Single Object Tracking Research Of The Feature Points Matching And Siamese Framework That Based On Optical Flow Information

With the rapid development of computer vision,the single-target tracking problem,which belongs to the field of visual tracking,has attracted more and more attention from the public.The development of tracking algorithms has been improved from the Kalman filter?particle filter and feature point matching generator models to discriminant models based on the correlation filter framework and the Siamese framework in recent years.The advantage of the algorithm based on feature point matching is that simple structure and no training process.However,there still exist some problems of low precision and disappearing of feature points during occlusion in these algorithms.The full-convolution network algorithm that based on Siamese framework has the advantages of high speed.However,it only considers the appearance characteristics,so it is difficult to track objects with intricate backgrounds and strenuous movements.In view of the above problems,this paper proposes feature point matching method based on optical flow information and single-target tracking on the basis of Siamese framework.Details are as follows:1.A method of virtual feature points and fuzzy weights is proposed.On the one hand,the tracking algorithm based on feature point matching is the most important implementation of the traditional model,but performance of algorithm is bad when the target is occluded.Therefore,based on the CMT algorithm,a method,with geometric uniformity,that supplements virtual-feature points is added to compensate for the lack of feature points when the target is occluded.On the other hand,the optical flow information of object is an important representation of the motion information,which has the advantages of high phase reliability and high robustness to illumination changes.Therefore,the triangulation method of fuzzy theory is proposed to balance the feature point matching frame and the optical-flow tracking frame.Last but not least,the tracking accuracy of the algorithm during occlusion is improved without reducing the tracking rate.2.A timing scoring model is proposed to integrate the optical flow network into the framework of the tracking algorithm.Although the Siam FC is one of the mainstreamsingle-target tracking algorithms,which proposes a double-way input and mutual convolution method for tracking.This paper integrates rich optical flow information into the tracking network based on the Siamese framework: using bilinear interpolation to fuse optical flow information to different video frames,in the meanwhile,integrating the different candidate detection frame by sequential scoring model.And above all,it solves the problem that difficult to track the blur target under severe motion through the application of optical flow network.3.A discriminator attention model is proposed.Under the Siamese tracking framework,the contribution of the template frame to each position on the discriminant response map is similar,making it possible to get the wrong response peak and loss the target.Therefore,three different attention models are presented in this paper,which are optical flow attention model,the normalized attention model and the hourglass attention model respectively.At the same time,the paper compares their effect on the accuracy of the tracking algorithmThe algorithm of this paper achieves real-time tracking by combining the advantages of off-line training and online tracking in deep network structure.In order to improve the accuracy of the algorithm,the algorithm trains on ILSVRC15.At the same time,the experiments are conducted on the 0TB(Object Tracking Benchmark)and VOT(Visual Object Tracking)tracking data sets,which achieve good results.The experimental results show that the proposed algorithm can follow targets efficiently and is robust to occlusion,strenuous motion and vanishing objects.