Research On Target Tracking Algorithm Based On Anchor-Free Framework

As a basic task of computer vision,target tracking has attracted extensive attention in recent years.In recent years,the target tracker based on anchor free has attracted extensive attention because it can achieve the same or even better accuracy as the Anchor-Based target tracker without setting a large number of super parameters.However,there are still many problems in the target tracker based on Anchor-Free.This work uses a novel structure to address a series of issues in target tracking based on Anchor-Free ideas.The following are the significant aspects of this paper’s work and innovation:(1)To address the issue of inconsistency between the inference and training phases caused by the algorithm’s shortening the performance gap by introducing extra branches,as well as the issue of prior information introduced by the Centerness strategy used by extra branches.In this paper,we analyze the reasons for the differences formed between AnchorBased and Anchor-Free idea target trackers,and we use a learnable localization quality estimation strategy to unify the extra task branches into the classification task,which not only solves the inconsistency problem between the training and inference phases of Anchor-Free series target tracking networks,but also the classification implications of the Anchor-Free series of target trackers are expanded,and the problem of a priori information caused by the extra branches is also eliminated.tracking networks.(2)To overcome the restrictions of modeling the regression branch as a Dirac δ distribution,this study introduces Riemann integration,allowing the bounding box regression task to learn the distribution of the data samples on its own.This research also makes extensive use of the probabilistic prediction of this bounding box distribution to guide the model’s classification task and improve the algorithm’s performance.(3)To solve the problem that the mutual correlation operation cannot match features with a global perspective,this work offers a lightweight target tracker based on the Transformer idea,which enables the tracker to estimate the target from a global view and enhance the algorithm’s performance.