Neural Network Approach to Crater Detection for Lunar Terrain Relative Navigation

Abstract

Terrain relative navigation can improve the precision of a spacecraft’s state estimate by providing supplementary measurements to correct for drift in an inertial measurement unit. This paper presents a crater detector, LunaNet, that uses a convolutional neural network (CNN) and image processing methods to detect craters from camera imagery taken by a spacecraft’s onboard camera. These detections are matched with known lunar craters, and these matches are used as visual landmark measurements in an extended Kalman filter (EKF). Our results show that, on average, LunaNet detects approximately twice the number of craters in an intensity image as two prior intensity-image-based crater detectors, and detects more accurate craters than the other two detectors as well. One of the challenges of using cameras for this task is that they can generate imagery with significant variations in image quality and noise levels. LunaNet is robust to four common types of image noise due to its incorporation of a CNN that is trained on diverse data. LunaNet also produces crater detections with better image persistence over a trajectory. These qualities contribute to a LunaNet-based EKF that results in consistently lower state estimation error and that outperforms the filters based on the other detectors.