Automated 3D Reconstruction and Pose Estimation of Space Objects Using Ground Based Telescope Imagery

Stellar Science, in collaboration with the University of Maryland, proposes to develop algorithms to demonstrate the feasibility of constructing three-dimensional (3D) models of satellites using ground-based imagery on a standard PC in under 15 minutes. The Phase I effort will address the challenge of registering two-dimensional (2D) images of satellites where traditional feature matchers frequently fail. We will conduct novel research into image alignment using both feature-based and featureless methods to support 3D pose estimation of a known satellite, laying the groundwork for Structure from Motion (SFM) and dense model reconstruction in Phase II. Our improved algorithms will draw on Stellar Sciences experience in computing 3D models from imagery and video acquired under less-than-ideal conditions, our experience in Space Surveillance and Space Situational Awareness (SSA), and the University of Marylands experience in deep learning methods for feature learning and pose estimation. The integrated C3PO-R2D2 (Computing 3D Pose and Object Reconstruction from 2D Data) system will allow dense 3D models to be built from a series of images in order to evaluate the identity, pose, and configuration of an unknown satellite in less than 15 minutes.