Research On Target Tracking Algorithm Based On Siamese Neural Network

Target tracking is a hot topic in current computer vision.At present,target tracking has been widely applied in military,unmanned driving,intelligent monitoring,humanmachine interaction,and medical imaging.The key issue is how to use the state information of the target in the image sequence to locate the target in the next frame of the image.The traditional target tracking method uses correlation filters to locate the position of the target in subsequent tracking by searching for response peaks in the predicted distribution map.In the past decade,great progress has been made in the research of target tracking,and various machine algorithms have been introduced into algorithm research,gradually surpassing traditional target tracking methods in terms of performance.At present,research on target tracking has always revolved around the basic issues of feature extraction,target apparent modeling,and tracking strategies.In feature extraction,the Siamese network model is a commonly used deep learning based method.However,in the actual tracking process,there are often changes in target attitude,background interference,occlusion,and other situations,which affect the tracking effect.The model results still have a certain gap with the ideal tracking effect.This article proposes two improved target tracking algorithms based on twin neural networks.(1)The performance of current Transformer based object tracking algorithms has been greatly improved,but only simple feature enhancement and fusion have been performed,which has higher computational complexity compared to common Siamese network model based object tracking algorithms.Therefore,this article proposes a tracking algorithm that combines the Da Siam RPN target tracking algorithm with the Trans T target tracking algorithm.The Hamming distance and response score are used to determine the target tracking effect of the current frame.If both the response score and Hamming distance score decrease,Trans T is used for target tracking to prevent the target tracking effect from deteriorating or even failing.The experimental results show that the algorithm achieved success rates of 72.0%,69.1%,and 67.1% on the GOT-10 k,OTB2015,and UAV123 datasets,respectively,and improved tracking speed by 3.8%and 0.7% compared to Trans T on the OTB2015 and UAV123 datasets,respectively.(2)The SiamCAR model decomposes visual tracking tasks into classification of pixel categories and regression of the bounding box of that pixel to solve tracking problems.In response to the issue that the tracking accuracy and speed of the SiamCAR model still need to be improved in more complex tracking environments,this article integrates two new modules on the basis of SiamCAR: the ODConv module improves the Res Net-50 feature extraction part on the framework of SiamCAR,improving the feature extraction ability;The Alpha-Refine module improves the accuracy of target scale estimation and optimizes the tracking results of SiamCAR prediction.The experimental results show that the success rates of the improved algorithm on the OTB2015,UAV123,and VOT2016 datasets are 65.4%,54.7%,and 57.2%,respectively,and the success rate on the OTB2015 dataset is 2.16% higher than that of the unimproved SiamCAR algorithm.