% higher than that of the benchmark algorithm BACF.Feature extraction,boundary effect,and model drift are deeply studied in the paper,and improvements and enhancements have been made to enable the algorithm to maintain good robustness and efficiency even in complex environments and under target appearance changes significantly.The success rates of the algorithm on the OTB2015,TC128,UAV123 and LaSOT datasets are 86.2%,76.3%,63.6% and 33.0% respectively,which are 10.3%,15.0%,7.9% and 5.3% higher than that of BACF respectively.Simultaneously,The efficiency is maintained at 12.1fps,which has a certain practical application value.