交通运输系统工程与信息

交通运输系统工程与信息 ›› 2026, Vol. 26 ›› Issue (3): 247-258.DOI: 10.16097/j.cnki.1009-6744.2026.03.023

• 智能交通系统与信息技术 • 上一篇    下一篇

无人机夜间航拍视角下小目标车辆精确检测方法

郑展骥1 ，廖方正1 ，李燊2 ，冯昌奎1 ，凃强1,3 ，张河山*1 ，徐进1   

  1. 1. 重庆交通大学，交通运输学院，重庆400074；2.延安市公路局，陕西延安716000； 3. 重庆市城投金卡信息产业（集团）股份有限公司，博士后科研工作站，重庆400074
  • 收稿日期:2025-10-29 修回日期:2025-12-09 接受日期:2026-01-06 出版日期:2026-06-25 发布日期:2026-06-23
  • 作者简介:郑展骥（1990—），男，浙江瑞安人，副教授，博士。
  • 基金资助:
    教育部人文社会科学研究青年基金(24YJCZH412)；重庆市社会科学规划项目 (2025PY39)。

Accurate Detection Method for Small Target Vehicles from Drone Night Aerial Photography Perspective

ZHENG Zhanji1, LIAO Fangzheng1, LI Shen2, FENG Changkui1, TU Qiang1,3, ZHANG Heshan*1, XU Jin1   

  1. 1. College of Transportation, Chongqing Jiaotong University, Chongqing 400074, China; 2. Yan'an Highway Bureau, Yan'an 716000, Shaanxi, China; 3. Postdoctoral Research Station, Chongqing City Investment & Golden Card Information Industry (Group) Co Ltd, Chongqing 400074, China
  • Received:2025-10-29 Revised:2025-12-09 Accepted:2026-01-06 Online:2026-06-25 Published:2026-06-23
  • Supported by:
    Youth Fund Project for Humanities and Social Sciences Research of China's Ministry of Education (24YJCZH412)； Chongqing Social Science Planning Project (2025PY39)。

PDF

摘要： 夜间无人机航拍图像存在照度低、噪声强和目标小等固有缺陷，致使车辆检测的误检率与漏检率持续偏高。为此，本文提出一种夜间航拍车辆检测算法NS-YOLOv8（Night Spin-You Only Look Once version8），旨在提升低照度环境下对航拍小目标车辆的检测精度。该算法在输入端采用两阶段图像增强策略，通过噪声抑制与色彩还原对图像进行修复与增强。在此基础上，网络主体引入一种精简的4层特征金字塔架构以减少冗余并优化特征表达 ；进而，在颈部网络嵌入可变注意力机制，增强模型对目标区域的关注能力；最终，通过对骨干与颈部网络的交叉融合，实现深、浅层级特征的高效互补与整合。试验结果表明，NSYOLOv8检测精度优于Oriented-R-CNN（Oriented Region-based Convolutional Neural Network）、YOLOv6OBB（You Only Look Once version 6- Oriented Bounding Box）等算法，模型的 Precision、mAP@0.5-0.95和F1_score分别达到了96.8%、80.3%和96.5%，参数量和计算量分别为1.3M和12.1G。可视化分析进一步表明，该算法有效减少了夜间客货车的误检与漏检问题，提升了检测置信度，适用于低照度环境下的小目标精确检测任务。

关键词: 智能交通, 夜间小目标检测, 夜间图像增强, 无人机航拍图像, YOLOv8

Abstract: Nighttime UAV(Unmanned Aerial Vehicle) aerial imagery suffers from inherent limitations such as low illumination, strong noise, and small targets, which leads to persistently high false detection and missed detection rates in vehicle monitoring. To address these issues, this paper proposes NS-YOLOv8 (NightSpin-You Only Look Once version 8), a vehicle detection algorithm designed to improve the detection accuracy of small target vehicles in low-light aerial scenarios. The method adopts a two-stage image enhancement strategy at the input end, which performs image restoration and enhancement through noise suppression and color restoration. Furthermore, a streamlined four-layer feature pyramid architecture is introduced into the backbone network to reduce redundancy and optimize feature representation. A deformable attention mechanism is embedded into the neck network to enhance the focus of model on target regions. Finally, cross-layer fusion between the backbone and neck networks achieves effective complementarity and integration of deep and shallow features. Experimental results demonstrate that NS-YOLOv8 outperforms methods such as Oriented R-CNN and YOLOv6-OBB in detection accuracy. It achieves Precision of 96.8%,mAP@0.5-0.95 of 80.3%, and an F1-score of 96.5%, with only 1.3 M parameters and 12.1 GFLOPs. Visualization analyses further confirm that the proposed algorithm effectively reduces false and missed detections of cars and trucks at night, while it improves detection confidence, and makes it suitable for precise small target detection in low-light environments.

Key words: intelligent transportation, nighttime small target detection, nighttime image enhancement, UAV aerial images, YOLOv8

中图分类号: