Scholarlink

<p>Detecting unmanned aerial vehicles (UAVs), especially micro-UAVs, in low-altitude surveillance remains challenging because targets are small and easily confused with birds and background clutter. To address this bird–UAV coexistence setting, we construct the Bird–UAV Fusion dataset by combining a laboratory-curated Bird–UAV image collection with a public UAV dataset. Based on YOLOv8s, we propose a micro-UAV detector that extends the original P3–P5 head to P2–P5, where the stride-4 P2 branch preserves fine spatial details without introducing a more expensive P1 branch. A Residual Channel–Spatial Attention Block (RCSAB) is inserted into the P2 path to enhance small-target responses, and the Feature Complementary Mapping Module (FCM) is incorporated in the backbone and P3<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\rightarrow \)</EquationSource> <EquationSource Format="MATHML"><math> <mo stretchy="false">→</mo> </math></EquationSource> </InlineEquation>P2 path to promote semantic–spatial interaction. On Bird–UAV Fusion, the proposed method achieves mAP@0.5:0.95 of 0.786 on the validation set and 0.752 on the test set, improving YOLOv8s by 0.085 and 0.074, respectively. For the UAV category, it reaches 0.709/0.673 on the validation/test sets. The model also achieves 29.6 frames per second, indicating a practical near-real-time accuracy–efficiency trade-off.</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

Micro-UAV detection under bird interference with an enhanced YOLOv8s

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • Deguo Yang,
  • Hao Pan,
  • Ruiguang Cao

摘要

Detecting unmanned aerial vehicles (UAVs), especially micro-UAVs, in low-altitude surveillance remains challenging because targets are small and easily confused with birds and background clutter. To address this bird–UAV coexistence setting, we construct the Bird–UAV Fusion dataset by combining a laboratory-curated Bird–UAV image collection with a public UAV dataset. Based on YOLOv8s, we propose a micro-UAV detector that extends the original P3–P5 head to P2–P5, where the stride-4 P2 branch preserves fine spatial details without introducing a more expensive P1 branch. A Residual Channel–Spatial Attention Block (RCSAB) is inserted into the P2 path to enhance small-target responses, and the Feature Complementary Mapping Module (FCM) is incorporated in the backbone and P3 \(\rightarrow \) → P2 path to promote semantic–spatial interaction. On Bird–UAV Fusion, the proposed method achieves mAP@0.5:0.95 of 0.786 on the validation set and 0.752 on the test set, improving YOLOv8s by 0.085 and 0.074, respectively. For the UAV category, it reaches 0.709/0.673 on the validation/test sets. The model also achieves 29.6 frames per second, indicating a practical near-real-time accuracy–efficiency trade-off.