海上无人艇雷达相机拓扑双向融合探测算法  

Topology-Based Bidirectional Radar-Optical Fusion Algorithm for USV Detection at Sea

作　　者：唐浩 彭煊[2] 熊伟 崔亚奇 胡剑秋 邢汇源 TANG Hao;PENG Xuan;XIONG Wei;CUI Yaqi;HU Jianqiu;XING Huiyuan(Institute of Information Fusion,Naval Aviation University,Yantai 264001,China;The People's Liberation Army 31092 Troops,Beijing 100000,China;Jiangsu Automation Research Institute,Lianyungang 222061,China;The People's Liberation Army 72506 Troops,Shijiazhuang 050000,China)

机构地区：[1]海军航空大学信息融合研究所,烟台264001 [2]中国人民解放军31092部队,北京100000 [3]江苏自动化研究所,连云港222061 [4]中国人民解放军72506部队,石家庄050000

出　　处：《光子学报》2025年第1期159-175,共17页Acta Photonica Sinica

基　　金：国家自然科学基金青年基金(No.62001499)。

摘　　要：针对传统IoU融合方法在无人艇航行晃动情况下多目标融合准确率低的问题,提出一种拓扑双向融合算法。使用改进的YOLOv7-tiny算法对图像进行检测,利用矩阵转换将雷达点投影至图像。利用PROSAC算法对投影后的雷达点进行拟合与偏转处理,以减小船体晃摇对融合的影响。为减小计算量,依据传感器的系统误差及位置误差损失设计了一种粗关联波门。对粗关联后的雷达和相机数据进行拓扑融合,设计一种拓扑融合指标,在三角形相似度的基础上,增加多边形角度相似度和中心点连线相似度,弥补了雷达投影导致的角度信息缺失。对未融合雷达点附近图像截取进行检测融合;对未融合光学检测框,选取方位线最近邻雷达点进行融合,实现双向融合。实测数据分析结果显示,改进的YOLOv7-tiny算法mAP@0.5从0.883提高到0.93,所提拓扑双向融合算法准确率达到92.76%,明显优于IoU算法。该研究对海上无人艇雷达相机融合探测领域具有一定的参考意义。In recent years,there have been significant developments in the field of Unmanned Surface Vehicles(USVs),with a notable increase in the number of applications in both military and civilian contexts.Equipped with radar and optical systems,USVs are designed to markedly enhance detection capabilities.While radars are capable of determining target distance and bearing in all weather conditions,they are limited in their ability to classify targets.In contrast,optical systems have strong colour perception and classification abilities,with angular resolution comparable to that of lidar systems.However,their ranging capabilities are limited,and they are susceptible to instability in adverse weather conditions.The fusion of radar and optical systems,by leveraging their complementary advantages,effectively augments the detection capabilities of USVs.Radar-optical fusion methods can be broadly categorised into two primary approaches:the linkage method and the matrix transformation method.The linkage method entails the rotation of the optical servo in accordance with the azimuth and elevation angles of radar-detected targets,thus enabling the optical system to capture target images.Subsequently,the data is transmitted to the data centre,where it is analysed and decisions are made by the relevant personnel.Although this method is particularly focused on angular transformation and is effective in scenarios with fewer,unobstructed targets,it is unable to accurately detect multiple targets simultaneously.In contrast,the matrix transformation method unifies radar and optical information within a single coordinate system through the application of mathematical operations.By focusing on radar-detected points to generate Regions Of Interest(ROIs),this approach employs the Intersection over Union(IoU)algorithm for association.The matrix transformation method is predominantly applied in the autonomous driving sector.It facilitates the fusion of millimetre-wave,lidar,and onboard camera data,enabling multi-target detection and 360°capt

关 键 词：拓扑双向融合 雷达相机融合探测 粗关联波门 改进YOLOv7-tiny 相似度 

分 类 号：TP212.9[自动化与计算机技术—检测技术与自动化装置]