Scholarlink

<p>Recently, 3D object detection systems have been deployed in hospitals to assist with labor-intensive tasks. However, accurate object detection in hospital environments remains challenging due to environmental complexity and privacy concerns. To address these challenges, we propose SALD-Net, a self-attention-integrated LiDAR-based 3D object detection network. Our framework consists of two stages: Stage 1 generates 3D box proposals using a point-based backbone, while Stage 2 refines them using the unified regional and grid (URG) RoI pooling head. We introduce a backbone-integrated self-attention mechanism (BAM) to extract representative points and enhance feature discrimination. Additionally, the RoI feature-based self-attention mechanism (RAM) within the URG RoI pooling head improves localization accuracy, particularly for occluded objects of varying sizes. A point cloud augmentation strategy further mitigates class imbalance, enhancing detection performance across all categories. Extensive experiments on a real hospital point cloud dataset demonstrate that our proposed method outperforms existing methods by a significant margin, improving both detection accuracy and robustness in cluttered hospital 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

SALD-Net: Self-attention-integrated LiDAR-based 3D object detection network in a crowded hospital environment

  • 期刊论文

  • 发表日期：

    2025年9月22日

查看全文
  • Goeun Kim,
  • Su Yang,
  • Ji Yong Han,
  • Jun-Min Kim,
  • Won-Jin Yi

摘要

Recently, 3D object detection systems have been deployed in hospitals to assist with labor-intensive tasks. However, accurate object detection in hospital environments remains challenging due to environmental complexity and privacy concerns. To address these challenges, we propose SALD-Net, a self-attention-integrated LiDAR-based 3D object detection network. Our framework consists of two stages: Stage 1 generates 3D box proposals using a point-based backbone, while Stage 2 refines them using the unified regional and grid (URG) RoI pooling head. We introduce a backbone-integrated self-attention mechanism (BAM) to extract representative points and enhance feature discrimination. Additionally, the RoI feature-based self-attention mechanism (RAM) within the URG RoI pooling head improves localization accuracy, particularly for occluded objects of varying sizes. A point cloud augmentation strategy further mitigates class imbalance, enhancing detection performance across all categories. Extensive experiments on a real hospital point cloud dataset demonstrate that our proposed method outperforms existing methods by a significant margin, improving both detection accuracy and robustness in cluttered hospital environments.