首页|期刊导航|飞控与探测|恶劣天气下基于可见光与红外融合的车辆目标检测方法

恶劣天气下基于可见光与红外融合的车辆目标检测方法OA

Method of Vehicle Target Detection Based on Visible-Infrared Fusion Under Adverse Weather Conditions

中文摘要英文摘要

针对雨、雪、雾等恶劣天气下车辆目标检测精度与鲁棒性不足的问题,提出一种基于伪孪生架构的可见光-红外双模态融合车辆目标检测方法.该方法以YOLOv8为基础框架,通过差异化设计双模态特征提取分支增强特征:可见光分支嵌入通道空间双增强注意力(Channel-Space Dual-Enhancement Attention,CSDEA)模块,强化车辆纹理语义特征辨识度;红外分支配置双维并行池化注意力(Dual-Dimensional Parallel Pooling Attention,DDPPA)模块,精准捕捉热辐射特征并抑制噪声干扰.同时,设计交叉注意力特征融合(Cross-Attention Feature Fusion,CAFF)模块,建立双模态特征的精准关联与双向增强,实现互补信息高效整合.在自制双模态数据集上的实验结果表明,所提方法在交并比(Intersection over Union,IoU)阈值为50%时的平均精度均值(Mean Average Precision,mAP)mAP50及交并比阈值从50%到95%、以5%为步长的平均精度均值mAP50:95指标上,均显著优于现有公开双模态算法,在小目标检出、遮挡目标识别、低能见度场景抗干扰等方面均表现出显著优势,有效提升了恶劣天气下车辆目标检测的精度与稳定性.

To address the issue of insufficient accuracy and robustness of vehicle target detection in adverse weather conditions such as rain,snow,and fog,a visible-infrared fusion vehicle target de-tection method based on a pseudo-twin architecture is proposed.This method takes YOLOv8 as the basic framework and enhances features through the differentiated design of dual-modal feature extraction branches:the visible light branch embeds the CSDEA module to strengthen the recogni-tion of vehicle texture semantic features;the infrared branch is equipped with the DDPPA module to accurately capture thermal radiation features and suppress noise interference.Meanwhile,a CAFF module is designed to establish precise correlation and bidirectional enhancement of dual-modal features,achieving efficient integration of complementary information.Experimental results on a self-made dual-modal dataset show that the proposed method significantly outperforms existing public dual-modal algorithms in terms of mAP50 and mAP50:95 metrics,demonstrating significant advantages in small target detection,occluded target recognition,and anti-interference in low visibility scenarios,effectively improving the accuracy and stability of vehicle target detec-tion in adverse weather conditions.

闫佳灿;田甜;赵天明;宋佳民

华中科技大学人工智能与自动化学院·武汉·430074||华中科技大学多谱信息智能处理技术全国重点实验室·武汉·430074华中科技大学人工智能与自动化学院·武汉·430074||华中科技大学多谱信息智能处理技术全国重点实验室·武汉·430074华中科技大学人工智能与自动化学院·武汉·430074||华中科技大学多谱信息智能处理技术全国重点实验室·武汉·430074华中科技大学人工智能与自动化学院·武汉·430074||华中科技大学多谱信息智能处理技术全国重点实验室·武汉·430074

信息技术与安全科学

车辆目标检测双模态融合伪孪生架构

vehicle target detectiondual-modal fusionpseudo-siamese architecture

《飞控与探测》 2026 (4)

93-103,11

国家自然科学基金(42071339)

10.20249/j.cnki.2096-5974.2026.04.008

评论