Camera-Pose Robust Crater Detection from Chang'e 5

TL;DR

This study tackles the problem of robust crater-detection for vision-assisted spacecraft pose estimation under off-nadir imagery. It evaluates a Mask R-CNN-based crater-detection and ellipse-regression pipeline, comparing pretraining on real lunar imagery versus simulated data, and introduces the CE5-CDA dataset to study off-nadir effects from the Chang'e 5 landing camera. The key finding is that pretraining on real nadir lunar mosaics, followed by finetuning on CE5-CDA, yields the best performance ($F_1 ext{ up to } $63.1$, $E_{ ext{IoU}} ext{ up to } 0.701$), while simulated data pretraining does not bridge the real-to-off-nadir domain gap. The work also provides the first annotated off-nadir crater dataset for Chang'e 5 and highlights the necessity of real-lunar data for robust CDA in off-nadir scenarios, with implications for dataset collection and domain-adaptation research.

Abstract

As space missions aim to explore increasingly hazardous terrain, accurate and timely position estimates are required to ensure safe navigation. Vision-based navigation achieves this goal through correlating impact craters visible through onboard imagery with a known database to estimate a craft's pose. However, existing literature has not sufficiently evaluated crater-detection algorithm (CDA) performance from imagery containing off-nadir view angles. In this work, we evaluate the performance of Mask R-CNN for crater detection, comparing models pretrained on simulated data containing off-nadir view angles and to pretraining on real-lunar images. We demonstrate pretraining on real-lunar images is superior despite the lack of images containing off-nadir view angles, achieving detection performance of 63.1 F1-score and ellipse-regression performance of 0.701 intersection over union. This work provides the first quantitative analysis of performance of CDAs on images containing off-nadir view angles. Towards the development of increasingly robust CDAs, we additionally provide the first annotated CDA dataset with off-nadir view angles from the Chang'e 5 Landing Camera.

Figures (3)

• Figure 1: Steps of pipeline, demonstrating bounding box detection and ellipse regression on image from the Chang'e 5 landing camera CE5.
• Figure 2: Sample images from CE5-CDA, with -regressed ellipses. Rows show two different samples from the CE5-CDA test set, while columns separate inference from each . Blue ellipses denote ground truth, while green and red ellipses show true positive and false positive predictions respectively.
• Figure 3: Reduction in detection and ellipse-regression performance over frame number of finetuned . The end of the training-set and beginning of the test-set is denoted with a grey line.