Scholarlink

<p>The ability to traverse off-road landscapes is essential for mobile robots and self-driving cars to navigate uneven terrain, particularly in rural or isolated areas lacking infrastructure. Contemporary segmentation solutions have high computational demands, making them unsuitable for low-resource embedded devices used in farming or rescue operations. To address this, we introduce LORNet, a lightweight unstructured off-road segmentation model designed for efficient operation on embedded systems. LORNet features a dual-path encoder, multi-modal cross fusion (MCF), criss-cross attention, and RepGhost bottleneck layers to improve segmentation accuracy while reducing complexity. Pointwise convolutions ensure low inference latency. Experiments on ORFD, RUGD, and RELLIS-3D demonstrate that LORNet achieves an mIoU of 0.913 on ORFD with only 80K parameters and runs at up to 20 FPS on embedded devices. These results show that LORNet outperforms existing models in both accuracy and efficiency, offering a practical solution for off-road navigation in resource-constrained environments.</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

LORNet: lightweight unstructured off-road segmentation on embedded devices for mobile robotic and intelligent vehicle

  • 期刊论文

  • 发表日期：

    2026年1月12日

查看全文
  • Igi Ardiyanto,
  • Mako Takahama,
  • Jun Miura

摘要

The ability to traverse off-road landscapes is essential for mobile robots and self-driving cars to navigate uneven terrain, particularly in rural or isolated areas lacking infrastructure. Contemporary segmentation solutions have high computational demands, making them unsuitable for low-resource embedded devices used in farming or rescue operations. To address this, we introduce LORNet, a lightweight unstructured off-road segmentation model designed for efficient operation on embedded systems. LORNet features a dual-path encoder, multi-modal cross fusion (MCF), criss-cross attention, and RepGhost bottleneck layers to improve segmentation accuracy while reducing complexity. Pointwise convolutions ensure low inference latency. Experiments on ORFD, RUGD, and RELLIS-3D demonstrate that LORNet achieves an mIoU of 0.913 on ORFD with only 80K parameters and runs at up to 20 FPS on embedded devices. These results show that LORNet outperforms existing models in both accuracy and efficiency, offering a practical solution for off-road navigation in resource-constrained environments.