Scholarlink

<p>Object detection in construction sites poses significant challenges due to densely packed targets, frequent occlusions, varying object scales, and unstructured scene layouts. These difficulties are exacerbated by extreme weather conditions—such as fog, rain, snow, and lightning—which degrade image clarity and obscure object boundaries. Compounding the problem is the scarcity of real-world datasets captured under such adverse conditions, hindering model development and robustness assessment. To overcome this limitation, we develop MultiWXGen, a multimodal weather synthesis module that generates realistic adverse weather scenes based on the MOCS dataset, enabling controlled and diverse robustness testing. We further propose HiPA-DETR, a weather-resilient detector built upon RT-DETR, which incorporates two novel components: HiMANet, a hierarchical semantic aggregator for enhanced multi-scale representation, and PAIFI, a polarity-aware interaction module that improves attention stability, particularly for small or low-texture objects. Extensive experiments validate that our HiPA-DETR achieves state-of-the-art performance on the MOCS and its weather-augmented variants. Furthermore, it demonstrates strong generalization to real-world adverse weather datasets, including BDD100K, DAWN, and ACDC, underscoring its effectiveness for safety-critical construction perception under extreme conditions. Code and pretrained models are available on GitHub: <a href="https://github.com/Zenghuang-Fu/Robust-Detection-in-Complex-Construction-Sites">https://github.com/Zenghuang-Fu/Robust-Detection-in-Complex-Construction-Sites</a>.</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

Robust detection in complex construction sites: HiPA-DETR with weather-aware and cross-domain generalization

  • 期刊论文

  • 发表日期：

    2026年5月14日

查看全文
  • Zenghuang Fu,
  • Jinming Yang,
  • Xiaofeng Han,
  • Zhe Feng,
  • Muyang Zhang,
  • Changwei Wang,
  • Weiliang Meng,
  • Jiguang Zhang,
  • Xiaopeng Zhang

摘要

Object detection in construction sites poses significant challenges due to densely packed targets, frequent occlusions, varying object scales, and unstructured scene layouts. These difficulties are exacerbated by extreme weather conditions—such as fog, rain, snow, and lightning—which degrade image clarity and obscure object boundaries. Compounding the problem is the scarcity of real-world datasets captured under such adverse conditions, hindering model development and robustness assessment. To overcome this limitation, we develop MultiWXGen, a multimodal weather synthesis module that generates realistic adverse weather scenes based on the MOCS dataset, enabling controlled and diverse robustness testing. We further propose HiPA-DETR, a weather-resilient detector built upon RT-DETR, which incorporates two novel components: HiMANet, a hierarchical semantic aggregator for enhanced multi-scale representation, and PAIFI, a polarity-aware interaction module that improves attention stability, particularly for small or low-texture objects. Extensive experiments validate that our HiPA-DETR achieves state-of-the-art performance on the MOCS and its weather-augmented variants. Furthermore, it demonstrates strong generalization to real-world adverse weather datasets, including BDD100K, DAWN, and ACDC, underscoring its effectiveness for safety-critical construction perception under extreme conditions. Code and pretrained models are available on GitHub: https://github.com/Zenghuang-Fu/Robust-Detection-in-Complex-Construction-Sites.