LUNA: Light-Based Urban Navigation and Anchor via Geometric Structural Constraints

Abstract

Robust visual localization for uncrewed aerial vehicles (UAVs) in urban environments is a critical capability, yet it remains challenging in low-light conditions where texture degradation compromises standard feature extraction. Existing solutions often rely on computationally intensive image enhancement or heavy multimodal sensors, which are unsuitable for resource-constrained platforms. Furthermore, traditional visual odometry methods struggle to maintain tracking consistency due to the low signal-to-noise ratio (SNR) and motion blur inherent in nighttime imagery. To address this, we propose light-based urban navigation and anchor (LUNA), a framework that exploits the high-contrast, repetitive window patterns of high-rise buildings as stable geometric anchors. Specifically, we introduce a hierarchical pipeline that extracts window primitives using multiscale maximally stable extremal regions (MSERs) and reconstructs the underlying facade lattice via a simulated annealing (SA)-based global consensus algorithm. These structural features are temporally tracked and integrated into a multistate constraint Kalman filter (MSCKF) to provide rigorous geometric constraints for attitude correction. Extensive real-world flight experiments demonstrate that LUNA achieves an absolute trajectory error (ATE) of 2.797 m, significantly outperforming methods, such as ORB (10.773 m), xFeats (8.197 m), and illumination-robust hybrid VIO (IRIH-VIO) (8.918 m). This work enables precise 6-DoF estimation in textureless environments by shifting reliance from unstable photometric textures to robust geometric structures.