Research And Implementation Of Object Tracking Method Based On IoU-Aware Siamese Network

Object tracking is an important research branch in computer vision,and it has been successfully applied to surveillance,human–computer interaction and intelligent transportation.Recently,several Siamese network-based trackers have drawn great attention due to their simple network structure and great performance.These methods use a large number of training data to train the tracking model offline,and then use the trained model to track the object online.Because the model is not updated online,these methods achieve a good balance between accuracy and speed.However,for the Siamese networkbased trackers,the classification and regression branches are relatively independent during the training phase,but the classification score is utilized to select the final box directly in the inference phase.The low correlation between the classification score and localization accuracy could lead to performance degradation.In addition,the existing Siamese networkbased trackers utilize fixed thresholds to obtain the positive and negative samples,which are not conducive to obtain superior performance model.In this thesis,an IoU-Aware Siamese Network is proposed.Main works are as follows.1.To address the problem that the correlation between the classification score and localization accuracy is weak in the existing Siamese network-based trackers,so this thesis proposes an IoU-guided siamese region proposal network.And in order to improve the robustness of the model,an IoU predictor is proposed to correlate the classification and regression branches in Siamese network-based trackers.Besides,an IoU-guided localization strategy is proposed to improve the localization accuracy.2.To address the problem that siamese trackers use fixed threshold to select positive and negative samples,which will lead to reduced tracking performance.In this thesis,an IoU-constrained siamese region proposal network is proposed to improve the robustness of the model.In detail,the fixed threshold is replaced by dynamic threshold.In addition,the IoU quality assessment branch is used to replace the classification branch,which could ensure the performance of the tracker and reduce the model parameters.