On-Orbit Relative Pose Initialization via Convolutional Neural Networks

Autonomous spacecraft rendezvous requires the relative pose (position and attitude) of the target spacecraft to be acquired and tracked for safe maneuvering. Model-based pose estimation methods, while shown to be highly accurate, typically require an initial position and attitude estimate to facilitate the subsequent model fitting process. Poor initial estimates can delay or hinder pose acquisition. Convolutional neural networks (CNNs) provide an alternative attitude initialization method by training them on labeled images and then classifying test images to attitude labels, resulting in a potentially faster and more robust attitude initialization process. This research seeks to answer several questions on the use of CNNs for spacecraft attitude initialization, including the necessary amount of training imagery and attitude label discretization as well as how on-orbit conditions such as lighting and image background affect CNN performance. A CNN framework that employs AlexNet-based transfer learning, systematic attitude label discretization, and a post-classification attitude refinement algorithm is tested on synthetic training and validation imagery produced in various on-orbit environments. Results indicate that CNNs can be designed and trained to reliably produce initial attitude estimates with the desired accuracy to expedite subsequent model-based pose estimation. This work also serves as a basis for future experiments that will employ real images collected on-orbit for testing.