基于锚框的远距离多尺度红外目标跟踪技术  

Long range multi-scale infrared target tracking technology based on anchor frame

作　　者：余云霞[1] 李毅鹏[2] 陈姝敏 YU Yun-xia;LI Yi-peng;CHEN Shu-min(Big Data Research Center,Jingchu University of Technology,Jingmen 448000,China;Zhongnan University of Economics and Law,Wuhan 430000,China;Jingchu University of Technology,Jingmen 448000,China)

机构地区：[1]荆楚理工学院大数据研究中心,湖北荆门448000 [2]中南财经政法大学,湖北武汉430000 [3]荆楚理工学院,湖北武汉448000

出　　处：《激光与红外》2024年第10期1594-1599,共6页Laser & Infrared

基　　金：中央高校基本科研业务费专项资金项目(No.2722023DK039);国家自然科学青年基金项目(No.62305141)资助。

摘　　要：在红外目标跟踪过程中,由于目标距离和视角的变化,红外目标的成像面积通常很小,使得红外图像中缺乏足够的像素信息,无法准确提取目标的特征,增加了目标跟踪的难度。因此,提出基于锚框的远距离多尺度红外目标跟踪技术。利用编码的方式对应追踪框图与真实框图,得到目标框中心坐标值。为保证计算的精准度,通过设定特征融合阈值判定提取信息与真实信息较高的重叠度,并按照锚框的边缘、中心以及顶点坐标对应划分像素网格,根据特征值输出分类输出向量,完成红外目标特征提取。由此采用损失函数给出锚框、目标框、实际框的类别损失以及候选框等不同类别损失函数,按照不同点在网格中的体素值,以编码形式逐一对比候选框和真实框数据,通过迭代实现远距离多尺度红外目标跟踪。实验结果表明,所提方法对远距离多尺度红外目标的识别效果较好,查全率曲线基本保持在0.9以上。说明所提方法具有良好的远距离多尺度红外目标跟踪效果。In the process of infrared target tracking,the imaging area of infrared targets is usually small due to changes in target distance and perspective,resulting in a lack of sufficient pixel information in the infrared image,and making it difficult to accurately extract target features and increasing the dfficulty of target tracking.Therefore,a long-distance multi-scale infrared target tracking technology based on anchor boxes is proposed.First of all,the tracking block diagram and the real block diagram are corresponded by encoding to get the coordinate value of the center of the target frame.To ensure the accuracy of the calculation,a feature fusion threshold is set to determine the high overlap between the extracted information and the true information,and the pixel grid is divided according to the edge,center,and vertex coordinates of the anchor box,and a classification output vector is output based on the feature values to complete the infrared target feature extraction.As a result,a loss function is used to provide the category loss of anchor boxes,target boxes,and actual boxes,as well as different category loss functions such as candidate boxes.According to the voxel values of different points in the grid,the candidate box and real box data are compared one by one in the form of coding,and the long-distance multi-scale infrared target tracking is achieved through iteration.The experimental results show that the proposed method has good recognition performance for long-distance multi-scale infrared targets,and the recall curve is basically maintained above 0.9.It indicates that the proposed method has good long-range multi-scale infrared target tracking performance.

关 键 词：红外目标 远距离多尺度 重叠度 特征提取 像素网格 

分 类 号：TN219[电子电信—物理电子学] TN976