End-to-End High-Speed CNN-Based Tracking Algorithm

Visual tracking is an important subject in the field of computer vision,which plays an important role in the application of security,automatic driving and smart robot.The current state-of-the-art tracking algorithms face several problems.One problem is that it is hard to balance the speed and precision.The second problem is that they are not powerful enough to absorb offline training data and hard to make full use big data.The third problem is the lack of effective ways to online update the algorithms,making it unable to track the target under deformation and background change.In practice,tracking algorithms need to run fast and have a high performance.Therefore,the tracking algorithms should balance the requirements of speed and accuracy.At the same time,different scenes will have different training data.By making full use of the advantages of big data,the tracking algorithms can easily and conveniently adapt to different scenes.In order to solve these three problems,this paper proposes the solution.This paper proposes an efficient end-to-end tracking algorithm and designs an elegant and lightweight network structure.Through the training framework,large offline samples can be well absorbed without the need to design complex hand-crafted features.At the same time,this paper proposes a meta-learning method to online update the parameters.It can quickly update network param-eters,and help tracking algorithm adapt to different tracking targets quickly.By utilizing offline training and online update,the potential of the network is fully explored,which increases its knowledge transfer capability and adapt-ability.Besides,it obtains the balance between speed and accuracy.Finally,through the design of scientific and reasonable experiments and the comparison with state-of-the-art tracking algorithms and ablation analysis,the effectiveness of the proposed method is verified.Compared with the state-of-the-art track-ing algorithm,my tracking algorithm has a good performance in the canonical datasets,such as OTB2013[1],OTB2015[2]and VOT2016[3],with the tracking speed of 60FPS which meets the real-time requirements.