Scholarlink

<p>In complex urban environments with spatial constraints, clutter, and electromagnetic interference, precise and reliable unmanned aerial vehicle landing suffers from positioning errors and communication disruptions, making vision-based landing essential. To address this challenge, we propose Star-YOLO, a lightweight landing-marker detector for UAV applications. The method replaces the backbone with StarNet, reducing parameters (− 62.9%) and FLOPs (− 65.9%); introduces SRCELAN (Split-RepConv Efficient Layer Aggregation Network) to improve multi-scale fusion and information propagation with minimal overhead; and employs LSAD (Light Softmax Adaptive-Weight Downsampling) to preserve fine details and boost small-object accuracy. On an AirSim-based visible-light dataset covering diverse urban scenes and illumination conditions, Star-YOLO achieves 96.2% mAP@0.5 (+ 1.6% over the baseline) and improves CSI from 0.869 to 0.898. The compact 3.00-MB model runs at 365.6 FPS on PC and, after TensorRT acceleration, reaches 95.1% mAP@0.5 at 33.6 FPS on NVIDIA Jetson Orin Nano, exceeding the typical 25–30 FPS threshold for real-time detection.</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

Star-YOLO: A lightweight detection algorithm for UAV landing markers

  • 期刊论文

  • 发表日期：

    2025年10月16日

查看全文
  • Deyang Yu,
  • Ruokun Qu,
  • Chenglong Li,
  • Xijun Liu

摘要

In complex urban environments with spatial constraints, clutter, and electromagnetic interference, precise and reliable unmanned aerial vehicle landing suffers from positioning errors and communication disruptions, making vision-based landing essential. To address this challenge, we propose Star-YOLO, a lightweight landing-marker detector for UAV applications. The method replaces the backbone with StarNet, reducing parameters (− 62.9%) and FLOPs (− 65.9%); introduces SRCELAN (Split-RepConv Efficient Layer Aggregation Network) to improve multi-scale fusion and information propagation with minimal overhead; and employs LSAD (Light Softmax Adaptive-Weight Downsampling) to preserve fine details and boost small-object accuracy. On an AirSim-based visible-light dataset covering diverse urban scenes and illumination conditions, Star-YOLO achieves 96.2% mAP@0.5 (+ 1.6% over the baseline) and improves CSI from 0.869 to 0.898. The compact 3.00-MB model runs at 365.6 FPS on PC and, after TensorRT acceleration, reaches 95.1% mAP@0.5 at 33.6 FPS on NVIDIA Jetson Orin Nano, exceeding the typical 25–30 FPS threshold for real-time detection.