A Monocular Pose Estimation Case Study: The Hayabusa2 Minerva-II2 Deployment

Andrew Price, Kazuya Yoshida; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2021, pp. 1992-2001

In an environment of increasing orbital debris and remote operation, visual data acquisition methods are becoming a core competency of the next generation of spacecraft. However, deep space missions often generate limited data and noisy images, necessitating complex data analysis methods. Here, a state-of-the-art convolutional neural network (CNN) pose estimation pipeline is applied to the Hayabusa2 Minerva-II2 rover deployment; a challenging case with noisy images and a symmetric target. To enable training of this CNN, a custom dataset is created. The deployment velocity is estimated as 0.1908 m/s using a projective geometry approach and 0.1934 m/s using a CNN landmark detector approach, as compared to the official JAXA estimation of 0.1924 m/s (relative to the spacecraft). Additionally, the attitude estimation results from the real deployment images are shared and the associated tumble estimation is discussed.