Scholarlink

<p>To address the limitations of traditional lane line image processing methods in complex traffic scenarios and the challenges of computational efficiency and global information integration faced by existing deep learning models in vehicle platform deployment, this paper proposes NeoLane, a multi-stage 2D lane detection framework for complex traffic scenarios. By integrating the RepVGGplus feature aggregation network, a sliding-window sampling-driven proposal generation mechanism, and a local feature interaction strategy, the framework achieves synergistic optimization of global information supplementation, instance differentiation enhancement, and geometric consistency modeling. Specifically: (1) A lightweight multi-level feature integration framework(Feature Integrate Module, FIM) is constructed using the RepVGGplus network, significantly reducing the computational complexity of feature integration. (2) A sliding-window-based local proposal generation method(Lane Navigate Module, LNM) is designed, which models global lane relationships through cross-attention, overcoming the modeling limitations of anchor-based methods for complex lane shapes. (3) A local feature interaction mechanism(Feature Refine Module, FRM) is proposed to fuse global information in a hierarchical progressive manner, avoiding the high computational overhead of traditional RNN/Transformer architectures. Experiments demonstrate that NeoLane achieves F1-scores of 98.13% and 80.1% on the Tusimple and CULane datasets, respectively. While maintaining detection accuracy comparable to state-of-the-art methods, it significantly improves inference speed with a computational complexity of 13.8 GFLOPs, achieving 177 FPS on an RTX 3080 GPU, providing an effective solution for real-time deployment on vehicle-mounted platforms.</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

Navigate, extract, and optimize: a multi-stage framework for 2D lane detection

  • 期刊论文

  • 发表日期：

    2025年9月22日

查看全文
  • Runshuai Ge,
  • Peicheng Shi,
  • Xinlong Dong,
  • Chadia Chakir,
  • Taonian Liang,
  • Aixi Yang

摘要

To address the limitations of traditional lane line image processing methods in complex traffic scenarios and the challenges of computational efficiency and global information integration faced by existing deep learning models in vehicle platform deployment, this paper proposes NeoLane, a multi-stage 2D lane detection framework for complex traffic scenarios. By integrating the RepVGGplus feature aggregation network, a sliding-window sampling-driven proposal generation mechanism, and a local feature interaction strategy, the framework achieves synergistic optimization of global information supplementation, instance differentiation enhancement, and geometric consistency modeling. Specifically: (1) A lightweight multi-level feature integration framework(Feature Integrate Module, FIM) is constructed using the RepVGGplus network, significantly reducing the computational complexity of feature integration. (2) A sliding-window-based local proposal generation method(Lane Navigate Module, LNM) is designed, which models global lane relationships through cross-attention, overcoming the modeling limitations of anchor-based methods for complex lane shapes. (3) A local feature interaction mechanism(Feature Refine Module, FRM) is proposed to fuse global information in a hierarchical progressive manner, avoiding the high computational overhead of traditional RNN/Transformer architectures. Experiments demonstrate that NeoLane achieves F1-scores of 98.13% and 80.1% on the Tusimple and CULane datasets, respectively. While maintaining detection accuracy comparable to state-of-the-art methods, it significantly improves inference speed with a computational complexity of 13.8 GFLOPs, achieving 177 FPS on an RTX 3080 GPU, providing an effective solution for real-time deployment on vehicle-mounted platforms.