Scholarlink

<p>Lane detection is a pivotal task in autonomous driving and advanced driver assistance systems, aiming to identify drivable areas and lane boundaries. Despite significant progress, challenges persist in recognizing thin distant lanes and handling dense-lane regions. This study introduces PolarNet, an improved lane detection model leveraging polar-coordinate modeling. PolarNet incorporates a polar-coordinate convolutional block attention module (Polar-CBAM) to enhance feature representation by decoupling spatial attention into angular and radial directions. Additionally, a bidirectional feature pyramid network (BiFPN) is employed for adaptive cross-scale feature fusion, improving the representation of small-scale lanes. A geometry-aware dynamic weighted focal loss is designed to assign dynamic training weights based on vanishing-point distance and lane density, enhancing learning capability for challenging regions. Experiments are conducted on three public benchmarks, including TuSimple, CULane, and CurveLanes. PolarNet achieves an F1@50-score of 81.56% on CULane, an accuracy of 96.41% on TuSimple, and an F1@50-score of 87.26% on CurveLanes, while maintaining computational complexity comparable to that of Polar R-CNN. Experimental results demonstrate the effectiveness of the proposed method. The source code is publicly available at <a href="https://github.com/aaaaalai/PolarNet">https://github.com/aaaaalai/PolarNet</a>. A permanent archived version is available via Zenodo at <a href="https://doi.org/10.5281/zenodo.20729047">https://doi.org/10.5281/zenodo.20729047</a>.</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

PolarNet: enhanced polar-coordinate lane detection via feature fusion and geometry-aware adaptive weighting

  • 期刊论文

  • 发表日期：

    2026年7月12日

查看全文
  • Xianying Lai,
  • Ling Li,
  • Yuxiang Zhang

摘要

Lane detection is a pivotal task in autonomous driving and advanced driver assistance systems, aiming to identify drivable areas and lane boundaries. Despite significant progress, challenges persist in recognizing thin distant lanes and handling dense-lane regions. This study introduces PolarNet, an improved lane detection model leveraging polar-coordinate modeling. PolarNet incorporates a polar-coordinate convolutional block attention module (Polar-CBAM) to enhance feature representation by decoupling spatial attention into angular and radial directions. Additionally, a bidirectional feature pyramid network (BiFPN) is employed for adaptive cross-scale feature fusion, improving the representation of small-scale lanes. A geometry-aware dynamic weighted focal loss is designed to assign dynamic training weights based on vanishing-point distance and lane density, enhancing learning capability for challenging regions. Experiments are conducted on three public benchmarks, including TuSimple, CULane, and CurveLanes. PolarNet achieves an F1@50-score of 81.56% on CULane, an accuracy of 96.41% on TuSimple, and an F1@50-score of 87.26% on CurveLanes, while maintaining computational complexity comparable to that of Polar R-CNN. Experimental results demonstrate the effectiveness of the proposed method. The source code is publicly available at https://github.com/aaaaalai/PolarNet. A permanent archived version is available via Zenodo at https://doi.org/10.5281/zenodo.20729047.