Photorealistic Image Generation for Satellite Pose Estimation Using Generative Adversarial Networks

Midshipman Researcher(s): 1/C Alec Engl

Adviser(s): Professor Tae Lim, Professor Randy Broussard, and Professor Gavin Taylor

Poster #6

Pose estimation, or the estimation of relative orientation and position of an object in space, is crucial for the autonomous rendezvous and servicing of a satellite. Convolutional Neural Network (CNN)-based methods show promise in satellite pose estimation, but are currently limited by the images used in their training. Few real images of satellites in orbit are available, so many turn to procedurally-generated imagery, despite the vast difference from the CNN’s perspective. This research gauges the ability of the Generative Adversarial Network (GAN) to refine synthetic imagery in order to better train CNN pose estimators.

Full Size Aerospace Engineering #1