基于多尺度注意力的YOLOv9红外车辆检测方法研究  

Research on Infrared Vehicle Detection Method Based on YOLOv9 with Multi-scale Attention

作　　者：段锐 龚加兴 韩坤林 斯新华 DUAN Rui;GONG Jiaxing;HAN Kunlin;SI Xinhua(Zhaotong Expressway Construction and Maintenance Engineering Co.,Ltd.,Yunnan Zhaotong 657013;School of Computer and Internet of Things,Chongqing Institute of Engineering,Chongqing 400056;China Merchants Chongqing Highway Engineering Testing Center Co.,Ltd.,Chongqing 400067)

机构地区：[1]昭通高速公路建设养护工程有限公司,云南昭通657013 [2]重庆工程学院计算机与物联网学院,重庆400056 [3]招商局重庆公路工程检测中心有限公司,重庆400067

出　　处：《公路交通技术》2025年第2期189-194,202,共7页Technology of Highway and Transport

摘　　要：为解决红外场景下车辆检测目标的准确性问题,促进道路交通安全的智能化与集成化,采用YOLOv9-c目标检测算法作为基准模型,通过加入多尺度注意力模块,改进了对红外车辆目标的特征提取能力,并将训练好的检测模型部署在边缘计算设备树莓派上验证红外车辆的检测任务。结果表明,YOLOv9-c-SEMA在添加多尺度注意力模块后,检测准确率提升了6.7个百分点,且检测速度仍能达100 FPS(帧数/s),可满足高速公路场景下的红外车辆检测需求。该方法可供红外车辆车型识别参考。In order to solve the accuracy problem of vehicle detection targets in infrared scenes and promote the intelligence and integration of road traffic safety,the YOLOv9-c target detection algorithm is used as the benchmark model.By adding a multi-scale attention module,the feature extraction capability of infrared vehicle targets is improved,and the trained detection model is deployed on the edge computing device Raspberry Pi to verify the infrared vehicle detection task.The results show that after adding the multi-scale attention module,the detection accuracy of YOLOv9-c-SEMA is improved by 6.7 percentage points,and the detection speed can still reach 100 FPS,which can meet the infrared vehicle detection needs in highway scenes and provide a method reference for infrared vehicle model recognition.

关 键 词：YOLOv9-c 多尺度注意力 边缘计算 车辆检测 

分 类 号：U495[交通运输工程—交通运输规划与管理]