中国空间科学技术 ›› 2022, Vol. 42 ›› Issue (3): 74-81.doi: 10.16708/j.cnki.1000-758X.2022.0038

• 论文 • 上一篇    下一篇

基于三维点云重建的助推器位姿估计

肖爱群,姜鸿翔   

  1. 1北京宇航系统工程研究所,北京100076
    2北京航空航天大学 宇航学院,北京100191
  • 出版日期:2022-06-25 发布日期:2022-06-22

Booster pose estimation based on 3D point cloud reconstruction

XIAO Aiqun,JIANG Hongxiang   

  1. 1Beijing Institute of Astronautical Systems Engineering, Beijing 100076, China
    2School of Astronautics, Beihang University, Beijing 100191, China
  • Published:2022-06-25 Online:2022-06-22

摘要: 助推器分离是运载火箭发射过程中的关键动作之一,常用的激光雷达姿态测量技术在助推器分离阶段受外界干扰严重,难以准确获得位姿。基于视觉的助推器位姿变化测量技术具有优秀的抗干扰能力,通过搭建三维点云重建网络,以图像为输入,三维点云为输出,在构建的助推器分离过程的图像点云数据上进行了训练和测试,对测试重建的助推器点云使用主成分分析的方法完成了位姿的估算。测试结果表明,所建立的三维点云重建网络可以根据仿真图像数据,精确测量助推器分离阶段的位姿变化,在R2score指标下,对三维坐标的预测分数均在0.98以上,姿态角平均误差约为21°,预测分数则均在0.80以上。

关键词: 神经网络, 三维点云重建, 生成式模型, 主成分分析, 位姿估计

Abstract:  Booster separation is one of the important actions in the launching process of the carrier rocket. The commonly used LiDAR pose measurement technology is severely affected by external factors during the separation stage for the booster,so it is difficult to accurately obtain the pose of booster. To improve the anti-interference ability of pose estimation,the vision-based pose measurement technology was utilized for booster. A 3D point cloud reconstruction network whose input was the image and output was corresponding 3D point cloud was built and trained on the imagepoint cloud dataset,which was constructed during the separation of booster. During testing,the pose estimation was completed via principal component analysis on the reconstructed booster point cloud. All the experimental results illustrate that pose changes can be measured precisely by the built network according to the simulation image data during the separation stage for booster. Under the R2score metric,the prediction scores for the three-dimensional coordinates are all above 0.98.For the attitude angle,the average error is about 21°,and the prediction scores are all above 0.80.

Key words: neural networks, 3D point cloud reconstruction, generative model, principal component analysis, pose estimation