Scholarlink

<p>Real-time, All-weather monitoring of unmanned aerial vehicles (UAVs) is crucial for ensuring low-altitude airspace security. Traditional single-modality detection systems often fail under challenging conditions such as night and adverse weather, leading to reduced accuracy. To address this, we propose LDD-YOLO, an improved YOLOv8 algorithm incorporating dynamic feature learning and enhanced feature fusion for dual-modality UAV detection. Our approach utilizes a dual-stream backbone to extract complementary features from infrared and visible modalities, a lightweight C2f-linear deformable convolution (LDC2f) module for improved feature extraction, and a dual feature enhancement (DFE) module to mitigate cross-modal interference. Additionally, we introduce a deformable convolution v4-Dynamic Head (DCNv4-DyHead) detection head to enhance multi-scale perception and localization accuracy. Experimental results on a self-constructed dataset of 11,490 paired infrared-visible UAV images demonstrate that the proposed LDD-YOLO model achieves real-time performance with only 6.43M parameters, demonstrating outstanding detection accuracy under adverse conditions and low-light environments. It achieves a map@50:95 of <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(75.4\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>75.4</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation>, surpassing the baseline by <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(5.2\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>5.2</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation>. Competitive performance is also observed on public benchmarks LLVIP and KAIST, showcasing strong generalization capabilities.</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

A UAV Detection in Complex Environments Method Based on Cross-Modal Fusion of Infrared and Visible images

  • 期刊论文

  • 发表日期：

    2026年1月30日

查看全文
  • Ershen Wang,
  • Jiayue Li,
  • Tengli Yu,
  • Song Xu,
  • Pingping Qu,
  • La Na,
  • Yunhao Chen,
  • Yuming Cheng

摘要

Real-time, All-weather monitoring of unmanned aerial vehicles (UAVs) is crucial for ensuring low-altitude airspace security. Traditional single-modality detection systems often fail under challenging conditions such as night and adverse weather, leading to reduced accuracy. To address this, we propose LDD-YOLO, an improved YOLOv8 algorithm incorporating dynamic feature learning and enhanced feature fusion for dual-modality UAV detection. Our approach utilizes a dual-stream backbone to extract complementary features from infrared and visible modalities, a lightweight C2f-linear deformable convolution (LDC2f) module for improved feature extraction, and a dual feature enhancement (DFE) module to mitigate cross-modal interference. Additionally, we introduce a deformable convolution v4-Dynamic Head (DCNv4-DyHead) detection head to enhance multi-scale perception and localization accuracy. Experimental results on a self-constructed dataset of 11,490 paired infrared-visible UAV images demonstrate that the proposed LDD-YOLO model achieves real-time performance with only 6.43M parameters, demonstrating outstanding detection accuracy under adverse conditions and low-light environments. It achieves a map@50:95 of \(75.4\%\) 75.4 % , surpassing the baseline by \(5.2\%\) 5.2 % . Competitive performance is also observed on public benchmarks LLVIP and KAIST, showcasing strong generalization capabilities.