Scholarlink

Different from the pedestrians from natural view, pedestrians from UAV view have smaller scale and complex motion state. Deformable DETR based on Transformer can effectively detect pedestrians, but still has some defects such as slow convergence speed and high hardware performance requirements. This paper improves these infect by omitting the decoder part, and integrating the local attention module and proposes DETR for UAV (UAV-DETR). Compared with Deformable DETR, the parameters of UAV-DETR network are reduced by 30 \(\%\) and the small object detection accuracy is improved by 10 \(\%\) percentage points. Experimental results on the VisDrone dataset demonstrate the effectiveness of UAV-DETR.

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

Improved DETR for Pedestrian Detection from the Perspective of UAV

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Zhengshan Wang,
  • Haoran Jin,
  • Chengyao Xue,
  • Tengchao Zhang,
  • Long Chen

摘要

Different from the pedestrians from natural view, pedestrians from UAV view have smaller scale and complex motion state. Deformable DETR based on Transformer can effectively detect pedestrians, but still has some defects such as slow convergence speed and high hardware performance requirements. This paper improves these infect by omitting the decoder part, and integrating the local attention module and proposes DETR for UAV (UAV-DETR). Compared with Deformable DETR, the parameters of UAV-DETR network are reduced by 30 \(\%\) and the small object detection accuracy is improved by 10 \(\%\) percentage points. Experimental results on the VisDrone dataset demonstrate the effectiveness of UAV-DETR.