On-board Small-Body Semantic Segmentation Based on Morphological Features with U-Net
Mattia Pugliatti, Michele Maestrini, Pierluigi Di Lizia, Francesco Topputo · Virtual Community of Pathological Anatomy (University of Castilla La Mancha) · 2021
Small-bodies such as asteroids and comets exhibit great variability in surface morphological features.These are often unknown beforehand but can be exploited for hazard avoidance during landing, autonomous planning of scientific observations, and for navigation purposes.The detection and classification of such features is a laborious task that requires extensive manual work done by experts in the field.This step renders online usage of images unfeasible for these applications.Such limitation could be overcome thanks to the recent advances in the field of neural networks, which allow to recognize features automatically from an acquired image.However, to train such networks, an annotated dataset needs to be generated with care by field experts, thus requiring once again extensive work and human-in-the-loop.In this work, a methodology that exploits an open-source rendering software, ray-tracing masking, and simple image processing techniques is illustrated, which allows to automatize the segmentation process and build up a robust database of labeled features (i.e.background, surface, craters, boulders, and the terminator region) for small-bodies.A procedural code is designed to generate images and their labels over 7 different small-body shapes for a total of 12, 550 images that are used to train a Convolutional Neural Network with a U-Net architecture in the task of semantic segmentation.The performances of the network are then analyzed in 4 different scenarios.First, the network is evaluated on a test set composed of 1, 050 new images belonging to bodies seen during training.Secondly, the network is evaluated on 3, 000 synthetic images from 2 models that have not been encountered in training.Afterward, one of these latter models is tested in a flyby trajectory scenario consisting of 56 images.The results of the first three tests show state of the art performances and the capability of this method to generalize features across synthetic data.Finally, the network's performances are qualitatively assessed with a set of 59 real images from previously flown missions, highlighting the current limits of this approach.These shortcomings suggest possible directions for future improvement, which are discussed in this work.