航空学报 > 2022, Vol. 43 Issue (1): 324854-324854   doi: 10.7527/S1000-6893.2020.24854

无人机视觉引导对接过程中的协同目标检测

王辉1,2, 贾自凯1,2, 金忍1,2, 林德福1,2, 范军芳3, 徐超1,4   

  1. 1. 北京理工大学 宇航学院, 北京 100081;
    2. 北京理工大学 无人机自主控制技术北京市重点实验室, 北京 100081;
    3. 北京信息科技大学 高动态导航技术北京市重点实验室, 北京 100085;
    4. 北京特种机电研究所, 北京 100012
  • 收稿日期:2020-10-12 修回日期:2020-12-10 出版日期:2022-01-15 发布日期:2020-12-03
  • 通讯作者: 金忍 E-mail:renjin@bit.edu.cn
  • 基金资助:
    国家自然科学基金(U1613225);北京信息科技大学高动态导航技术北京市重点实验室资助项目(HDN2020105);北京信息科技大学高动态导航技术北京市重点实验室开放课题基金(HDN2020101)

Cooperative object detection in UAV-based vision-guided docking

WANG Hui1,2, JIA Zikai1,2, JIN Ren1,2, LIN Defu1,2, FAN Junfang3, XU Chao1,4   

  1. 1. School of Aerospace Engineering, Beijing Institute of Technology, Beijing 100081, China;
    2. Beijing Key Laboratory of UAV Autonomous Control, Beijing Institute of Technology, Beijing 100081, China;
    3. Beijing Key Laboratory of High-Dynamic Navigation Technology, Beijing Information Science and Technology University, Beijing 100085, China;
    4. Beijing Institute of Special Mechanic-Electric, Beijing 100012, China
  • Received:2020-10-12 Revised:2020-12-10 Online:2022-01-15 Published:2020-12-03
  • Supported by:
    National Natural Science Foundation of China(U1613225); Funded Project of Beijing Key Laboratory of High Dynamic Navigation Technology of Beijing Information Science and Technology University(HDN2020105); Open Fund Project of Beijing Key Laboratory of High Dynamic Navigation Technology of Beijing Information Science and Technology University(HDN2020101)

摘要: 无人机空中自主回收是未来的发展趋势,对空中载具的自动识别是实现视觉引导回收的关键技术之一。目前对于空中关联目标检测的研究局限于单个目标个体,没有充分利用关联目标之间的信息。本文针对空中高动态对接中的关联目标检测问题,提出了一种母机与挂载对接物体的单阶段快速协同目标检测算法,包括相关类别并行独立分支目标检测、相关类别掩模增强检测以及相关类别的特征一致性约束,这些模块能够共同提升检测表现。实验表明,在测试集中该算法相比于YOLOv4能够提升4.3%的平均精度,相比YOLOv3-Tiny能够提升31.6%的平均精度。同时,该算法已应用在MBZIRC2020的高动态空中对接项目上,实现机载图像在线实时处理,团队借此斩获冠军。

关键词: 无人机自主回收, 视觉引导, 目标检测, 关联目标, 卷积神经网络

Abstract: Autonomous aerial recovery of UAV is a future development trend, and automatic detection of aerial vehicles is one of the key technologies to realize vision-guided recovery. At present, the research on the detection of aerial related objects is limited to individual objects, and the information between correlated objects is not fully utilized. For the problem of related object detection in high-dynamic aerial docking, this paper proposes a single-stage fast cooperative algorithm for detection of the master and the mount, including detection of sibling independent head of related category, detection of mask enhancement of related category, and constraints on consistency of features of related categories. These modules can improve the detection performance jointly. Experiments show that in the test dataset, the algorithm can obtain a 4.3% increase of the average precision of compared with YOLOv4, and can obtain a 31.6% increase of the average precision compared with YOLOv3-Tiny. At the same time, this algorithm has been applied to the high dynamic aerial docking project of MBZIRC2020 to achieve online real-time processing of airborne images, and our team won the championship.

Key words: UAV autonomous recovery, visual guided, object detection, related objects, convolutional neural networks

中图分类号: