VGPNet: A Vision-aided GNSS Positioning Framework with Cross-Channel Feature Fusion for Urban Canyons

This paper proposes VGPNet, a vision-aided GNSS positioning framework for urban canyon environments. Instead of classifying and excluding non-line-of-sight (NLOS) satellite observations, VGPNet preserves all satellite observations and uses sky-pointing fisheye imagery together with GNSS measurements to estimate adaptive per-satellite weights and bias corrections.

The core component is a Cross-Channel Feature Fusion Module (CCFFM), which projects visual and GNSS features into a shared latent space and enables cross-modal interaction through a cross-channel self-attention mechanism. This design helps the model exploit both line-of-sight and NLOS observations while using environmental context from fisheye images.

Experiments on the public KLTDataset and a self-collected real-world urban dataset show that VGPNet improves both 2D and 3D positioning accuracy compared with existing GNSS positioning methods.

Recommended citation: H. Xue, H. Zhu, L. Wang, S.-C. Kuok, B. Li, H. Zhang, G. Li, and H. Leung. (2025). "VGPNet: A Vision-aided GNSS Positioning Framework with Cross-Channel Feature Fusion for Urban Canyons." IEEE Transactions on Instrumentation and Measurement, 74, 1-13. doi:10.1109/TIM.2025.3635311.
Download Paper