A Deep Learning-based Crater Detector for Autonomous Vision-Based Spacecraft Navigation

The paper proposes the use of Cascade Mask R-CNN for the detection of craters from monocular images. Crater detection is a challenging task being the images prone to changes in lighting and noise conditions. Besides, the crater appearance is strongly modified according to the region of interest, being the shadows strongly affected by the sun vector inclination. To tackle these issues, the paper exploits the generalizability of modern deep learning architectures to create a highly reliable crater detector. The dataset used for transfer learning the model comprises more than 800 real lunar monocular images obtained from the lunar reconnaissance orbiter (LRO) cameras. Results confirm the performance reached by the multi-stage object detection architecture both in equatorial and polar regions, its robustness, and the validity of this crater detection scheme for planetary navigation tasks.