Research On Discriminative Siamese Tracking Network Integrated With Online Update

Object tracking is a branch of computer vision and widely used in autonomous driving,drones,intelligent robots,and intelligent security.Recently,the siamese tracking network has become one of the best tracking algorithms by balancing speed and accuracy well.However,these trackers still have shortcomings that need to be solved urgently.Like the failure to make good use of context information,inability to handle the deformation and occlusion of the target,and less precise in bounding box regression,etc.To solve the problems mentioned above,the main contributions are as follows:(1)Multi-classifier guided discriminative siamese tracking network with online classifiers is proposed to exploit video-specific context information.The proposed algorithm is consists of four parts: siamese tracking network with offline training,online classifier network,template update and bounding box selection mechanism.Combined with online and offline training,the algorithm has a strong discriminative and robustness.(2)Cascaded siamese region proposal network with an online classifier is proposed to deal with deformation,occlusion and lack of robustness.The algorithm is based on Siam DW and weights the classifiers scores in cascaded siamese tracking network.then,the algorithm introduces a novel sample selection mechanism to improve the robustness.Finally,the online classifier is integrated to further improve the discriminative and the ability to handle the deformation and occlusion.(3)Siamese dense local regression network is proposed to fully use local information.The algorithm aims to realize the local proposal region bounding boxes regression,and further improve the accuracy of the baseline tracker.The designed module is generally applicable to anchor-based or point-based tracking algorithms.(4)Point set representation for siamese tracking network is proposed to enhance regression accuracy.The algorithm replaces the siamese region proposal network with a point set network,it enhances the discrimination of the classifier by introducing an elliptical region for sample selection.A multi-level point set augmentation is used to improve the accuracy of bounding box regression.Finally,an online classifier module is applied to further improve the performance of the tracker.