Research On Long-term Target Tracking Algorithm Based On Correlation Filtering And Siamese Networks

Computer vision plays a prominent role in daily life and industrial production,and target tracking has become one of the key research directions in this field,and is widely used in security monitoring,intelligent transportation,national defense industry and automatic driving and other fields.In practical applications,the target tracking will face scale,rotation,shade,light changes,challenges,such as deformation and rapid movement,stable and efficient long-term target tracking algorithm as one of the core problems in target tracking,nature is also a major direction of academic research hot spot,it ensures the target tracking in the complex tracking can rapidly and accurately in the scene for target tracking tasks.In recent years,researches on long-term target tracking algorithm in complex scenes mainly involve two directions:long-term target tracking algorithm based on correlation filter and long-term target tracking algorithm based on siamese network.For various problems existing in the tracking process,this thesis makes a detailed analysis and research on the two kinds of algorithms,and does the following work:1.In order to solve the problem of target loss in the process of long-term target tracking,this thesis,based on the basic model of feature complementary learning algorithm,proposes a long-term target tracking algorithm that relies on the mechanism of loss redetection.The feature matching GMS(grid-based Motion Statistics)algorithm is used as the detector to judge the target loss in the process of long-term target tracking,call GMS to re-detect and update the template,so that the tracker can reposition the initial target,and effectively maintain the stability and accuracy of the tracking algorithm.The feasibility of the algorithm is verified by quantitative and qualitative experiments.2.Aiming at the tracking failure problems caused by illumination change,scale change and occlusion in long-term target tracking,a long-term target tracking algorithm combining scaling and redetection is proposed based on the motion and appearance model of long-term correlation tracking algorithm.The algorithm innovatively uses the principal component analysis dimensionality reduction strategy to reduce the computational cost of the fast Fourier transform,and establishes a high confidence sample set to predict the target position.When the target is blocked or lost for a long time,the online classification detector and best-buddies Similarity(BBS)matching are started through the adaptive threshold,the target position is repositioned,and the template is balanced and updated.The performance of the algorithm in this chapter was evaluated quantitatively and qualitatively on partial sequences of standard datasets such as OTB-2015 and Temple Color-128,and the effectiveness and robustness of the algorithm were verified.3.Aiming at the problem of anti-jamming in complex scenes which is common in target tracking algorithms based on siamese Networks,a long-term tracking algorithm of Siamese Region Proposal Networks(SiamRPN)network based on interference discriminant is designed.On the basis of SiamRPN,the discriminant model is introduced to enhance the adaptability of the algorithm,and the fast search strategy is used to re-detect and locate the missing target.Through quantitative and qualitative analysis experiments,it is verified that the proposed algorithm has a great improvement compared with SIAMRPN target tracking algorithm,and its target tracking robustness is strong in complex scenes.Finally,the research content,innovation and experimental results of this thesis are summarized,and the problems found in the research and the future work are prospected.Figure[15]table[8]reference[75]...