Research On Vision Object Tracking Based On Deep Learning

Target tracking has made remarkable progress in recent years.The depth neural network model optimized to recognize objects in images shows unprecedented task training accuracy and good generalization ability.In this paper,the twin network in-depth learning method is used to study wooden plaque tracking.The specific research contents are as follows:(1)According to the different methods of feature extraction,the algorithms are divided into two categories: the artificial design feature tracking algorithm and the deep learning tracking algorithm.Their method ideas and model structures are analyzed in detail.Then choose the most popular data set to evaluate and compare the algorithms and analyze the reasons for the different differences of tracking results caused by different methods.It is convenient for researchers to choose the appropriate methods for their desired performance,so that readers can clearly understand the broad prospects and research significance of the application of deep learning methods in target tracking and computer vision.(2)Target tracking algorithm based on deep learning has many advantages over target tracking algorithm based on artificial design features,but it also has some shortcomings.For example,the lack of supervised training data in deep learning and the poor real-time performance of deep learning algorithms are common problems.In order to solve these problems,a universal tracker is trained by twin network structure so that tracking does not need to be trained in the face of different videos.Relevant filters are used instead of SGD to speed up tracking,and relevant regular terms are added to correlative filters to optimize their performance.In OTB benchmark,the improved performance gain is obtained by comparing with the benchmark frame tracker,and then compared with other trackers using artificial design features and depth features,it can be concluded that the use of depth features can significantly improve the tracking performance compared with the tracker using artificial design features.(3)Target detection technology is introduced,and target tracking is carried out by using target location method.Generally,candidate blocks are randomly placed around the previous target location and estimated target location,which results in a lot of redundancy when extracting features.A new method to accelerate the tracker operation based on convolution neural network is proposed in this paper.The whole region of interest is input intoconvolution neural network once to eliminate the redundancy operation of random candidate blocks.Firstly,each candidate block is classified as object or background,and then the target block is separated as rough location.Finally,bilinear interpolation of convolutional neural network features is used as fine location.The experimental results on OTB100 show that it not only keeps the same accuracy and robustness as the top-level tracker,but also achieves several times of acceleration.(4)The latest target tracking work integrates the advantages of both tracking and detection methods.Fusion of generative and discriminant methods significantly improves the tracking ability of the tracker.In this paper,a new method is proposed,which combines the advantages of discriminant framework and generative framework,to meet the requirements of tracking deformable and occluding targets effectively.At the same time,the regional recommendation optimization(RPO)method is introduced to remove unqualified regional recommendation,which greatly improves the positioning accuracy.Dynamic modeling is used to simulate the change of object appearance caused by motion and occlusion,so that the algorithm can update the object model adaptively,thus improving the robustness to the change of object appearance caused by high motion and occlusion.