Pose estimation of chaotic motion of Didymos' moon using CNN-based Image Processing algorithm

Abstract

The Asteroid Impact and Deflection Assessment (AIDA) is an international collaboration between the European Space Agency (ESA) and the National Aeronautics and Space Administration (NASA) aiming to investigate the binary asteroid system‎(65803) Didymos and to demonstrate asteroid deflection technique with kinetic impact. NASA’s Double Asteroid Redirection Test (DART) mission successfully impacted Dimorphos, the moon of the binary system, in September 2022. ESA’s contribution is the Hera mission that will rendezvous with Didymos and observe the impact effects closely. The Close Observation Phase (COP) is the proximity operation of Hera mission with the objective of obtaining high-resolution images of Dimorphos and fully characterizing the impact crater. Autonomous optical navigation system is designed for this phase based on line-of-sight and range measurements from both the primary body and Dimorphos in order to estimate the relative position of the spacecraft. The close distance between the primary and the spacecraft during the COP allows the implementation of feature tracking relative navigation to solve the primary’s relative attitude. Nevertheless, the relative attitude of Dimorphos remains unsolved as it requires closer distance. This paper develops an innovative methodology to estimate the continuous six degree of freedom pose (position and attitude) of Dimorphos during the COP using a Convolutional Neural Networks (CNN)-based Image Processing (IP) algorithm. For the attitude, we implement an appearance-based method that consists of two stages. In the first stage, we use CNNs with the images captured by the spacecraft on-board camera to regress a set of keypoints segmenting Dimorphos from its background. In the second stage, we use Neural Networks (NN) to map these keypoints to the three Euler angles representing the relative rotation matrix of Dimorphos with respect to the spacecraft. The estimated keypoints are also used to estimate the position of the centroid of Dimorphos and its relative distance with respect to the spacecraft, which together provides the relative position vector of the spacecraft. For the distance, the shape of Dimorphos is approximated to an ellipse of size and shape depending on its relative attitude with respect to the spacecraft. The regressed keypoints are used to evaluate the apparent semi-minor and semi-major axes of Dimorphos, which are used to estimate the range from the spacecraft using the pinhole camera model. The High-Resolution Network (HRNet) is used as CNN architecture as it represents the state-of-the-art technology in keypoint detection with its capability of maintaining high resolution representations of the input images by connecting multiple subnetworks in parallel. For attitude navigation, the appearance-based method is selected for three main reasons. Firstly, it has the main advantage of a reduced dependency on the distance that is the main driver for a feature tracking relative navigation technique. Secondly, other methods such as model-based ones require sufficiently regular shapes of the targets, which is not the case of near-earth objects such as asteroids. Thirdly, the appearance-based method does not depend on prior knowledge of spinning axis and rate of the target, which appears to be chaotic for Dimorphos after the impact with DART. The training, validation and testing datasets consist of synthetic images generated with the software Planet and Asteroid Natural scene Generation Utility (PANGU) at different epochs of the COP trajectory provided by ESA. Additional images generated from different trajectory segments around the binary system are used to augment the training database. Therefore, our develop algorithm is expected to solve the overall pose estimation and improve the efficiency and the robustness of the autonomous navigation of the proximity operations of Hera mission.