Robust Perspective-n-Crater for Crater-based Camera Pose Estimation

Sofia Mcleod, Chee Kheng Chng, Tatsuharu Ono, Yuta Shimizu, Ryodo Hemmi, Lachlan Holden, Matthew Rodda, Feras Dayoub, Hirdy Miyamoto, Yukihiro Takahashi, Yasuko Kasai, Tat-Jun Chin; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 6760-6769

Lunar survey missions require accurate estimation of satellite and/or sensor pose (position and attitude) to achieve precise surface measurement. Crater-based pose estimation (CBPE) holds promise to achieve the desired accuracy. However current pose estimation methods suffer from one or more weaknesses such as loose coupling of position and attitude optimisation not accounting for wrong crater matches and using geometrically invalid objective functions for estimation. To conclusively address these drawbacks we develop a robust perspective-n-crater pose estimation method that employs geometrically meaningful and information-rich elliptical representation of craters in combination with M-estimators to account for incorrect crater identifications. To enable evaluation we construct an extensive labelled dataset of synthetic lunar images taken under diverse conditions over the Moon's surface. Results on the dataset demonstrate that our work addresses the drawbacks of previous methods and raises the achievable accuracy of CBPE. As another contribution we will also release our dataset to stimulate further research.