Research On Three-branch Infrared Target Tracking Algorithm Based On Local Structrure And Global Sematics

Target tracking is a highly concerned research field in computer vision in recent years.Its task is to detect a specific target in the surveillance scene and track the specified target.Visible light target tracking algorithm has been applied in many fields such as video surveillance,unmanned driving,unmanned aerial vehicle,etc.,and its effect is getting better and better.However,under the condition of low light or even no light at night,the target tracking effect is not so good.This means that when the target tracking is in a scenario with poor lighting conditions,the target tracking task may fail,resulting in irreparable losses.Especially in complex scenes such as security and rescue,low-light or even completely dark scenes are often more worthy of attention.In these scenes,visible light imaging quality is poor,and even effective images cannot be obtained.Infrared detector,due to its imaging characteristics,can be a good solution to the object imaging in low light or even completely dark scenes.This paper mainly studies the target tracking method under the infrared scene.Aiming at the two stages of target tracking,target feature extraction and discrimination,this paper does the following two works.Infrared detector can work well at night because of its characteristics of imaging at night,but it also has its own problems,such as low resolution,high noise and other characteristics of imaging quality.Therefore,it is necessary to design highperformance target tracking algorithm to solve the existing problems.Therefore,a thermal infrared target tracking algorithm based on local structure and global semantic mining is proposed in this paper.In this algorithm,when dealing with the problem of low resolution,a local structure feature enhancement module is designed to obtain the infrared target features with higher discrimination,so that the infrared target can be tracked well in the low resolution image.In order to deal with the problem of high noise,a global semantic enhancement module is designed to alleviate the effect of noise on infrared target tracking.The thermal infrared tracking algorithm based on local structure and global semantic mining improved the expected average overlap rate(EAO.)of the infrared target tracking data set VOT-TIR 2017 by 23.56%compared with the benchmark SIAMFC.The effectiveness and feasibility of thermal infrared target tracking algorithm based on local structure and global semantic mining are verified.In the discrimination stage,two branching structures are commonly used,including the pre-background classification branch and the target box regression branch.However,these models all have the problem of the positioning accuracy of the target frame,among which the most important factor is that the classification effect of the pre-background classification branch is poor.In order to solve this problem,this paper puts forward a three-branch structure.The three-branch structure includes the pre-background classification branch,the quality assessment branch and the target box regression branch.The new quality estimation branch is to correct the classification results of the pre-background classification branch,so as to obtain more accurate classification results,so as to alleviate the problem of low positioning accuracy of the target frame.The experiment showed that the expected mean overlap ratio(EAO.)of the three-branched discriminant model was 3.14% higher than that of the two-branched discriminant model on the VOT-TIR 2017.The experimental results verify the effectiveness of the three-branch discriminant network,which has reached the current advanced level.