改进YOLOv8算法的机场外来物检测研究  

Research on Airport Foreign Object Detection Based on Improved YOLOv8 Algorithm

作　　者：郭九霞[1] 李金润 王义龙 李静远 唐锐 GUO Jiuxia;LI Jinrun;WANG Yilong;LI Jingyuan;TANG Rui(School of Air Traffic Management,Civil Aviation Flight University of China,Guanghan 618307;Recruitment Office,Civil Aviation Flight University of China,Guanghan 618307;Flight Planning Department,Civil Aviation Administration Operation Monitoring Center,Beijing 100710)

机构地区：[1]中国民用航空飞行学院空中交通管理学院,广汉618307 [2]中国民用航空飞行学院招生处,广汉618307 [3]民航局运行监控中心飞行计划处,北京100710

出　　处：《舰船电子工程》2025年第3期119-125,共7页Ship Electronic Engineering

基　　金：国家自然科学基金项目(编号:72201268);四川省自然科学基金项目(编号:U2333207);四川省社会科学基金项目(编号:SCJJ23ND186);中央高校基本科研业务费专项资金(编号:PHD2023-041);民航教育人才类项目(编号:MHJY2023010);中央高校教育教学改革专项资金(编号:E2024024)资助。

摘　　要：为解决机场外来物检测方法存在检测稳定性差、漏检的问题,论文使用YOLOv8算法进行改进。首先,使用动态卷积ODConv,通过引入可学习的形变模块,动态调整卷积核的形状、大小及通道维度,优化卷积过程并专注于机场外来物的形状大小和尺度变化,实现对图像特征信息的高效提取;其次,设计了C2f_DAConv模块,降低了算法的参数量;然后,在PANet网络架构的基础上,融合主干网络的P2特征层,并将PANet网络架构更改为BiFPN,该网络实现了底层细节特征信息和高层语义特征信息的高效融合,减少了外来物目标特征的信息丢失;最后,为解决预测框与目标框之间的定位误差问题,更改损失函数为Inner SIoU,优化了算法的计算过程,加快了算法训练的收敛速度,同时提升了算法的检测精度。实验结果表明,改进的算法相比原YOLOv8算法,其参数量降低了35.5%,平均精度均值(mAP)达到97.3%,提升了2.0%,召回率(Re-call)为95.5%,提升了5.2%;对比分析F1曲线、P-R曲线和Recall曲线,表明改进的算法在检测稳定性方面有显著提升,能有效解决机场外来物的漏检问题。In order to solve the problems of poor detection stability and missed detection in the airport foreign object detection method,this paper improves the YOLOv8 algorithm.Firstly,dynamic convolution(ODConv)is used by introducing a learnable de-formation module that dynamically adjusts the shape,size,and channel dimension of the convolution kernel.This optimization of the convolution process focuses on the shape,size,and scale changes of airport foreign objects,achieving efficient extraction of im-age feature information.Secondly,the C2f_DAConv module is designed to reduce the number of parameters in the algorithm.Then,based on the PANet network architecture,the P2 feature layer of the backbone network is fused,and the PANet network architec-ture is changed to BiFPN.The network realizes the efficient fusion of low-level detail feature information and high-level semantic feature information,reducing information loss of foreign object target features.Finally,to solve the positioning error problem be-tween the prediction box and the target box,the loss function is changed to Inner SIoU,optimizing the calculation process of the al-gorithm,accelerating the convergence speed of algorithm training,and improving the detection accuracy of the algorithm.The exper-imental results show that the improved algorithm has 35.5%fewer parameters than the original YOLOv8 algorithm.The mean aver-age precision(mAP)reaches 97.3%,an increase of 2.0%,and the recall rate(Recall)is 95.5%,an increase of 5.2%.Compara-tive analysis of the F1 curve,P-R curve,and Recall curve shows that the improved algorithm significantly improves detection stability and effectively solves the problem of missed detection of foreign objects at airports.

关 键 词：改进YOLOv8算法 FOD检测 动态卷积 机场安全 

分 类 号：X949[环境科学与工程—安全科学] V351.11[航空宇航科学与技术—人机与环境工程]