Siamese Object Tracking Algorithm Combined With The Intersection Over Union

Object tracking is a basic research direction of computer vision,which is commonly used in application scenarios such as automobile autonomous driving and traffic monitoring systems.With the rise of deep learning,various algorithms with excellent performance have been proposed one after another,Siam RPN is one of them.Siam RPN draws on the detection module in the target detection task,proposes a siamese region proposal network,and realizes the foreground and background classification prediction and the bounding box regression prediction.However,the detection module of Siam RPN uses the L1 norm loss function,it does not consider the intersection over union(Io U)relationship between the prediction box and the true value,so the prediction of the bounding box is rough,and it is prone to deviation in the interference scene.To solve this problem,based on the improvement of Siam RPN,this thesis proposes a siamese network object tracking algorithm combined with Io U.The specific innovations are as follows:(1)Because of the Io U relationship is not considered in regression prediction,the prediction accuracy of bounding box is not enough.In order to solve this problem,the IOU loss optimization regression prediction is proposed.Due to the large number of parameters whitch are unevenly distributed in Siam RPN,it is difficult to optimize the network.If the Io U loss is directly used as the regression loss function,the network can not converge.Therefore,the Io U & smooth L1 joint optimization strategy is designed,whitch apply the Io U loss to revise regression prediction of the algorithm.The joint optimization strategy of Io U & smooth L1 selects the sum of the Io U loss of the best matching positive sample and the L1 loss of the other positive samples as the regression loss,and makes up for the insufficient training problem of calculating the Io U loss from a single sample by random displacement sampling of the search image.(2)The classification of positive samples is sampled in the same object,which makes the similar positive samples lack enough distinction,so the algorithm is difficult to evaluate the classification reasonably,and is easily affected by the prior parameter setting.In view of this problem,the paper proposes the strategy of Io U weighted classification.Based on the prediction results of regression branch,the Io U of prediction box and ground-truth box is calculated,and it is used as the weight of the positive sample of classification,so that the more accurate positive sample can get more weight in classification.(3)Due to the lack of correlation between the regression branch and the classification branch of Siam RPN,the optimal results of the prediction outputs of the two branches may not match in the tracking process,which makes the algorithm prone to drift in the interference scene.Therefore,the regression branch is associated with the classification branch by using Io U as the weight.In the Io U weighted classification,because the classification prediction refers to the weight calculated from the regression prediction results,the two branch prediction has relevance,which further improves the stability of the algorithm and makes its performance better than Siam RPN under interference conditions.This thesis compares Siam Io U with Siam RPN and other mainstream algorithms on the OTB2013,OTB2015 and VOT2018 tracking benchmark.The results show that our algorithm can effectively rise the tracking performance and has stronger antiinterference ability in certain tracking scenarios.