Bennunet—An Update on Applying Deep Learning for Minor Bodies Optical Navigation

Abstract

This article presents Bennunet, a hybrid neural network-based method, devoted to on-board spacecraft relative position and attitude estimation in the vicinity of minor bodies using monocular vision. It is a follow-up investigation of Churinet, which set up the basis for using neural networks for pose estimation, offering a lightweight and robust alternative to the computationally expensive traditional methods, which may fail under adverse illumination conditions. In this case, the asteroid Bennu has been chosen as the target of the investigation given the extensive data derived from the OSIRIS-REx mission. Multiple shape models of Bennu have been used to produce synthetic image training sets covering the whole range of camera position, attitude, illumination conditions, camera field-of-view, image resolution, and target albedo map variation, allowing to study the impact of different geometries and image effects in the network performance and making it more robust. Moreover, modified state-of-the-art architectures have been implemented for Bennunet, substantially improving its performance compared to the baseline convolutional neural network (CNN) used in previous works. In addition, the implementation of a time distributed CNN, taking as input a sequence of images, has further improved the model accuracy. Multiple data augmentation techniques have been implemented to further extend the image sets during training. Finally, the trained networks have been validated with real images of Bennu. The obtained results show that the network is able to maintain the same accuracy achieved with synthetic images without any degradation.