Scholarlink

<p>In this study, we investigate the visual detection challenge of ground-to-air drones. With the widespread use of drones in civilian domains, the demand for counter-drone technologies has increased significantly. Detection serves as a foundation for countermeasures. In the ground-to-air drone visual detection domain, we address two primary challenges: the absence of a comprehensive open-source dataset for drone detection and the lack of multifactor analysis based on such a dataset. To address these challenges, we propose GA-Fly, a new dataset containing 10,800 high-resolution images capturing a DJI Mini 4 Pro drone at various angles, distances, and lighting conditions. The proposed dataset facilitates research on small target drone detection. We assessed the performance of eight representative deep learning algorithms, namely, Cascade R-convolutional neural networks (CNN), Faster R-CNN, Faster R-CNN-FPN, RTMDet, YOLOv8, YOLOv10, SSD, and YOLOv3, using the GA-Fly dataset. Our experiments included evaluating the effects of grayscale images, shooting angles, target scales, and cropping on detection accuracy. Our findings reveal that Faster R-CNN-FPN (0.894) achieves the best mean average precision, which is closely followed by YOLOv3 (0.823), whereas RTMDet (0.370) and YOLOv8 (0.458) achieve the worst and second worst performance, respectively. This study highlights the necessity for specialized algorithms tailored for drone detection given the distinct challenges posed by small targets and diverse environmental conditions. The GA-Fly dataset and the insights gained from these experiments can serve as a benchmark for future research in this area, contributing to the advancement of more effective drone detection systems. The proposed dataset is available at <a href="https://github.com/ballballubsmart/GA-FLY">https:/github.com/ballballubsmart/GA-FLY</a>.</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

Experimental study and evaluation of ground-to-air drone visual detection based on deep learning

  • 期刊论文

  • 发表日期：

    2025年9月5日

查看全文
  • Jun Li,
  • Ban Wang,
  • Qitian Cui,
  • Feng Tian,
  • Maoying Zhou,
  • Rougang Zhou

摘要

In this study, we investigate the visual detection challenge of ground-to-air drones. With the widespread use of drones in civilian domains, the demand for counter-drone technologies has increased significantly. Detection serves as a foundation for countermeasures. In the ground-to-air drone visual detection domain, we address two primary challenges: the absence of a comprehensive open-source dataset for drone detection and the lack of multifactor analysis based on such a dataset. To address these challenges, we propose GA-Fly, a new dataset containing 10,800 high-resolution images capturing a DJI Mini 4 Pro drone at various angles, distances, and lighting conditions. The proposed dataset facilitates research on small target drone detection. We assessed the performance of eight representative deep learning algorithms, namely, Cascade R-convolutional neural networks (CNN), Faster R-CNN, Faster R-CNN-FPN, RTMDet, YOLOv8, YOLOv10, SSD, and YOLOv3, using the GA-Fly dataset. Our experiments included evaluating the effects of grayscale images, shooting angles, target scales, and cropping on detection accuracy. Our findings reveal that Faster R-CNN-FPN (0.894) achieves the best mean average precision, which is closely followed by YOLOv3 (0.823), whereas RTMDet (0.370) and YOLOv8 (0.458) achieve the worst and second worst performance, respectively. This study highlights the necessity for specialized algorithms tailored for drone detection given the distinct challenges posed by small targets and diverse environmental conditions. The GA-Fly dataset and the insights gained from these experiments can serve as a benchmark for future research in this area, contributing to the advancement of more effective drone detection systems. The proposed dataset is available at https:/github.com/ballballubsmart/GA-FLY.