Scholarlink

<p>Lane detection is a critical perception component for autonomous vehicles. Despite the emergence of sophisticated deep networks, real-world deployment remains hindered by the conflict between high computational demands and the limited resources of embedded platforms. This paper challenges the conventional trend of increasing architectural depth by proposing LRNet, a minimalist yet robust framework for instance lane detection. Unlike existing hybrid models that rely on heavy feature fusion, LRNet strategically integrates a single, optimized residual block into the terminal downsampling stage of a modified U-Net. This refined synergy allows the model to capture complex global lane topologies while maintaining an exceptionally small memory footprint. Comprehensive experiments on TuSimple and LLAMAS datasets demonstrate that LRNet achieves state-of-the-art accuracy—particularly in challenging scenarios like severe occlusion and unmarked lanes—while occupying only 3.15&#xa0;MB of parameter memory. Our findings reveal that for lane detection, a “leaner” structural intervention at the critical semantic junction can yield superior performance-to-efficiency ratios compared to over-parameterized alternatives.</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

LRNet: a lightweight and robust network for instance lane detection

  • 期刊论文

  • 发表日期：

    2026年6月17日

查看全文
  • Jiyong Zhang,
  • Bo Wang,
  • Si Si

摘要

Lane detection is a critical perception component for autonomous vehicles. Despite the emergence of sophisticated deep networks, real-world deployment remains hindered by the conflict between high computational demands and the limited resources of embedded platforms. This paper challenges the conventional trend of increasing architectural depth by proposing LRNet, a minimalist yet robust framework for instance lane detection. Unlike existing hybrid models that rely on heavy feature fusion, LRNet strategically integrates a single, optimized residual block into the terminal downsampling stage of a modified U-Net. This refined synergy allows the model to capture complex global lane topologies while maintaining an exceptionally small memory footprint. Comprehensive experiments on TuSimple and LLAMAS datasets demonstrate that LRNet achieves state-of-the-art accuracy—particularly in challenging scenarios like severe occlusion and unmarked lanes—while occupying only 3.15 MB of parameter memory. Our findings reveal that for lane detection, a “leaner” structural intervention at the critical semantic junction can yield superior performance-to-efficiency ratios compared to over-parameterized alternatives.