改进YOLOv5的无人机影像车辆检测方法  

Vehicle Detection Method of UAV Imagery Based on Improved YOLOv5

作　　者：范江霞 张文豪 张丽丽 余涛 钟林汕 FAN Jiangxia;ZHANG Wenhao;ZHANG Lili;YU Tao;ZHONG Linshan(School of Remote Sensing and Information Engineering,North China Institute of Aerospace Engineering,Langfang,Hebei 065000,China;Hebei Province Space Remote Sensing Information Processing and Application of Collaborative Innovation Center,Langfang,Hebei 065000,China;National Engineering Laboratory for Remote Sensing Satellite Applications,Aerospace Information Research Institute,Chinese Academy of Sciences,Beijing 100094,China;China Science and Technology Space Information(Langfang)Research Institute,Langfang,Hebei 065001,China)

机构地区：[1]北华航天工业学院遥感信息工程学院,河北廊坊065000 [2]河北省航天遥感信息处理与应用协同创新中心,河北廊坊065000 [3]中国科学院空天信息创新研究院遥感卫星应用国家工程实验室,北京100094 [4]中科空间信息(廊坊)研究院,河北廊坊065001

出　　处：《遥感信息》2023年第3期114-121,共8页Remote Sensing Information

基　　金：高分辨率对地观测系统重大专项(30-Y30F06-9003-20/22);国家重点研发计划项目(2019YFE0127300);国家自然科学基金项目(41907192);河北省自然科学基金项目(D2020409003);河北省高等学校科学技术研究项目(ZD2021303);民用航天预研项目(D040102);国防基础科研项目(JCKY2020908B001);北华航天工业学院博士科研启动基金项目(BKY-2021-31);北华航天工业学院硕士研究生创新资助项目(YKY-2021-29)。

摘　　要：针对传统遥感影像车辆目标检测算法易受干扰、鲁棒性较差且在实际应用当中会产生一定的漏检与误检现象等问题,提出了一种基于改进YOLOv5s的轻量级无人机遥感影像车辆目标检测算法。以YOLOv5s为基线模型,根据车辆目标长宽比相对固定的特点,对锚框尺寸进行修正,提高了锚框与车辆目标的契合度;针对无人机影像中车辆密集情况,进行了加权框融合改进,对检测框合并,解决了预测框计数不准确的问题;由于车辆目标具有多变性,通过增加注意力机制网络,提升了模型识别车辆的速度和准确性。研究表明,改进的YOLOv5s模型可以实现实时准确的无人机影像车辆检测。Aiming at the problems that the traditional remote sensing image vehicle target detection algorithm is easy to be interfered,has poor robustness,and will produce certain missing and false detection phenomena in practical applications,a lightweight Unmanned Aerial Vehicle(UAV)remote sensing image vehicle target detection algorithm based on improved YOLOv5s is proposed.Taking YOLOv5s as the baseline model,according to the relatively fixed aspect ratio of the vehicle target,the size of the anchor frame is modified to improve the fit between the anchor frame and the vehicle target.In view of the dense vehicles in the UAV image,the weighted frame fusion is improved,and the detection frame is merged to solve the problem of inaccurate prediction frame count.Due to the variability of vehicle targets,the speed and accuracy of vehicle recognition are improved by adding attention mechanism network.The research results show that the improved YOLOv5s model can achieve real-time and accurate vehicle detection of UAV images.

关 键 词：车辆 目标检测 YOLOv5s 轻量级 无人机影像 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]