Research On Infrared Target Tracking Algorithm Based On Siamese Network

Most target tracking algorithms are designed based on visible scenes,but in some cases,infrared target tracking has advantages that visible light does not have.Infrared devices use radiation from the object itself to image,do not require an additional light source,can show a target in low-light or dark scenes.However,infrared images also have some defects,such as the boundary between target and background is not clear,the image is fuzzy and even the background is messy.Especially,some images in public infrared datasets are rough,which has a certain negative impact on the training of data-driven infrared target tracking algorithm.From the perspective of image properties,compared with visible light tracking,target tracking in infrared scenes lacks color,texture and other information.Moreover,the contrast of infrared images is low,so the problem of interference caused by similarity between targets and similarity between targets and background is particularly prominent.At present,infrared target tracking methods based on deep learning have occupied an important position in the field of infrared tracking.However,such methods not only lack the use of target details in the above infrared scenes,but also have the problem of poor generality caused by repeated manual debugging of fixed anchor frame to train and predict the target size.In addition,in order to make the neural network pay more attention to infrared target features,most current algorithms add a large number of attention mechanisms into the network,which seriously affects the computational efficiency.Secondly,in the tracking stage,especially in the long-term tracking,the tracker is vulnerable to the interference of the surrounding similarity.When there are similar targets in the tracking scene and the background is messy,most of the trackers cannot effectively update the template of tracked target.If the historical information of the target cannot be fully utilized,the tracker will eventually lose the target,resulting in poor robustness in the long-term tracking.Finally,most infrared tracking algorithms lack a re-detection mechanism after losing target,which leads to poor tracking effect after occlusion or blurring.In order to solve the above problems,based on the siamese network tracker,this paper makes the following improvements:(1)In order to solve the problems such as unclear target and background details in infrared scenes,rough image of infrared target dataset and inability of infrared target tracker to utilize historical information of target,this paper proposes an infrared target tracking algorithm based on attention and adaptively update of target model.First,based on the anchor-free algorithm,a fast attention enhancement module is added to process infrared images in parallel,improve the difference between infrared targets and background and enhance the target details without losing the original information.Then,the extracted features are fused to the middle layer of the backbone network.Finally,the target adaptive update network is used to learn the feature change trend of infrared targets,and the middle and high level features of targets are dynamically updated,which strengthens the long-term tracking ability of infrared tracker.(2)The current infrared tracking method is basically that after the target is selected manually,the tracker realizes long-term tracking by using the historical features of the target.However,when similar interference and target occlusion occur,these methods lack a redetection mechanism to re-determine the target.If the contaminated template is continued to be used to update the target position,errors will be generated and the tracking will fail eventually.To solve these problems,an infrared tracking algorithm based on scene-aware classification and re-detector is proposed in this paper.Firstly,Siam BAN is taken as the main frame and scene-aware classifier is proposed.The classifier will judge according to the image features of the current image and output the coefficients most suitable for the scene.According to these coefficients,the similarity calculation results of the three BAN are dynamically weighted to improve the feature utilization ability of different infrared scenes.Secondly,the infrared target re-detector is designed.The classifier based on multi-domain convolutional network is used as the judgment module,and the first frame target feature is used for initialization,so that it can adapt to the characteristics of the current target.In the follow-up continuous tracking,the tracking results of Siam BAN are judged by the judgment module,and the target with high confidence is used as the sample training judgment module.When it is judged that the tracked target is missing,the structure of the judgment module is reused into a re-detector,and the randomly collected targets are classified and sorted according to the confidence level,ultimately finding the target and completing the tracking.This method improves the accuracy of judging the tracking state,and provides an efficient and simple solution to the lost target searching,with higher localization and regression accuracy.