基于红外-反转红外图像的双分支无人机目标跟踪算法  

Dual-branch Algorithm for Tracking UAVs with Infrared and Inverted Infrared Image

作　　者：李邵港 高晋 王刚 王以政 李椋 LI Shaogang;GAO Jin;WANG Gang;WANG Yizheng;LI Liang(University of South China,Hengyang 421001,China;Institute of Military Cognition and Brain Sciences,Academy of Military Medical Sciences,Beijing 100850,China;Institute of Automation,Chinese Academy of Sciences,Beijing 100190,China)

机构地区：[1]南华大学,湖南衡阳421001 [2]军事医学研究院军事认知与脑科学研究所,北京100850 [3]中国科学院自动化研究所,北京100190

出　　处：《火力与指挥控制》2023年第6期19-27,共9页Fire Control & Command Control

基　　金：国家自然科学基金面上项目(61972394);科技创新2030—“新一代人工智能”重大基金资助项目(2020AAA0105802)。

摘　　要：热红外成像会随场景各个部分的温度变化产生灰度波动,小型无人机目标在建筑等大型背景中穿梭时易受干扰造成目标区域和背景区域之间对比度反转,导致现有的跟踪方法失效。针对该问题,提出一种双分支目标跟踪框架,利用不同分支分别提取原始红外图像及反转红外图像特征,在搜索图像中分别对原始图像与反转图像中的目标模板进行匹配。此外,提出一种互相关特征融合方法,将两个分支的特征进行融合,增强目标区域特征以获得更精准的目标框回归。在ICCV2021 Anti-UAV数据集上进行实验,该算法基于边界框重叠率阈值(0.5)和中心点位置误差阈值(20像素)的跟踪成功率和精确率分别为78.39%和80.07%,比基准算法分别提升5.84%和4.52%,高于TransT、SiamRPN++、SiamMask等算法。结果表明双分支能够有效提升红外目标跟踪的性能。Thermal infrared imaging has the effect of gray-scale value fluctuation due to the temperature changes of different parts in a scenario.The imaging for small UAV targets is easily disturbed when they shuttle between different surrounding regions in the large background such as buildings,resulting in target contrast inversion,the existing tracking methods are invalid.To tackle this problem,a dual-branch target tracking framework is proposed,which extracts the features of the original infrared images and inverted infrared images respectively with different branches,and matches the target templates in the original images and inverted images respectively during searching images.In addition,a cross-correlation feature fusion method is proposed,which fuses the features of the two branches to enhance the features of the target area so as to obtain more accurate bounding box regression.The experiment is carried out on the ICCV2021 Anti-UAV dataset,it has achieved tracking success rate of 78.39%and precision of 80.07%respectively when the algorithm is based on bounding box overlap threshold of 0.5 and central point location error threshold(20 pixels),which are 5.84%and 4.52%respectively higher than the benchmark algorithm and are also higher than that of other algorithms such as TransT,SiamRPN++,SiamMask and so on.The results have shown dual-branch can effectively improve the performance of infrared target tracking.

关 键 词：红外目标跟踪 对比度反转 无人机 特征融合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]