Close-range relative navigation with EPOS 6D pose estimator

Abstract

The EPOS6D method is used in a vision-based relative navigation setting to estimate the pose of an uncooperative target satellite in monocular camera images. EPOS6D is a pose estimation system composed of multiple sequentially acting components. The first stage is a neural network performing simultaneously object segmentation, part segmentation, and keypoint regression. The part segmentations and keypoints belong to specific fixed areas on the surface of the target object and allow establishing correspondences between 2D keypoints in the image and 3D keypoints on target object models, which is done via a Progressive-X implementation of PnP-RANSAC in the EPOS6D system. The EPOS6D method is applicable for target objects exhibiting symmetry or some other form of visual ambiguity that causes multiple pose hypotheses to be plausible. Previously, the EPOS6D method has been demonstrated on the T-LESS, YCB-V, and LM-O datasets from the general computer vision field. Using the method for pose estimation in space brings with it new challenges like large distance and lighting variations. Therefore, EPOS6D is tested on new datasets that reflect such spaceborne conditions, featuring synthetic and real images of a spacecraft exhibiting symmetry. The first focus of the experiments is on determining parameters and pose estimation system design aspects that allow training the network on synthetic images while actually developing it for use on the real camera images, or in other words, overcoming the "sim2real" problem. The second focus is on determining the accuracy and reliability potential of the system given the difficult visual conditions featured in the datasets. Finally, an attempt is made to try and initialize a visual tracker with the EPOS6D system in an attempt to improve the refresh rate of the pose estimations.