Infrared Target Tracking Method In Ground Scene

In the current production and life,infrared video,as an important medium for obtaining temperature field information,has derived many industrial or daily applications.As a very important technology in computer vision,the target tracking algorithm is widely used in continuous,accurate and fast capture of the calibrated target in the image sequence.The infrared-to-ground target tracking algorithm that combines the two is a practical challenge,which has attracted wide attention from researchers from various countries.There are many shortcomings in the existing algorithms.For ground targets,the quality of infrared video is relatively low,resulting in poor video tracking performance,and the diversity of information and the complexity of motion make the accuracy and robustness of the tracking algorithm have room for improvement.This article focuses on the difficulties faced by infrared video target tracking.First,the reasons for the poor quality of infrared video and the improved methods are analyzed,and then two algorithms based on infrared video self-attention and infrared video boundary enhancement based on visible light video are proposed to solve these problems.The main research contents of this paper are as follows:(1)Analyze the characteristics of infrared video.First,analyze the reasons for the poor quality of infrared video from the perspective of video acquisition;then,analyze the difficulties and challenges of poor quality video from the perspective of the most popular technology at the moment.Finally,the influence of the current evaluation index on the target tracking algorithm of infrared video is analyzed.(2)Aiming at the weaker boundary of infrared video,a self-attention mechanismbased image preprocessing mechanism is proposed.In order to enhance the boundary of the infrared video,through the relationship between each pixel and the surrounding pixels,the structural information of the pixel is obtained,and then the target tracking framework based on the twin network and the candidate region extraction network is used for subsequent processing,and a set of the complete end-to-end network architecture improves the accuracy and robustness of tracking,the average accuracy reached 0.768,and the average number of losses dropped to 0.75.(3)Aiming at the shortcoming of insufficient boundary information extracted by infrared video itself,a supervisory network based on visible light video is proposed,which uses visible light boundary information to enhance infrared video boundary information,so that the structure information of infrared video is enhanced during tracking..At the same time,a loss function is proposed to measure structural information and train and regularize the network.The accuracy and robustness of the supervision mode are improved,which the algorithm achieved an average accuracy of 0.707 and a follow-up rate of 0.714 in the more challenging test set,and no additional algorithm overhead is required.