Research On Target Tracking Based On Siamese Network

Target Tracking is one of the most important tasks in computer vision,and it has been widely applied in various fields,such as intelligent transportation,humancomputer interaction,autonomous driving,and automatic navigation of aircraft.With the rapid development of deep learning technology and its widespread application in the field of image processing,the use of deep learning methods to study the field of target tracking has become a research trend.Among them,Siamese Network achieves good performance and has become a hot topic in the field of target tracking.The target tracking field is currently facing problems such as scale variation,illumination variation,background clutter,deformation,and occlusion.In the target tracking algorithm based on the Siamese framework,both SiamRPN and SiamRPN++are trained through a large amount of data to obtain superior tracking performance.Although they have good tracking performance,they still have some shortcomings.The SiamRPN tracking model has the following problems:(1)It has a higher response value for objects similar to the target,and the target response value decreases when the target state changes.(2)In the entire tracking frameworkt,SiamRPN fixes the target feature of the first frame as a template feature,and lacks an update strategy.SiamRPN++ uses ResNet50 convolutional neural network as the backbone of feature extraction.Although more effective features can be extracted,the large feature extraction backbone reduces the model tracking rate.In order to solve the problems of the above two models and improve the tracking performance of the model,the main research of this paper is as follows:1)In response to the above-mentioned problems with SiamRPN model,and in order to reduce the negative effects caused by the interference of similar objects and the state change,based on the Siamese network,this paper proposes a channel positive and negative feedback network(CPFN)for target tracking.In this paper,the Gaussian kernel is used to quickly select the positive feedback feature channel and the interference feature channel,and effectively combine these deep features.In addition,this article also designed a novel update strategy to avoid template pollution caused by simple template update strategy.2)In order to ensure real-time tracking rate while maintaining tracking performance,this paper proposes a simple end-to-end Siamese network(SiamMSF)based on multi-scale feature fusion.By modifying the network structure and adding a multi-scale feature fusion module,the model tracking rate is more efficient.Specifically,the part that does not significantly improve the tracking results but has a huge amount of parameters is removed from the feature extraction backbone,which reduces the complexity of model training and improves the tracking speed.And the experiment verified the necessity of adding the convolutional layer for feature learning before the feature matching of the regression branch.This paper also designed a multi-scale feature fusion module based on rectangular convolution of different scales to further aggregate more effective features and improve the performance of the model.In addition,we use focal loss to focus the model on the learning of difficult samples,avoiding the decline of the model's generalization ability caused by the learning of a large number of simple samples.The tracking models based on the Siamese network proposed in this paper have been experimentally tested on various mainstream target tracking test sets,and the results show that the models have achieved state-of-the-art performance compared with the same type of the models.