基于RDB-YOLOv5的遥感图像车辆检测  

Remote sensing images vehicle detection based on RDB-YOLOv5

作　　者：周力 惠飞 张嘉洋 戚建 杨景超 唐翠仁 ZHOU Li;HUI Fei;ZHANG Jia-yang;QI Jian;YANG Jing-chao;TANG Cui-ren(School of Electronic Control,Chang'an University,Xi'an 710064,Shaanxi,China;China Construction Eighth Engineering Division CORP.,LTD.,Xi'an 710001,Shaanxi,China;Xi'an Outer Ring Branch of Shaanxi Transportation Holding Group Co.Ltd,Xi'an 710061,Shaanxi,China)

机构地区：[1]长安大学电控学院,陕西西安710064 [2]中国建筑第八工程局有限公司,陕西西安710001 [3]陕西交通控股集团有限公司西安外环分公司,陕西西安710061

出　　处：《长安大学学报（自然科学版）》2024年第3期149-160,共12页Journal of Chang’an University(Natural Science Edition)

基　　金：国家重点研发计划项目(2021YBF2501200);陕西省交通运输厅科研项目(22-04x)。

摘　　要：为了解决遥感图像中目标密集、目标小车辆检测难的问题,提出基于YOLOv5的改进模型RDB-YOLOv5,并将其首次应用于遥感图像车辆检测领域。首先,针对遥感图像车辆方向任意的问题,对现有基于旋转边界框(rotated bounding box)的目标检测方法环形平滑标签(circular smooth label,CSL)进行改进;其次,针对遥感图像中背景信息复杂,车辆尺寸小导致检测精度降低的问题,提出基于注意力的多尺度目标检测方法,在骨干网络中添加双注意力机制(dual attention mechanism)将局部特征与全局特征相结合,并用空洞卷积进行改进;再借鉴双向特征融合网络(BiFPN)的思想,加入新的浅层特征和深层特征信息传递路径,更好融合浅层中车辆的位置信息,并设计新的检测头,提高网络对小目标车辆的检测能力。研究结果表明,改进的RDB-YOLOv5相比YOLOv5的均值平均精度(mAP)增长2.7%,特别是小型车辆检测提高了3.5%,相较于传统模型RCNN等mAP整体平均提高了10%。RDB-YOLOv5能够在通用数据库上取得较高的检测精度,同时在遥感图像的复杂场景中能有效解决水平框检测带来的重叠以及漏检问题,并提高了小型车辆目标的检测准确率。To solve the problem of dense target and difficult detection of small target vehicle in remote sensing image,an improved model called RDB-YOLOv5 based on YOLOv5 was proposed and applied it in remote sensing image vehicle detection for the first time.Firstly,to address the problem of arbitrary vehicle orientation in remote sensing images,the existing rotation bounding box-based object detection method CSL(circular smooth label) was improved.Secondly,a multi-scale object detection method based on an attention mechanism was proposed to tackle the problem of complex background information and reduce detection accuracy due to small vehicle sizes in remote sensing images.A dual attention mechanism was introduced in the backbone network to combine local and global features,and improvement was made using dilated convolutions.Furthermore,inspired by the idea of bidirectional feature pyramid network,a new shallow feature and deep feature information transmission paths were added,it was incorporated better to integrate the positional information of vehicles in shallow layers,and a new detection head was designed for enhance the detection capability of small target vehicles in the network.The results show that RDB-YOLOv5 achieves a 2.7% increase in mean average precision(mAP) compared to the improved YOLOv5,especially with a 3.5% improvement in small vehicle detection.Compared to traditional models like RCNN,the overall map is improved by an average of 10%.RDB-YOLOv5 can achieve high detection accuracy on public databases and effectively solve the issues of overlap and missed detections caused by horizontal bounding box detection in complex scenes of remote sensing images,and the detection accuracy of small vehicle targets also improves.8 tabs,9 figs,32 refs.

关 键 词：交通工程 数字图像处理 遥感图像 车辆检测 旋转边界框 双注意力机制 双向特征融合网络 

分 类 号：U495[交通运输工程—交通运输规划与管理]