FINLOS-TriSeg: Fine-Grained Fisheye Semantic Segmentation for GNSS NLOS Identification in Consumer-Grade Measurement Systems

In challenging environments, such as urban canyons, Global Navigation Satellite System (GNSS) signals frequently undergo non-line-of-sight (NLOS) propagation due to blockage and reflection from buildings and vegetation, which severely degrades positioning accuracy, particularly in consumer-grade measurement systems. Upward-looking fisheye cameras provide panoramic environmental visibility cues and are therefore well suited for satellite-visibility inference. However, most existing vision-aided approaches adopt a binary paradigm (sky versus nonsky), which cannot effectively distinguish semitransparent vegetation from solid buildings, leading to high-NLOS false-alarm rates in complex scenes, especially under canopy cover. To address this limitation, we propose FINLOS-TriSeg, a fine-grained fisheye semantic segmentation framework that establishes a three-class perception paradigm for sky, foliage, and building understanding. To support this task, we build and open-source the Fisheye-SLBIK dataset with high-precision pixel-level annotations. Methodologically, we design a teacher-student collaborative learning framework for semi-supervised training, integrating explicit boundary modeling with cross pseudo supervision (CPS) to improve segmentation reliability near sky-object interfaces. Experimental results show that the proposed method achieves competitive segmentation performance while providing a favorable accuracy-efficiency tradeoff for lightweight deployment-oriented applications. Furthermore, by jointly analyzing the predicted semantic masks and satellite geometric trajectories, we show that FINLOS-TriSeg can more effectively identify subtle NLOS signals than conventional signal-domain indicators and improve real-time kinematic (RTK) positioning robustness in complex urban environments. This study provides high-quality environmental perception priors and reliable signal-quality labels for multisource fusion navigation in challenging urban scenarios.