Parallel inversion of neural radiance fields for spacecraft trajectory estimation

Details

Supervisors

De Vleeschouwer, Christophe

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] en sciences informatiques, à finalité spécialisée

Abstract

This thesis focuses on spacecraft trajectory estimation. Knowing the 6D pose (position and orientation) of an unknown uncooperative space object relative to an on-board camera facilitates space missions such as debris removal, re- supplying and servicing. Making it possible for some of these missions to be executed without human intervention. Our method leverages recent researches around neural radiance fields and their inversion to perform 6D pose estimation of an uncooperative space object based on sequences of RGB images. The state of the art method for NeRF inversion is the parallel inversion technique. It focuses on estimating the pose of an object by minimizing the error between pixels rendered from a NeRF, for different starting poses, and the corresponding pixels of an observed RGB image iteratively. The main issue when considering this approach for space missions is the computational power required. This kind of computational power is not available in a spacecraft. In this work, we show that we can improve the efficiency of these kind of methods when dealing with moving objects. As most space missions imply a spacecraft navigating on orbit around a target object we can leverage this observation and adapt the parallel inversion of NeRFs to modify the target image during execution. We demonstrate on synthetic terrestrial images that the method converges fast to a subsequent image of the sequence when it is able to predict a pose relatively close to the target for the previous image of the trajectory. The particular conditions characterizing the space environment such as occlusion, symmetrical and almost texture less objects make it hard for most pose estimation techniques to achieve accurate results on spacecraft images. In this work we show how our method performs in these particular conditions both on synthetic and real images. We show that the method performance drops significantly when dealing with synthetic spacecraft images compared to the ideal case and even more when dealing with real images. Finally, we study the impact of the starting pose candidates on the perfor- mance of the method.