Research On Single Target Tracking Algorithm Based On Siamese Network

In recent years,with the development of artificial intelligence,deep learning-based object tracking algorithms have become more and more widely used in the fields of unmanned aerial vehicles,autonomous driving,sports events,and public safety.The object tracking algorithm has been able to effectively adapt to a variety of scenes in real life.By extracting the characteristics of the target object's color,angle,trajectory,and scale change,the object tracking network has the ability to distinguish the target's foreground and background and predict the target location.Although the current object tracking algorithms have many solutions to problems such as occlusion,deformation,jitter,fast movement,etc.,when the target is deformed in a complex environment,the accuracy of the tracker's position prediction will be greatly reduced.This paper proposes two single-object tracking algorithms based on the siamese network.Compared with the current method,they can effectively improve the accuracy of the tracker and adapt to the accuracy of the position prediction when the object is deformed.The main contributions of this paper are as follows:1.The non-anchor method is introduced based on the region-generating siamese network to solve the problem of too long calculation time when the model is tracking tasks and the problem of high tracking failure rate when the object is deformed.In the branch of predicting the position of the object,this paper directly predicts the distance between the center point and the periphery of the frame.Under the premise that the accuracy is almost unchanged,compared with the original anchor-based target tracking algorithm,the parameter amount is 5 times less,which is effective the calculation time required for object tracking is shortened.The experimental results on the VOT dataset show that this method is higher in accuracy and speed than the traditional method model.2.Propose a multi-layer feature fusion method for fusing deep features and shallow feature information,combining the feature maps of the multi-layer network to form a continuous feature space,and fusing the high-latitude features of the deep network on the premise of preserving the image feature details.It enriches the image feature information extracted by the network model,improves the diversity of feature extraction,and reduces the amount of network training parameters.The experimental results on the VOT dataset show that the tracking accuracy of this method is higher than that of Siam FC,Siam VGG,DSiam,and other methods.