Siamese Network Based Visual Tracking Algorithm

Target tracking technology is an important branch of research in computer vision.There are many application scenarios,which can be used in video surveillance,human-computer interaction,competitive sports and other fields.The target tracking algorithm is to predict the position and size of the target in subsequent frames and make a mark according to the initial state of the target given in the initial frame of the video.In recent years,with the continuous development of target tracking algorithms,the tracking accuracy has been continuously improved.However,when the target is subject to severe deformation,occlusion,and fast movement interference,tracking failure may occur.In addition,tracking speed is also an important evaluation index for tracking algorithms.Therefore,it is important to study the tracking algorithm more deeply.The mainstream tracking algorithms can be divided into two methods.One is a tracking algorithm based on correlation filtering,which is characterized by fast tracking speed and can meet the requirements of real-time tracking,but the tracking accuracy is low.The other is a deep learning based tracking algorithm,which is characterized by slower tracking speed but higher tracking accuracy.The tracking algorithm based on the Siamese network uses the method of deep learning to effectively improve the tracking speed while ensuring the tracking accuracy.The fast online target tracking and segmentation algorithm(Siam Mask)is based on the highly accurate algorithm in the siamese network tracking framework,but Siam Mask does not need to learn online during tracking,The template features are only extracted based on the pre-processed first frame image,It is not updated during the tracking process,Although the speed is excellent and meets the requirements of real-time tracking,the template feature cannot be well adapted to the change of the target appearance model;In addition,in the mask branch of Siam Mask,the shallow features of the template path contain redundancy,which will pollute the generated target mask and affect the algorithm's positioning of the target.In order to improve theaccuracy of the tracking algorithm based on the siamese network,two optimization methods are proposed in the Siam Mask framework:(1)In order to improve the expression ability of the template features,by setting the learning rate,the template features extracted from the initial frame are merged with the template features extracted from each frame,the template features are updated,the semantic information of the template features is enriched,improved judgment ability of tracker.(2)In the process of obtaining the target mask,by setting the weight,the convolution features of the template branch dimensionality reduction are diluted,the redundancy of the dimensionality reduction convolution features is reduced,and the tracking speed can be achieved in real time,While improving the tracking accuracy.In order to verify the effectiveness of the improved algorithm in this paper,tests were performed on the VOT2016 and VOT2018 datasets.The expected average overlap rate is 0.450 and 0.390,the accuracy is 0.649 and 0.618,and the robustness is0.205 and 0.272.It is higher than other comparison algorithms to varying degrees,and the tracking speed of 34.66 frames/s meets the requirements of real-time tracking.The proposed algorithm effectively improves the accuracy of tracking,and can complete the tracking task well in a complex tracking environment.