Research On Object Tracking Method Based On Siamese Network

Target tracking is an important subject in the field of computer vision,and has a wide application prospect in security monitoring,visual navigation,medical diagnosis,military guidance and so on.However,due to the multi-scale variation,occlusion interference and other problems in the process of target movement,the existing target tracking algorithm can not effectively deal with multiple problems at the same time.This thesis focuses on the multi-scale change and occlusion interference of the target in the tracking process,aiming to improve the accuracy of the tracker by referring to the relevant technologies in the target detection field,and realize the effective and stable operation of the tracker.(1)Aiming at the problem of multi-scale variation in the process of target tracking and combining with the idea of multi-layer feature fusion,a target tracking algorithm based on improved feature extraction network is proposed in this thesis.Feature extraction part adopts the algorithm Mobile Net V3 lightweight convolution neural network as the foundation,high and low dimensional characteristics with two-way fusion,to achieve an increase in less amount of calculation under the premise of low resolution semantic information fusion strong characteristic figure and high resolution with weak semantic information but the characteristics of the space information rich figure,so as to enhance the capacity,for the detection of target Effectively avoid tracking failure due to multi-scale changes.Experimental results show that compared with Siam RPN algorithm,the accuracy and success rate of the proposed algorithm on OTB data sets are improved by 2.5% and 1.1%,respectively,and the accuracy,robustness and average overlap rate on VOT data sets are improved by 17.5%,5.9% and 9.5%,respectively.(2)Aiming at the problem of occlusion interference in the process of target tracking,the guidance anchor frame generation network is introduced,and a target tracking algorithm based on the improved anchor frame generation network is proposed in this thesis.The algorithm is based on(1)algorithm,the anchor box production part adopts guide anchor box production network GA-RPN as the foundation,with the help of a triple loss function training,implementation with the help of a target object feature information guidance of anchor box generated,eliminate redundant anchor box,by reducing the target characteristic and the information of the anchor box,reduce amount of calculation.Thus,the detection ability of occluded target can be improved and tracking failure caused by occluded interference can be effectively prevented.Experimental results show that compared with algorithm(1),the accuracy and success rate of the proposed algorithm on OTB data sets are improved by 1.3%and 0.7% respectively,and the accuracy,robustness and average overlap rate of the proposed algorithm on VOT data sets are improved by 17.1%,6.1% and 7.8% respectively.Based on the target tracking algorithm mentioned above,this thesis designs and implements target tracking system to provide target tracking service for security monitoring system.The system has the functions of target detection,target selection and target tracking,and is optimized according to the characteristics of hardware and software architecture.The system in this thesis is connected to the intelligent road Monitoring System,the information center dual prevention system and other security monitoring systems for testing.The test results show that the system can run effectively and stably,and the real-time running speed can reach 30 frames per second.