A Study Of Correlation Filtering Tracking Methods Incorporating Timing

Correlation filtering is one of the methods for single-target video tracking,which is fast,accurate and has high compatibility.However,its simple framework is also accompanied by performance degradation when scenes such as occlusion and severe deformation occur.In recent years,the correlation filtering algorithm usually starts from changing the objective function in order to improve the precision and success rate of trakers,so finding an excellent objective function,i.e.,the correlation filtering framework,has become the mainstream task of the correlation filtering class of algorithms.In this paper,we mainly focus on the research of tracking algorithms from the perspective of proposing an objective function with better tracking performance,adaptive feature fusion strategies and update model.The main research contents are as follows.1.Research for the subject,the basic theory,evolution process of the correlation filter class algorithm under single target tracking,analysis of the advantages and disadvantages of the correlation filter class algorithm,and the most worthy research content of the correlation filter algorithm to find the entry point close to the development trend.2.The impact of adaptive feature fusion and adaptive update strategy on performance under target tracking is studied,and a feature fusion model with the proportion of each feature response is proposed,and a model updating method for adaptively changing filter learning rate by response map frame difference method.The experimental results manifest that the tracking performance can be improved based on the original algorithm,specifically: the success rate on the OTB100 dataset is better than the original algorithm.In scenes with complex backgrounds,the precision is improved by 4.6% and the success rate is improved by 2.6% compared to the original algorithm;the success rate is improved by 10% when fading out of the field of view;and the success rate is improved by 2.3% when the scale changes.Meanwhile,the algorithm outperforms BACF and the original algorithm in in-plane rotation and complex background environment.3.For correlation filtering in tracking accuracy is high but there is room for improvement,from the method of changing the objective function on the basis of ARCF proposed a tracking algorithm with a new objective function that introduces a doublebounded regular term from the timing perspective.The proposed algorithm is tested on three datasets,OTB100,DTB70 and TColor-128,and the mainstream and latest algorithms with better performance are selected for comparison.The experimental results indicate that on OTB100 dataset,the success rate of the proposed algorithm outperforms ECO,ARCF,BACF,CACF,etc.,and improves 2.8% compared with ARCF;in terms of accuracy,it improves 1.6% compared to ARCF.On the UAV dataset,the success rate and accuracy of the algorithm in this paper outperform ARCF,LADCF,STRCF,ECO,BACF,CACF,etc.on the UAV DTB70 dataset;on the TColor-128 dataset,the success rate and precision of the algorithm in this paper outperform ARCF,Autotrack,BACF,AMCF,Staple,etc.The success rate is 2.5% higher than that of ARCF,3.1% higher than that of Autotrack,and 8.0% higher than that of BACF;the accuracy is 2.7% higher than that of Autotrack,3.8% higher than that of ARCF,and12.8% higher than that of BACF.In addition,the experiments show that the proposed algorithm can improve the problems caused by the background complexity.To further prove this conclusion,the proposed algorithm framework is ported to the LADCF,and the experimental results manifest that it can solve the background complexity problem to a certain extent indeedly.Subsequently,the robustness of the algorithm is analyzed,and the TRE index of this paper is better than that of ARCF on many datasets.