无人机高空航拍视角下小尺度车辆精确检测方法  

Accurate Detection Method of Small-scale Vehicles from Perspective of Unmanned Aerial Vehicle High-altitude Aerial Photography

作　　者：张河山[1,2] 谭鑫 范梦伟 潘存书 徐进[1,2] 张羽 ZHANG Heshan;TAN Xin;FAN Mengwei;PAN Cunshu;XU Jin;ZHANG Yu(School of Traffic&Transportation,Chongqing Jiaotong University,Chongqing 400074,China;Chongqing Key Laboratory of"Human-Vehicle-Road"Cooperation and Safety for Mountain Complex Environment,Chongqing Jiaotong University,Chongqing 400074,China;School of Economics and Management,Chongqing Jiaotong University,Chongqing 400074,China)

机构地区：[1]重庆交通大学,交通运输学院,重庆400074 [2]重庆交通大学,山区复杂道路环境“人-车-路”协同与安全重庆市重点实验室,重庆400074 [3]重庆交通大学,经济管理学院,重庆400074

出　　处：《交通运输系统工程与信息》2024年第3期299-309,共11页Journal of Transportation Systems Engineering and Information Technology

基　　金：国家自然科学基金(52172340);重庆市教育委员会青年项目(KJQN202200710);重庆市博士后科学基金(CSTB2022NSCQ-BHX0731)。

摘　　要：无人机高空航拍图像中车辆像素占比极低,目标可视化信息较少,在目标检测任务中容易漏检和误检。因此,本文提出一种基于改进YOLOX(You Only Look Once X)的无人机高空航拍视角下小尺度车辆精确检测方法。首先,为增强网络对低级特征的提取能力,在原始YOLOX预测头部增加一个160 pixel×160 pixel的浅层特征提取网络;其次,在骨干网络后端嵌入基于归一化的注意力机制模块(Normalization-based Attention Module,NAM),以抑制冗余的非显著特征表达;最后,为了增大小尺度车辆的相对像素比,提升网络捕捉有效特征信息的能力,提出一种基于滑动窗口的图像切分检测方法。试验结果表明,改进YOLOX网络表现出良好的检测效能,检测精度达到了84.58%,优于典型的目标检测网络Faster R-CNN(79.95%)、YOLOv3(83.69%)、YOLOv5(84.31%)及YOLOX(83.10%)。此外,改进YOLOX能够有效解决无人机高空航拍图像中小尺度车辆的漏检和误检问题,且预测框更贴合车辆的实际轮廓;同时,在不同航拍高度的目标检测任务中具有较高的鲁棒性。The proportion of vehicle pixels in high-altitude aerial images taken from unmanned aerial vehicle(UAV)is low,and there are limited visual information of can be identified for the targets,which result in missed or false detection in the detection tasks.This paper proposes a detection method of small-scale vehicles based on improved YOLOX(You Only Look Once X)from the perspective of high-altitude aerial photography.First,to enhance the network's ability to extract low-level features,this study added a shallow feature extraction network of 160 pixel×160 pixel to the original YOLOX prediction head.Then,a Normalization-based Attention Module(NAM)was embedded after the backbone network to suppress redundant non-significant feature expression.To increase the relative pixel ratio of small-scale vehicles and improve the ability of the network to capture effective feature information,an image segmentation detection method was proposed based on sliding window.The experiment results show that the improved YOLOX network shows good detection performance,and the detection accuracy reaches 84.58%,which is better than the typical target detection network Faster R-CNN(79.95%),YOLOv3(83.69%),YOLOv5(84.31%),and YOLOX(83.10%).In addition,the improved YOLOX can effectively solve the problem of missed detection and false detection of small-scale vehicles in high-altitude aerial images of UAV,and the prediction box is more suitable for the actual contour of the vehicle.At the same time,it has high robustness in target detection tasks at different aerial heights.

关 键 词：智能交通 小尺度车辆检测 YOLOX 无人机 注意力机制 浅层特征提取网络 

分 类 号：U491[交通运输工程—交通运输规划与管理]