摘要
道路缺陷检测作为保障自动驾驶安全的重要问题,现有道路缺陷检测方法无法满足自动驾驶精准检测道路缺陷需求。且现有的评估模型仅考虑道路缺陷的严重程度,没有考虑到缺陷距离的影响。为了解决这一问题,提出一种改进YOLOv8的道路缺陷检测和自动驾驶风险评估方法。在网络中不同位置引入注意力机制,改进YOLOv8网络并融合了双目视觉和SGBM算法。通过对比实验,得出符合自动驾驶场景的最优模型,比原始网络平均准确度提高1.31%,实现实时检测道路缺陷的类型及距离。根据检测和评估结果,对道路缺陷进行自动驾驶风险等级判定,制定了基本行驶策略。
Road defect detection is crucial to ensure the safety of autonomous driving.However,current road defect detection methods fail to meet the demand of automatic driving to accurately detect road defects.Moreover,existing evaluation models primarily consider the severity of road defects without taking the impact of defect distance into account.To address these issues,this paper proposes an approach for road defect detection and autonomous driving risk assessment using improved YOLOv8.The attention mechanism is incorporated at different network positions.The optimal model for autonomous driving scenarios is identified,improving by 1.31%in mean average precision compared to the original network.Furthermore,our approach enables real-time detection of road defect types and distances.Based on the detection and evaluation results,the risk level for autonomous driving on defect roads is determined and a fundamental strategy for shunning road defects is formulated.
作者
潘明章
袁乐艺
万振华
梁璐
苟轩源
曹鑫鑫
PAN Mingzhang;YUAN Leyi;WAN Zhenhua;LIANG Lu;GOU Xuanyuan;CAO Xinxin(College of Mechanical Engineering,Guangxi University,Nanning 530004,China)
出处
《重庆理工大学学报（自然科学）》
CAS
北大核心
2024年第9期67-74,共8页
Journal of Chongqing University of Technology：Natural Science
基金
国家自然科学基金区域联合重点项目(U23A202599)
广西大学甘蔗专项科研项目(2022GZB008)。
关键词
道路缺陷
自动驾驶
神经网络
注意力机制
双目视觉
road defect
autonomous driving
neural network
attention mechanism
binocular vision
