Lights Out: A Nighttime UAV Localization Framework Using Thermal Imagery and Semantic 3D Maps

Researchers Ryan Allen and Melissa Greeff have introduced Lights Out, a novel framework for nighttime UAV localization that leverages thermal imagery and semantic 3D maps. Published on arXiv and accepted to ICUAS 2025, the system addresses the critical challenge of GNSS-denied navigation at night. Traditional methods fail due to the severe modality gap between daytime RGB maps and nighttime thermal images. Lights Out overcomes this by formulating localization in a shared semantic domain, using a symmetric bidirectional reprojection objective with confusion-aware weighting to handle segmentation uncertainty. The framework relies on a globally referenced, semantically labeled 3D map built from daytime RGB data, aligning segmented thermal observations without appearance-based correspondence.

Evaluated offline across 6.5 km of real-world nighttime UAV flights in urban and semi-structured environments, the system achieved impressive accuracy: a bias-corrected RMSE2D of 2.18 meters and a median RMSE2D of 1.52 meters, relative to RTK GNSS ground truth. The study found that localization performance strongly correlates with the availability of semantic edge evidence, and large-error events are spatially localized to semantically ambiguous areas rather than uniformly distributed. This work demonstrates that semantic reprojection offers a promising pathway for globally referenced nighttime UAV localization using only thermal imagery, with potential applications in search and rescue, surveillance, and autonomous drone operations in low-visibility conditions.