航空学报 > 2022, Vol. 43 Issue (5): 325298-325298   doi: 10.7527/S1000-6893.2021.25298

基于Transformer模型的卫星单目位姿估计方法

王梓1, 孙晓亮1, 李璋1, 程子龙2, 于起峰1   

  1. 1. 国防科技大学 空天科学学院, 长沙 410073;
    2. 中国航天员科研训练中心, 北京 100094
  • 收稿日期:2021-01-21 修回日期:2021-02-05 发布日期:2021-03-09
  • 通讯作者: 孙晓亮 E-mail:alexander_sxl@nudt.edu.cn
  • 基金资助:
    国家自然科学基金(62003357);湖南省研究生科研创新项目(CX20200024,CX20200025,CX20200088)

Transformer based monocular satellite pose estimation

WANG Zi1, SUN Xiaoliang1, LI Zhang1, CHENG Zilong2, YU Qifeng1   

  1. 1. College of Aerospace Science and Engineering, National University of Defense Technology, Changsha 410073, China;
    2. China Astronaut Research and Training Center, Beijing 100094, China
  • Received:2021-01-21 Revised:2021-02-05 Published:2021-03-09
  • Supported by:
    National Natural Science Foundation of China (62003357); Postgraduate Scientific Research Innovation Project of Hunan Province (CX20200024, CX20200025, CX20200088)

摘要: 鉴于测量精度高、设备成本低等优点,基于单目图像的卫星位姿估计方法在交会对接、空间攻防等应用中具有广泛前景。得益于强大的特征提取与表达能力,卷积神经网络在目标单目位姿估计应用中取得了明显优于传统方法的性能表现,但已有基于卷积神经网络的方法存在诸如归纳偏置、绝对距离描述不直接、长距离建模能力不足等问题。聚焦卫星单目位姿估计应用需求,针对以上问题,创新地将Transformer模型应用于卫星目标位姿估计中,提出了一种新颖的端到端卫星单目位姿估计方法。首先,设计了一种基于关键点集合的卫星目标表示方法,并构建了基于该表示方法的损失函数,进一步,结合关键点回归任务特点,设计了端到端的关键点回归网络模型,并改进了用于特征提取的主干网络结构。在公开数据集上实验测试结果表明:本文方法实现了可靠、高效的卫星目标单目位姿估计,并取得了优于已有同类方法的性能。

关键词: 卫星位姿估计, Transformer模型, 非合作目标, 注意力机制, 关键点回归

Abstract: With the advantages of measurement accuracy and low equipment cost, the satellite pose estimation method based on monocular image has a broad prospect in rendezvous and docking, space attack-defense and other applications. Due to the strong power of feature extraction and representation, the convolutional neural network has achieved significantly better performance than traditional methods in monocular pose estimation. However, the existing methods based on convolutional neural network have some problems, such as inductive bias, indirect description of absolute distance, and lack of long-distance modeling ability. Considering the application requirements of satellite monocular pose estimation, this paper applies the transformer model for satellite pose estimation innovatively to overcome the problems above, and proposes a novel end-to-end satellite monocular pose estimation method. A satellite target representation method is proposed based on the set of key points, and the loss function based on the representation method is established. Then, an end-to-end key point regression network model is developed based on characteristics of the key point regression task, and the backbone network structure for feature extraction is improved. Experimental results on public datasets show that the proposed method can achieve reliable and efficient monocular pose estimation of satellite targets, demonstrating better performance than existing similar methods.

Key words: satellite pose estimation, transformer model, no-cooperative target, attention mechanism, key points regression

中图分类号: