Monocular Visual Localization via Conic Geometric Cues for Astrobee Free-Flyers

Minji Kim, Suyoung Kang, Seonwook Yeom, Ryan Soussan, Brian Coltin, Pyojin Kim

Keywords: Localization, Space Robotics and Automation, Vision-Based Navigation, Computational Geometry

TL;DR: Monocular visual localization for ISS robots using circular landmark geometry and Manhattan world regularities, robust to illumination changes without pre-built maps.

Abstract:

We present a 6-DoF monocular visual localization method for free-flying robots inside the ISS, leveraging time-invariant conic geometric cues and Manhattan world (MW) regularities in structured scenes for globally aligned poses.
Existing visual localization methods for free-flying robots rely on handcrafted or learned keypoints matched against pre-built maps of the ISS, making them vulnerable to illumination changes, occlusions, and frequent environmental reconfigurations.
To address these limitations, we exploit time-invariant conic primitives from circular landmarks and MW regularities in human-made environments.
A Perspective-Two-Circle (P2C) solver recovers an initial pose from two detected ellipses using only their known radii.
The resulting conic normals are aligned with MW lines for drift-free orientation.
All poses are jointly optimized through conic bundle adjustment (CBA) with algebraic conic residuals.
Experiments on Astrobee datasets show that our method outperforms state-of-the-art Astrobee localizers.
It further achieves accuracy comparable to feature-map-based approaches under severe illumination changes.

Submission Number: 31