Scholarlink

Vision-based unmanned aerial vehicle (UAV) target detection has significant application value in many scenarios. Obtaining a suitable dataset to provide sample support for the training of deep learning algorithms is the key issue and a prerequisite for the algorithm to be implemented. This paper first proposes a method for generating complex multi-scene UAV image dataset in a simulation environment. AirSim is used in Unreal Engine to propose an automatic data collection and labeling method to collect multi-scene UAV position dataset, and we collect a dataset containing (a total of 59,330) images. Second, we use representative target detection algorithms such as YOLOv3 and CenterNet to train and test in the dataset. Finally, we compare the results of the experiments obtained and make a corresponding analysis.

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

Ground-to-Air Visual Detection of UAV: AirSim-Based Dataset Generation and Deep Learning Evaluation

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Jingming Yan,
  • Jiaqi Zhou,
  • Xiangyu Zhu,
  • Dongjie Zhou,
  • Zhoujingzi Qiu,
  • Yong Wang

摘要

Vision-based unmanned aerial vehicle (UAV) target detection has significant application value in many scenarios. Obtaining a suitable dataset to provide sample support for the training of deep learning algorithms is the key issue and a prerequisite for the algorithm to be implemented. This paper first proposes a method for generating complex multi-scene UAV image dataset in a simulation environment. AirSim is used in Unreal Engine to propose an automatic data collection and labeling method to collect multi-scene UAV position dataset, and we collect a dataset containing (a total of 59,330) images. Second, we use representative target detection algorithms such as YOLOv3 and CenterNet to train and test in the dataset. Finally, we compare the results of the experiments obtained and make a corresponding analysis.