Target Tracking Network Based On Dynamic Anchor

In recent years,many excellent target tracking algorithms have been proposed by researchers,and have made great progress in the popular data sets.However,most of these methods focus on target classification,but the target estimation has not been paid enough attention.Most target tracking networks estimate the target scale in the following ways: multi-scale test,preset fixed scale region proposal network(RPN),reverse iterative optimization.These target estimation methods do not consider the potential information between adjacent frames.To better estimate the target scale,the main contents and innovations of this paper can be summarized as follows:First,to fully consider the relationship between adjacent frames,this paper proposes a dynamic anchor-based target tracking network.Different from the way of preset fixed scale RPN,this paper uses the scale of the previous frame as the initialization scale of the current frame anchor,so as to make full use of the scale information between adjacent frames and predict the new scale.Second,because each anchor has no fixed initialization box,the network cannot learn the semantic range of each anchor.To solve this problem,this paper proposes a boundary support module.This module maps the features of different initialization boxes to the corresponding anchor so that the network can learn the corresponding semantic range of each anchor,which is complementary to the dynamic anchor scheme.Third,because the target center predicted by DCF is a rough position,to further improve the positioning accuracy,this paper adds an offline center correction branch.Based on DCF Prediction Center,further,adjust the center position,get a more accurate target center,and then improve the performance of target tracking.To verify the effectiveness of the methods mentioned above,this paper evaluates the algorithm on five main target tracking datasets.Among them,UAV123,La SOT,and Tracking Net use one pass evaluation(OPE)to evaluate the tracker.VOT2018 and VOT2019 use the expected average overlap(EAO)method to evaluate the tracker.Experiments show that the performance of target tracking can be effectively improved on five popular datasets.Among them,the best result is obtained on the data set La SOT with smoother adjacent scale changes.Besides,to prove the contribution of the dynamic anchor,boundary support module,and central modification branch in the final performance,this paper has carried out sufficient experiments on different modules.