Scholarlink

The proliferation of Unmanned Aerial Vehicles (UAVs) poses escalating security and privacy threats, underscoring the critical need for advanced detection technologies. To address the scarcity of robust visual solutions for ​small-target UAV detection, this work introduces the ​Small-target UAV Dataset—a significant expansion and refinement of the DUT Anti-UAV benchmark. This dataset contains 15,302 high-quality visible-light images, rigorously partitioned into non-overlapping training (10,200 images), validation (2,551), and test sets (2,551), eliminating redundancy prevalent in existing datasets. It notably enriches challenging scenarios, including multi-scale targets against complex backgrounds (e.g., dense foliage) and ultra-small “pixel-level” drones. Leveraging this dataset, we pioneer the application of ​YOLOv11—the state-of-the-art YOLO variant—to anti-UAV detection. Without architectural modifications, our trained model (YOLOv11-SmallUAV) achieves breakthrough performance: ​0.939 precision, ​0.588 recall, ​77.1% single-class AP, and ​48.7% mAP 50–95​ while sustaining 608.2 fps. These metrics significantly surpass baseline models (YOLOv8: 19.4% AP; YOLOv11 trained on DUT: 61.2% AP), demonstrating the dataset’s efficacy and YOLOv11’s adaptability.

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

Vision-Based Anti-UAV Target Detection

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Wang Zhendong,
  • Wei Junyu,
  • Su Shaojing,
  • Zhao Zongqing,
  • Li Shiqi,
  • Huang Jiangjiang

摘要

The proliferation of Unmanned Aerial Vehicles (UAVs) poses escalating security and privacy threats, underscoring the critical need for advanced detection technologies. To address the scarcity of robust visual solutions for ​small-target UAV detection, this work introduces the ​Small-target UAV Dataset—a significant expansion and refinement of the DUT Anti-UAV benchmark. This dataset contains 15,302 high-quality visible-light images, rigorously partitioned into non-overlapping training (10,200 images), validation (2,551), and test sets (2,551), eliminating redundancy prevalent in existing datasets. It notably enriches challenging scenarios, including multi-scale targets against complex backgrounds (e.g., dense foliage) and ultra-small “pixel-level” drones. Leveraging this dataset, we pioneer the application of ​YOLOv11—the state-of-the-art YOLO variant—to anti-UAV detection. Without architectural modifications, our trained model (YOLOv11-SmallUAV) achieves breakthrough performance: ​0.939 precision, ​0.588 recall, ​77.1% single-class AP, and ​48.7% mAP 50–95​ while sustaining 608.2 fps. These metrics significantly surpass baseline models (YOLOv8: 19.4% AP; YOLOv11 trained on DUT: 61.2% AP), demonstrating the dataset’s efficacy and YOLOv11’s adaptability.