With the growing demand for unmanned aerial vehicle (UAV) localization systems that remain robust under challenging weather and lighting conditions, mid-wave infrared (MWIR) imagery has emerged as a promising alternative to conventional visible light (VL) imagery. However, existing VL-based localization methods are not directly applicable to MWIR images due to their fundamentally different visual characteristics. In this paper, we propose a novel UAV localization framework specifically designed for MWIR imagery. Our approach begins by detecting line features from MWIR frames, which are then matched to a predefined 3D line map of geologically meaningful structures. To establish robust correspondences under significant viewpoint changes, we introduce a 2D-3D line matching model trained using a synthetic dataset generated via a novel data augmentation strategy. The UAV’s pose is subsequently refined by minimizing the reprojection error between matched 2D and 3D lines. We validate our method using simulated MWIR flight sequences rendered from 3D model of the Sinjin Island. Unlike state-of-the-art VL-based baselines, which suffer from degraded performance in the MWIR domain, our method achieves more accurate and robust localization. In addition, the proposed framework runs in real-time, making it suitable for practical deployment in onboard UAV systems.
Huh et al. (Thu,) studied this question.