宇航学报

宇航学报 ›› 2021, Vol. 42 ›› Issue (10): 1283-1292.doi: 10.3873/j.issn.1000-1328.2021.10.009

• 制导、导航、控制与电子 • 上一篇    下一篇

地球同步轨道目标物深度学习检测方法

黄西尧，何羿霆，杜华军，曾祥远，刘天赐，单文婧，程林   

  1. 1. 北京理工大学自动化学院，北京 100081；2. 北京航天自动控制研究所宇航智能控制技术国家级重点实验室，北京 100854；3. 北京航空航天大学宇航学院，北京 102206

  • 收稿日期:2020-10-09 修回日期:2021-03-25 出版日期:2021-10-15 发布日期:2021-10-15
  • 基金资助:
    国家重点研发计划（2017YFB1300101）；国家自然科学基金（11972075）

Geostationary Orbit Object Detection Based on Deep Learning

HUANG Xi yao, HE Yi ting, DU Hua jun, ZENG Xiang yuan, LIU Tian ci, SHAN Wen jing, CHENG Lin   

  1. 1. School of Automation, Beijing Institute of Technology, Beijing 100081, China;2. National Key Laboratory of Science and Technology on Aerospace Intelligent Control,Beijing Aerospace Automatic Control Institute, Beijing 100854, China;3. School of Astronautics, Beihang University, Beijing 102206, China
  • Received:2020-10-09 Revised:2021-03-25 Online:2021-10-15 Published:2021-10-15

PDF

213

摘要： 针对欧空局SpotGEO竞赛中地球同步轨道目标物的检测问题，提出面向低精度CCD空间图像的深度学习检测方法。在图像预处理环节，分别采用高斯过程回归和模板匹配实现前景/背景分割和多帧图像配准。根据地球同步轨道物体的运动特征，采用拓扑扫描提取候选目标物。在此基础上，提出一套基于深度学习的目标物筛选方法。该方法利用卷积神经网络，依次对拓扑扫描前后候选目标物进行筛选，显著减少噪声点数量，提高检测效率。仿真结果表明，该方法达到98%的目标检测准确率，适用于存在光污染、云层遮挡等干扰的复杂环境。


关键词: 空间目标检测, 地球同步轨道, 深度学习, 拓扑扫描

Abstract: A deep learning based method is proposed to detect GEO objects from the low precision CCD images for the ESA “SpotGEO” competition. The Gaussian process regression and template matching method are adopted in the image data preprocessing step. According to the motion characteristics of GEO objects, the topological sweeping method is used as a preliminary step. To reduce the noise effect, an object filtering method is proposed. Two additional data filters are set before and after the topological sweeping respectively using the convolutional neural network. They significantly decrease the number of noise points and increase the detection accuracy. Results show that this method can reach a high detection accuracy of 98%, which is suitable for the sophisticated environment with light pollution and clouds covering.


Key words: Space object detection, Geostationary orbit, Deep learning, Topological sweeping

中图分类号: