Scholarlink

<p>Monitoring and maintaining railway corridors requires accurate, high-resolution spatial data to ensure operational safety and efficiency. However, data-driven railway infrastructure assessment has been limited by the scarcity of large-scale, finely annotated 3D benchmarks. To address this gap we first introduce <i>SemanticRail3D</i>, a mobile-LiDAR dataset of 438 high-resolution point clouds (approximately 2.8 billion points) in 200 m segments, annotated via a heuristic rule-based segmentation method into 12 semantic classes and grouped into instance labels, with per-point intensity information. Building on this foundation, we present <i>SemanticRail3D-V2</i>, featuring a Machine Learning (ML)-ready preprocessing pipeline that aligns and sections raw scans into uniform blocks, and a novel evaluation protocol combining metric-based anomaly detection with a probabilistic validity analysis of class distributions and spatial relationships. Railway domain experts then reviewed each block to remove low-quality scans and divide the dataset into five training shards-selected for annotation accuracy and complexity-plus separate validation and held-out test sets. The <i>SemanticRail3D dataset, together with the V2 enhancements,</i> offers a rigorously curated, richly annotated benchmark for semantic and instance segmentation in railway environments, supporting research in condition monitoring and asset management.</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

SemanticRail3D - A Mobile LiDAR Benchmark for Semantic and Instance Segmentation of Railway Corridors

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月18日

查看全文
  • Arshia Ghasemlou,
  • Mario Soilán,
  • Belén Riveiro

摘要

Monitoring and maintaining railway corridors requires accurate, high-resolution spatial data to ensure operational safety and efficiency. However, data-driven railway infrastructure assessment has been limited by the scarcity of large-scale, finely annotated 3D benchmarks. To address this gap we first introduce SemanticRail3D, a mobile-LiDAR dataset of 438 high-resolution point clouds (approximately 2.8 billion points) in 200 m segments, annotated via a heuristic rule-based segmentation method into 12 semantic classes and grouped into instance labels, with per-point intensity information. Building on this foundation, we present SemanticRail3D-V2, featuring a Machine Learning (ML)-ready preprocessing pipeline that aligns and sections raw scans into uniform blocks, and a novel evaluation protocol combining metric-based anomaly detection with a probabilistic validity analysis of class distributions and spatial relationships. Railway domain experts then reviewed each block to remove low-quality scans and divide the dataset into five training shards-selected for annotation accuracy and complexity-plus separate validation and held-out test sets. The SemanticRail3D dataset, together with the V2 enhancements, offers a rigorously curated, richly annotated benchmark for semantic and instance segmentation in railway environments, supporting research in condition monitoring and asset management.