Scholarlink

<p>UAV surveillance systems require high-performance computing (HPC) to process multi-stream imagery in real time. Small object detection in UAV imagery faces significant challenges: sparse pixels, limited features, complex backgrounds, and strict latency requirements (sub-100&#xa0;ms). We propose MD-DRIFPN, a framework based on YOLOv11s, optimized for real-time small-object detection. The framework includes three key modules: DRIB expands receptive fields through multi-scale dilated convolutions and dual attention; MD-SFPN performs efficient multi-scale feature fusion; and LSDH optimizes detection accuracy via parameter sharing. We also introduce the Focaler-IoU loss function to improve bounding box localization. Experiments on VisDrone datasets show MD-DRIFPN improves baseline performance by 40.9% (AP<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(_{0.50:0.95}\)</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mrow /> <mrow> <mn>0.50</mn> <mo>:</mo> <mn>0.95</mn> </mrow> <mrow /> </mmultiscripts> </math></EquationSource> </InlineEquation> of 0.248) while reducing parameters by 23.9% and achieving 104 FPS on NVIDIA RTX 3090. The model also performs well on DIOR and AI-TOD datasets, demonstrating strong generalization. With only 21.7 GFLOPs computational cost, MD-DRIFPN supports deployment on edge devices and enables parallel processing across multiple UAV streams. The lightweight design and GPU acceleration make it suitable for distributed multi-UAV coordination where synchronized real-time processing is essential.</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

MD-DRIFPN: dilated multi-directional FPN for small drone object detection

  • 期刊论文

  • 发表日期：

    2025年12月12日

查看全文
  • Houyu Luan,
  • Shuobo Xu,
  • Dishi Xu,
  • Lele Liu,
  • Mengwei Guo,
  • Shaoqing Huang

摘要

UAV surveillance systems require high-performance computing (HPC) to process multi-stream imagery in real time. Small object detection in UAV imagery faces significant challenges: sparse pixels, limited features, complex backgrounds, and strict latency requirements (sub-100 ms). We propose MD-DRIFPN, a framework based on YOLOv11s, optimized for real-time small-object detection. The framework includes three key modules: DRIB expands receptive fields through multi-scale dilated convolutions and dual attention; MD-SFPN performs efficient multi-scale feature fusion; and LSDH optimizes detection accuracy via parameter sharing. We also introduce the Focaler-IoU loss function to improve bounding box localization. Experiments on VisDrone datasets show MD-DRIFPN improves baseline performance by 40.9% (AP \(_{0.50:0.95}\) 0.50 : 0.95 of 0.248) while reducing parameters by 23.9% and achieving 104 FPS on NVIDIA RTX 3090. The model also performs well on DIOR and AI-TOD datasets, demonstrating strong generalization. With only 21.7 GFLOPs computational cost, MD-DRIFPN supports deployment on edge devices and enables parallel processing across multiple UAV streams. The lightweight design and GPU acceleration make it suitable for distributed multi-UAV coordination where synchronized real-time processing is essential.