无人机视角下车辆小目标智能检测算法研究  

Research on Intelligent Detection Algorithms for Small Vehicle Targetsfrom the Perspective of Unmanned Aerial Vehicle

作　　者：王海玮 蔡耿基 张为 游峰[2] WANG Haiwei;CAI Gengji;ZHANG Wei;YOU Feng(Guangdong Communication Polytechnic,Guangzhou 510650,Guangdong,China;School of Civil Engineering,South China University of Technology,Guangzhou 510641,Guangdong,China;State Key Lab of Intelligent Transportation System,Beijing 100088,China)

机构地区：[1]广东交通职业技术学院,广东广州510650 [2]华南理工大学土木与交通学院,广东广州510641 [3]车路一体智能交通全国重点实验室,北京100088

出　　处：《广东交通职业技术学院学报》2025年第1期57-63,共7页Journal of Guangdong Communication Polytechnic

基　　金：广东省自然科学基金(编号:2024A1515012115);中国高校产学研创新基金(编号:2021ZYB04003);车路云重点实验室开放课题(编号:2024-B011);广东省普通高校特色创新类项目(编号:2022KTSCX256)。

摘　　要：随着低空经济与无人机技术的快速发展,无人机可作为获取交通信息的新型载具。针对无人机视角下车辆小目标数据集较为稀缺的问题,构建了包括客车、货车和巴士三类的车辆小目标检测数据集。对YOLO系列网络在无人机场景中的适配性进行系统对比,最终优选YOLOv8网络,作为车辆小目标检测的核心框架。为全面评估算法性能,论文构建了基于准确率、召回率、平均精度均值、Giga浮点运算量及权重文件大小等的多维度评价指标。在此基础上,在YOLOv8m模型上增加了适用于小目标检测的优化层,构建YOLOv8m-p2网络。实验结果表明,YOLOv8m-p2在保持较低运算与存储开销的同时,大幅提升了无人机航拍下车辆小目标的识别精度与实时性,并在小目标捕捉能力与预测准确度方面予以平衡。With the rapid development of the low-altitude economy and UAV technology,unmanned aerial vehicles(UAVs)have emerged as novel platforms for acquiring traffic information.To address the scarcity of small vehicle target datasets from the UAV perspective,a dataset was constructed comprising three vehicle categories:cars,trucks,and buses.A systematic comparison of the YOLO series networks was conducted to evaluate their suitability for UAV scenarios,ultimately selecting YOLOv8 as the core framework for small vehicle target detection.To comprehensively assess algorithm performance,a multi-dimensional evaluation metric was developed,incorporating precision,recall,mean average precision(mAP),Giga Floating Point Operations(GFLOPs),and model weight file size.On this basis,the YOLOv8m model was enhanced with additional layers optimized for small target detection,resulting in the YOLOv8m-p2 network.The experimental results demonstrate that YOLOv8m-p2 significantly improves the recognition accuracy and real-time performance of small vehicle targets in UAV imagery while maintaining low computational and storage overheads,effectively balancing small target capture capability and prediction accuracy.

关 键 词：无人机 车辆数据集 目标检测 YOLO 低空经济 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] U495[自动化与计算机技术—计算机科学与技术]