Research On Visual Tracking Algorithm Based On Siamese Network

Computer vision is to use computers and photographing equipment to replace human eyes and brains to identify,detect and track objects,and to analyze and process the obtained image information.As one of the important basic researches of computer vision,object tracking technology has broad application prospects in the fields of robot control,unmanned automatic driving and intelligent monitoring.The single object tracking algorithm studied in this paper is to determine the position of the tracking object according to the first frame,and the subsequent frames use the tracker to predict the position information of the object,and then complete the tracking task.In recent years,with the development of machine learning and deep learning,visual tracking algorithms based on Siamese networks have gradually become popular.Such algorithms have certain advantages in terms of tracking success rate and accuracy,but large-scale morphological changes and object occlusion appear in the target.In complex scenarios,such as loss of target location information,low tracking success rate and precision,etc.In order to improve the shortcomings of existing Siamese network tracking algorithms,this paper proposes an attention-based and anchor-free Siamese network visual tracking algorithm.A deeper convolutional neural network is used as the feature extraction backbone network,and the multi-layer feature extraction network is fused to obtain richer semantic information to express multi-scale features.A transformation network method based on attention mechanism is introduced,using for the fusion and matching of the features of the search area and the template area,the perception ability of the global field of view is obtained.The classification and regression algorithm based on the anchor free is used to complete the accurate prediction of the target center position and bounding box.To verify the effectiveness of the algorithm,it is compared with mainstream Siamese network tracking algorithms on some single object tracking test datasets.In the OTB100 dataset,the success rate reached 69.3%,and the precision reached 91.1%.In the UAV123 dataset,the success rate reached 62.3%,and the precision reached80.9%.Compared with Hi FT,which also used the transformer network,the success rate was increased by 3.4%,and the tracking accuracy was improved by 2.2%.The Siam CAR with anchor free regression method improves the success rate by 0.9% and the tracking accuracy by 4.9%.At the same time,in the VOT 2019 dataset,in the current experimental environment,the tracking speed reached 50.2 frames per second,meeting the requirements of real-time tracking.By comparing with the mainstream Siamese network visual tracking algorithm on some single object tracking datasets,the algorithm in this paper has certain advantages in accuracy and robustness,and has better performance in complex scenes such as object occlusion,large-scale morphological changes of objects,and interference with similar objects.