航空学报 > 2021, Vol. 42 Issue (11): 525045-525045   doi: 10.7527/S1000-6893.2021.25045

基于深度图推理的卫星背板部件检测方法

陈奥1, 解永春1,2, 王勇1,2, 李林峰1   

  1. 1. 北京控制工程研究所, 北京 100190;
    2. 空间智能控制技术重点实验室, 北京 100190
  • 收稿日期:2020-12-02 修回日期:2021-03-14 发布日期:2021-05-20
  • 通讯作者: 陈奥 E-mail:chenao526@foxmail.com
  • 基金资助:
    国家自然科学基金(U20B2054)

Deep graph reasoning method for satellite component detection

CHEN Ao1, XIE Yongchun1,2, WANG Yong1,2, LI Linfeng1   

  1. 1. Beijing Institute of Control Engineering, Beijing 100190, China;
    2. Science and Technology on Space Intelligent Control Laboratory, Beijing 100190, China
  • Received:2020-12-02 Revised:2021-03-14 Published:2021-05-20
  • Supported by:
    National Natural Science Foundation of China (U20B2054)

摘要: 在轨加注是一种典型的在轨服务操作,它对于降低空间运输成本和任务风险起着重要的作用,视觉感知系统可以感知操作任务周围环境并提供给控制系统。目前在轨加注依赖于人,在人员监控下完成或通过遥操作完成,缺乏自主性。本文围绕未来高自主性的基于深度强化学习的在轨加注方法,对基于深度学习的视觉感知方法展开了研究,针对基于深度学习的方法对相似实例的检测存在精确率低、对光照变化敏感等缺点,提出了基于深度图推理的卫星背板部件检测方法。提出的方法可以有效地检测复杂形状的目标,不依赖于手工设计的特征;提高了复杂光照环境下部件的检测正确率;可以有效区分外形相似的不同部件;其有效性在数学仿真和物理仿真中均得到了验证。

关键词: 在轨加注, 视觉感知, 复杂光照, 相似实例, 深度学习, 知识图谱

Abstract: On-orbit refueling is a typical on-orbit service operation, and plays an important role in reducing space transportation costs and mission risks. The visual perception system can perceive the surrounding environment of the operation task, and provide the information to the control system. Nowadays, on-orbit refueling relies on human labor, and is done under the supervision of personnel or by remote operation, lacking autonom. Centered around the future high autonomy on-orbit refueling method based on deep reinforcement learning, this paper conducts research on deep learning based visual perception. Because of the problems of low accuracy of detection of similar instances and being sensitive to changes in illumination, deep learning based methods cannot be directly used for on-orbit refueling. A deep graph reasoning method is proposed for satellite component detection in this paper. The proposed method can effectively detect complex-shaped targets without relying on manually-designed features. It can improve the detection accuracy of parts under complex illumination, and effectively distinguish different parts with similar appearances. Its effectiveness has been verified by mathematical simulation and physical simulation.

Key words: on-orbit refueling, visual perception, complex illumination, similar instance, deep learning, knowledge graph

中图分类号: