Research On Target Tracking Algorithm Based On Deep Siamese Network

With the continuous progress and development of computer science,the direction of computer vision has become more and more important.Among them,target tracking is one of the important research directions in computer vision.Target tracking is highly valued in many fields,such as: video surveillance,Human-driven cars,robots,and ocean exploration all play extremely important roles.However,current target tracking tasks are often interfered by various factors,such as illumination,occlusion of similar objects,and tracking target background.Therefore,it is very challenging to establish a tracking model with higher accuracy and better robustness.Today’s twin networks continue to shine in the field of target tracking.This article focuses on the research of target tracking algorithms based on deep twin networks..First,it studies and analyzes Siam RPN,a twin network tracking algorithm.It uses the residual network to expand the backbone network depth of the model,and introduces the structure of the cascade network to extract better features to improve the tracking effect.On this basis,this paper proposes an Anchor-free twin tracking model Siam AFN(Siamese Anchor-free FPN Network,Siam AFN),adding feature pyramids to the model to build a new twin feature pyramid network structure,and extracting features The top-down path and horizontal connection are added to obtain a more accurate and robust feature extractor;and although the traditional area suggestion network improves the accuracy of the tracker,the traditional area suggestion network requires a preset anchor frame.Too many hyperparameters are introduced for the algorithm,which requires many times of training and debugging to obtain better performance.This paper constructs an Anchor-free based regional suggestion network to replace the traditional regional suggestion network.The constructed Anchor-free regional suggestion network is composed of regression and classification.The regression branch estimates the scale and classification of the target.The branch predicts the foreground and background probabilities,and includes a parallel centrality branch to remove anchor boxes with lower scores.The Siam AFN method proposed in this paper has achieved real-time tracking effects on the single-target tracking public data sets OTB100 and VOT2018,and has been compared with the current classic single-target tracking algorithms Siam RPN,ECO-HC,Siam FC,CFNet,SINT,DSST and Staple.Comparing experimental analysis,it can be verified that the proposed method reduces the hyperparameters that need manual debugging,and the model has better performance improvement and achieved good target tracking effects.