Cross-domain adaptive Siamese framework with benchmark for UAV self-positioning in low-light environments

Vision-based autonomous positioning has emerged as a reliable alternative for Unmanned Aerial Vehicles (UAVs) positioning in GNSS-denied environments. Compared to other methods, vision-based positioning offers distinct advantages including strong autonomy, low cost, and the unique capability to provide rich semantic understanding of the environment. However, existing studies, almost exclusively grounded in a single visible light modality, are incapable of satisfying the positioning needs of UAVs in low-light environments. To address these issues, a Cross-Domain Adaptive Siamese Framework with Benchmark are proposed. In the proposed framework, a feature extraction network known as CDFD-Net is initially investigated, with the objective of achieving the cross-domain separation of radiometric and structural characteristics. Second, an adaptive quality-weighted prediction head is proposed, which constructs a quality-aware adaptive weighted loss to better leverage the feature space information hidden within the samples, solving the problem of feature interference caused by repetitive textures. For the experiment, a novel dataset called LLE-UAV was constructed, which contains six categories of different scene attributes and over 14.9 × 10 3 pairs of strictly registered infrared and visible light images, and corresponding evaluation metrics MA@ K and RMAS were also introduced.