北京交通大学学报2026，Vol.50Issue(2)：190-199,10.DOI:10.11860/j.issn.1673-0291.20250080
基于改进YOLOv5的车载红外目标检测
Vehicle-mounted infrared object detection based on improved YOLOv5
摘要
Abstract
Infrared imaging technology is not limited by lighting conditions and exhibits strong anti-interference capabilities,enabling stable operation at night and in complex,harsh environments.Therefore,it is a critical technology for vehicle-mounted intelligent systems to achieve all-weather en-vironmental perception.However,infrared images generally suffer from blurred edges and a lack of de-tailed features,making it difficult to accurately detect small-scale targets such as pedestrians and fail-ing to meet the real-time detection requirements of vehicle-mounted scenarios.To address the issues of feature attenuation and ambiguous localization in infrared images,a global context space object de-tection network named Global Context Space-You Only Look Once(GCS-YOLO)is proposed.First,a Global Adaptive Feature Extraction Module(GFEM)is designed to improve the backbone net-work.By introducing a Global Channel Attention(GCA)mechanism and adopting a residual structure with progressively selected kernels,the model adaptively modulates its receptive field to extract fea-ture information at various scales.Second,a Multi-Spatial Attention Feature Pyramid Network(MSA-FPN)is designed.By using a coordinate attention module to enhance deep feature maps con-taining positional information and introducing lateral skip connections to incorporate shallow-layer in-formation,the detection accuracy for small and weak targets is improved.Finally,the Minimum Point Distance Intersection over Union(MPDIoU)loss function is introduced.By minimizing the Euclidean distances between the top-left and bottom-right corners of the predicted and ground-truth bounding boxes,the accuracy of bounding box regression is enhanced.Experimental results demonstrate that,compared to YOLOv5,GCS-YOLO achieves an mAP@0.5 of 81.1％and an mAP@0.5:0.95 of 49.5％on an open-source infrared dataset,representing improvements of 10.4％and 7.4％,respec-tively.Furthermore,it operates at a processing speed of 26.4 FPS,successfully meeting the real-time detection requirements for vehicle-mounted applications.Compared to existing algorithms,GCS-YOLO demonstrates significant advantages in infrared object detection accuracy,providing effective technical support for the all-weather operation of intelligent driving systems.关键词
深度学习/红外图像/目标检测/注意力机制/车载系统Key words
deep learning/infrared image/object detection/attention mechanism/vehicle-mounted system分类
信息技术与安全科学引用本文复制引用
罗红轩,张远航,阮雅端..基于改进YOLOv5的车载红外目标检测[J].北京交通大学学报,2026,50(2):190-199,10.基金项目
江苏省交通运输科技项目(2021Y04-2)Jiangsu Provincial Transportation Science and Technology Project(2021Y04-2) （2021Y04-2）
