Satellite Relative Pose Estimation via Transfer Learning and Environment-adaptive Conditional Features

초록

This paper presents a transfer-learning framework for 6D satellite relative pose estimation that augments a FiLM-based conditional feature extraction and environment adaptation module. Our adaptation module constructs a 12-dimensional condition vector—combining background brightness, V-channel histogram, and RGB-channel mean absolute deviation (MAD)—and embeds it via a compact MLP to generate FiLM parameters that dynamically modulate intermediate CNN features, compensating for background and contrast–induced distortions to improve pose accuracy. Validation on unseen satellite imagery confirms robustness to severe illumination changes and supports critical operations such as rendezvous, docking, and capture of non-cooperative debris. © ICROS 2025.

키워드