Scholarlink

<p>Lane line detection is crucial for road environment understanding and traffic monitoring, especially in UAV-based road inspections. This study introduces an improved UNet model, DE-UNet, incorporating a dual-branch attention convolution (DBAC) module and an efficient multi-scale attention aggregation (EMSAA) module. The DBAC module enhances feature extraction by integrating parallel dilated and standard convolutions with attention mechanisms, while the EMSAA module fuses multi-level encoder features for cross-scale representation. To evaluate the model’s performance, this study created the UAV-Laneline3K dataset with more than 3000 finely annotated lane line images captured by DJI Mavic 3&#xa0;T at 80–100&#xa0;m and 100–120&#xa0;m. Experiments demonstrate that the proposed method achieves superior performance with an F1-score of 84.11% and mIoU of 75.16% at 80–100&#xa0;m altitudes, and 82.90% and 72.40% at 100–120&#xa0;m, outperforming traditional methods. This research contributes to the development of robust lane line detection systems in complex road scenarios.</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

DE-UNet: an enhanced UNet with dual-branch attention convolution and efficient multi-scale attention aggregation for UAV lane line segmentation

  • 期刊论文

  • 发表日期：

    2026年2月15日

查看全文
  • Yuehao Wang,
  • Haiqing Liu,
  • Lei Yu,
  • Mengmeng Zhang,
  • Haonan Wang

摘要

Lane line detection is crucial for road environment understanding and traffic monitoring, especially in UAV-based road inspections. This study introduces an improved UNet model, DE-UNet, incorporating a dual-branch attention convolution (DBAC) module and an efficient multi-scale attention aggregation (EMSAA) module. The DBAC module enhances feature extraction by integrating parallel dilated and standard convolutions with attention mechanisms, while the EMSAA module fuses multi-level encoder features for cross-scale representation. To evaluate the model’s performance, this study created the UAV-Laneline3K dataset with more than 3000 finely annotated lane line images captured by DJI Mavic 3 T at 80–100 m and 100–120 m. Experiments demonstrate that the proposed method achieves superior performance with an F1-score of 84.11% and mIoU of 75.16% at 80–100 m altitudes, and 82.90% and 72.40% at 100–120 m, outperforming traditional methods. This research contributes to the development of robust lane line detection systems in complex road scenarios.