IRoLSD: Illumination Robust line segment detection for spacecraft relative navigation

Line segments are used extensively in computer vision tasks, since they capture robust geometric features from structural information in the scene. In spacecraft relative navigation applications, images often contain prominent geometric structures, rendering line features valuable for downstream tasks such as localization and mapping. However, the performance of both learning-based and traditional line detection algorithms is significantly degraded under the harsh illumination conditions encountered in orbit. In this work, we introduce a novel line detection framework designed to robustly detect line segments in spacecraft imagery. Using images with perturbed lighting conditions and by incorporating depth maps to reconnect fragmented lines, we propose IRoLSD, an algorithm that reliably detects line segments across diverse lighting environments and realistic trajectories. We extensively evaluated the performance of IRoLSD in a series of experiments, proving that the proposed training method outperforms the state-of-the-art line detection algorithms in the especially challenging space domain. We also demonstrate the generalizability of IRoLSD to achieve superior performance across synthetic and real-world images, as well as across different spacecraft models. • We present IRoLSD, a self-supervised line segment detector trained for illumination-robustness in spacecraft navigation. • We develop a hybrid training scheme utilizing visual and depth information to ensure accurate and complete self-supervision. • IRoLSD achieves higher detection repeatability and accuracy compared to state-of-the-art line detectors under challenging lighting conditions. • We demonstrate consistently increased illumination-invariance through evaluation on both synthetic orbital imagery and hardware-in-the-loop experiments. • We present line segment tracking results proving more reliable in scenarios where traditional line segment detectors fail.