Scholarlink

<p>Lane detection is a crucial component in autonomous driving systems, yet challenges persist in achieving accurate detection and lightweight deployment of 3D lane detection on highways. To address these issues, we propose a novel striped lane representation that more closely reflects the real-world characteristics of lane lines. Additionally, we introduce a double-branch cross-layer refinement network, designed to enhance model robustness while accelerating training convergence. To further enhance detection performance, we develop the stripes IOU loss function, specifically tailored for evaluating striped lane representations. To simplify deployment, we implement a prior-based spatial projection correction mechanism, effectively mapping 2D detection results into 3D space. Our algorithm achieves a 97% accuracy on the TuSimple and an average F1-score of 79.9 on the CULane, all while maintaining remarkable runtime efficiency. Furthermore, on the 3D OpenLane, it delivers robust detection performance in critical adjacent regions while sustaining high detection efficiency. These results offer substantial support for the effectiveness and practical feasibility of our approach on highway.</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

DCRnet: an effective 2D and 3D lane detection method on highway

  • 期刊论文

  • 发表日期：

    2025年5月22日

查看全文
  • Long Zhao,
  • Xiaoye Liu,
  • Linxiang Li

摘要

Lane detection is a crucial component in autonomous driving systems, yet challenges persist in achieving accurate detection and lightweight deployment of 3D lane detection on highways. To address these issues, we propose a novel striped lane representation that more closely reflects the real-world characteristics of lane lines. Additionally, we introduce a double-branch cross-layer refinement network, designed to enhance model robustness while accelerating training convergence. To further enhance detection performance, we develop the stripes IOU loss function, specifically tailored for evaluating striped lane representations. To simplify deployment, we implement a prior-based spatial projection correction mechanism, effectively mapping 2D detection results into 3D space. Our algorithm achieves a 97% accuracy on the TuSimple and an average F1-score of 79.9 on the CULane, all while maintaining remarkable runtime efficiency. Furthermore, on the 3D OpenLane, it delivers robust detection performance in critical adjacent regions while sustaining high detection efficiency. These results offer substantial support for the effectiveness and practical feasibility of our approach on highway.