Research On Real-time Target Tracking Technology Based On Correlation Filtering And Siamese Network

As one of the key directions in computer vision research,visual target tracking technology has broad application prospects in many fields such as video surveillance,intelligent driving,and human-computer interaction.In recent years,target tracking technology has been rapidly developed,the accuracy of target tracking is increasingly improved,and the practicability is constantly enhanced.However,due to the complexity and variability of the actual environment,achieving accurate and robust tracking in practical applications is still a great challenge.In view of this,this article conducts in-depth research on target tracking algorithms based on correlation filtering and twin networks(Siamese networks).The main contents are as follows:(1)Introduced the basic process of target tracking and the corresponding phase division,and expounded the basic principles of common target tracking algorithms,including classic target tracking algorithms and the current mainstream target tracking algorithms based on correlation filtering and Siamese networks.In addition,the algorithm evaluation test sets(OTB and VOT)used in this paper and their corresponding evaluation criterion are also introduced.(2)A target tracking framework based on correlation filtering and Siamese network is designed.Among current mainstream tracking algorithms,algorithms based on correlation filtering have the disadvantage of model drift,while algorithms based on the Siamese network cannot be learned online.This paper designs a target tracking framework based on correlation filtering and Siamese network,and uses the proposed adaptive fusion strategy based on average peak-to-correlation energy(APCE).The fusion of correlation filtering and Siamese network algorithm is realized,and the tracking performance is effectively improved.In order to comprehensively evaluate the actual effect of the proposed framework,this paper uses the OTB dataset to verify the effectiveness of the framework designed in this paper.Under the premise of ensuring real-time performance,the algorithm in this paper also takes into account the robustness and accuracy of tracking,which is better than other algorithms for comparison.(3)An improved Siam Mask algorithm based on spatial channel reliability is designed.Siam Mask is a tracking algorithm based on the Siamese network,which realizes the prediction of the target position,target bounding box and target mask at the same time,but it also has the disadvantage of not being able to learn online.Therefore,an improved Siam Mask algorithm based on spatial channel reliability was proposed in this paper.By designing a correlation filter module based on spatial channel reliability using shallow depth features and adopting a model update strategy based on mask reliability,the target appearance changes were captured in real time.Then it was introduced into Siam Mask by using the response graph fusion strategy based on the average peak correlation energy,which enhanced the robustness of the algorithm and improved the overall performance of the algorithm.The test results on the VOT dataset show that the improved algorithm proposed in this paper is better than other algorithms participating in the comparison in indicators such as EAO(Expect Average Overlap Rate).