Research On Cross-view Object Continuous Tracking

This thesis mainly focuses on cross-view object continuous tracking problem,and divides it into two parts: object tracking and object re-identification.Furthermore,this paper proposes a robust single object tracking algorithm and an effective object re-identification method to tackle single-view tracking troubles and cross-view object re-identification,respectively.High-performance Siamese network is adopted and combines correlation filters to achieve fast,effective object tracking in single object tracking.To handle the problem of model drift in the tracking process,a tracking reliability criterion and a model adaptive updating method are proposed.Whether the model is updated or not is determined by the reliability criterion and associated with the tracking weight.Furthermore,the occlusion situation is solved by a robust part-based tracking method which uses motion vector of reliable parts and their distances with target center to vote for the center and target size of the object in the new frame,respectively.Additionally,a SVM detector is trained to further improve the tracking performance and a detector-tracker parallel mechanism is proposed to guarantee the accurate and real-time tracking.Also,this paper introduces a cross-view asymmetric metric learning method,which gathers samples in different camera views more closely by learning projections of different camera views and finding shared feature space of objects in different camera,to avoid the effect of different camera view in the object re-identification process.Then,a Siamese network is used to associate objects in different views to achieve object re-identification.Abundant experiments on public object tracking and object re-identification datasets show that the proposed tracking algorithm and cross-view object re-identification method are effective and robust,and further prove the feasibility and effectiveness of the whole continuous tracking system.