Deep Learning-Based Relative Navigation About Uncooperative Space Objects

The space sector is nowadays experiencing an increase in the demand for missions involving proximity operations. This trend can be partially attributed to the raised awareness of the space debris problem, as an increasingly higher number of Active Debris Removal (ADR) and on-orbit servicing missions are being planned. Consequently, the need for an accurate onboard relative navigation system has become growingly relevant in the industry. This work proposes a software pipeline for spacecraft relative navigation, which leverages deep learning techniques to obtain the relative pose measurements and uses Kalman filtering to reconstruct the relative dynamics and to improve the robustness of the pipeline. Furthermore, a testing procedure involving a Blender-based spaceborne image generator has been devised and applied to validate the results in the case of a realistic image sequence of a rendezvous scenario. The image processing pipeline is based on a Convolutional Neural Network architecture that scored excellent results in a pose estimation challenge organized by ESA. This architecture has demonstrated centimeter-level position accuracy and degree-level attitude accuracy, along with considerable robustness to changes in background and lighting conditions. In order to reconstruct the relative state, a set of Kalman filters has been developed to tackle the attitude and position problems separately. For the relative distance, an Extended Kalman Filter has been implemented, as the underlying relative dynamics can be described by a linearized model. Instead, for the more complex attitude problem, the choice fell on an Unscented Kalman Filter due to its superior robustness to high non-linearities. In addition, proving the robustness of the filtering algorithms was taken as a priority, with thousands of tests aimed at identifying and counteracting the most common failure modes. Moreover, some techniques were also developed for the detection and rejection of measurement outliers. The whole navigation pipeline was then tested on a set of synthetic image sequences of the TANGO spacecraft in free tumbling conditions. The frames were obtained from a Blender-based spaceborne image generation platform, exploiting a 3D model of the target and relying on an accurate propagation of the relative dynamics. The proposed filtering pipeline proved to substantially improve the accuracy of the raw deep learning-based measurements by leveraging sequential information, while also increasing the overall robustness.