Scholarlink

<p>Lane detection is a critical component in autonomous driving, driver assistance systems, and traffic monitoring systems. However, due to factors, such as lane line wear, occlusion, and large curvature, lane detection still faces significant challenges in balancing accuracy and real-time performance. This paper comprehensively considers the impact of local lane angles and sharp-curvature lanes on anchor-based lane detection methods. It proposes a comprehensive lane Intersection over Union (IoU), which effectively eliminates these influences and is used to calculate loss functions and sample assignment costs, thereby improving the accuracy of confidence scores. Additionally, this paper utilizes lane prior information to initially estimate the local direction of lane lines, generating lane anchor lines based on prior information to more accurately fit the lane line shapes. Finally, a cross-attention mechanism is employed to focus on information from different lane segments, emphasizing the relationships between foreground features and enhancing the network's tolerance to lane variations. The approach is validated on the public datasets CULane and Tusimple, achieving excellent results.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Research on the Curved Lane Detection Method Based on Comprehensive Lane IOU

  • 期刊论文

  • 发表日期：

    2025年8月4日

查看全文
  • Mingheng Zhang,
  • Xinghai Yi,
  • Juntao Li,
  • Xiaoyu Wang,
  • Baozhen Yao

摘要

Lane detection is a critical component in autonomous driving, driver assistance systems, and traffic monitoring systems. However, due to factors, such as lane line wear, occlusion, and large curvature, lane detection still faces significant challenges in balancing accuracy and real-time performance. This paper comprehensively considers the impact of local lane angles and sharp-curvature lanes on anchor-based lane detection methods. It proposes a comprehensive lane Intersection over Union (IoU), which effectively eliminates these influences and is used to calculate loss functions and sample assignment costs, thereby improving the accuracy of confidence scores. Additionally, this paper utilizes lane prior information to initially estimate the local direction of lane lines, generating lane anchor lines based on prior information to more accurately fit the lane line shapes. Finally, a cross-attention mechanism is employed to focus on information from different lane segments, emphasizing the relationships between foreground features and enhancing the network's tolerance to lane variations. The approach is validated on the public datasets CULane and Tusimple, achieving excellent results.