Scholarlink

<p>Real-time perception on unmanned surface vehicles (USVs) necessitates accurate yet lightweight processing under strict latency and power constraints. Although LiDAR acquires sparse 3D point clouds, efficient real-time detectors typically operate on structured, image-like tensor representations derived from raw points, such as point-wise feature maps or sparse feature volumes. In this paper, we propose Attention-guided Reweighted Distillation (ARD), a lightweight framework for USVs detection that jointly distills spatial attention heatmaps and intermediate features within this image-like feature space. Specifically, ARD first aligns the student’s spatial attention with the teacher’s to highlight informative regions, and subsequently applies an attention-guided reweighting strategy to emphasize discriminative features while suppressing background interference during distillation. Experiments on a real-world ship point-cloud dataset collected via the "Dolphin-1" intelligent vessel platform demonstrate that ARD achieves consistent gains over multiple distillation baselines while maintaining real-time inference speed, supporting low-power onboard deployment.</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

ARD: attention-guided reweighted knowledge distillation on image-like feature representations for real-time unmanned surface vehicles detection from LiDAR point clouds

  • 期刊论文

  • 发表日期：

    2026年2月6日

查看全文
  • Jun Bai,
  • Shiyang Li,
  • Guihua Xia,
  • Chunsheng Yang

摘要

Real-time perception on unmanned surface vehicles (USVs) necessitates accurate yet lightweight processing under strict latency and power constraints. Although LiDAR acquires sparse 3D point clouds, efficient real-time detectors typically operate on structured, image-like tensor representations derived from raw points, such as point-wise feature maps or sparse feature volumes. In this paper, we propose Attention-guided Reweighted Distillation (ARD), a lightweight framework for USVs detection that jointly distills spatial attention heatmaps and intermediate features within this image-like feature space. Specifically, ARD first aligns the student’s spatial attention with the teacher’s to highlight informative regions, and subsequently applies an attention-guided reweighting strategy to emphasize discriminative features while suppressing background interference during distillation. Experiments on a real-world ship point-cloud dataset collected via the "Dolphin-1" intelligent vessel platform demonstrate that ARD achieves consistent gains over multiple distillation baselines while maintaining real-time inference speed, supporting low-power onboard deployment.