ESA GNC Conference Papers Repository

Pose Estimation of a Non-Cooperative Target Based on Silhouette Imagery Using Convolutional Neural Networks
Anne Bettens, Anthea Comellini, Emmanuel Zenou, Vincent Dubanchet
Presented at:
Virtual Conference 2021
Full paper:

Autonomous pose estimation of spacecraft is the ability to calculate the state of a target satellite in order to navigate in close proximity for on-board servicing or debris removal. The research presented proposes using a Convolution Neural Network (CNN) to determine a spacecraft?s pose estimate from silhouette images. Using computer vision techniques, a silhouette of an object of interest or non-cooperative satellite can be obtained from multi-spectral images; these images are robust against illumination conditions. Extracting a silhouette of the target satellite for use in a CNN provides a novel approach to pose estimation. Using a hybrid classification-regression based approach to a CNN framework allows for numerical attribution of estimates, where the network is capable of providing an output based on a 2D image. Furthermore, image pre-processing led to the generation of an original data set large enough for training a CNN to perform accurate machine learning. The method presented demonstrates that a CNN is capable of extracting a pose estimate of a spacecraft from silhouette imagery.