Satellite Pose Estimation

Abstract

6D Satellite pose estimation is a crucial step of many on orbit servicing missions. Either for docking or debris removal, one need to know the pose of the target satellite. Throughout the time many methods were proposed to solve this problem, till the recent surge of monochrome based deep learning methods. We particularly get interested in data demanding direct regression methods using neural networks. We generate photorealistic data using Blenderproc and we divide many experiments to probe some architectural designs improving the networks performance, such as the choice of representations, hyperparameters and loss functions. Using all the feedback from previous experiments we go on optimizing the results.