LF-GNSS: Towards More Robust Satellite Positioning with a Hard Example Mining Enhanced Learning-Filtering Deep Fusion Framework
Jianan Lou, Rong Zhang
TL;DR
LF-GNSS addresses the challenge of robust satellite positioning in urban environments by fusing learning with filtering: a deep network analyzes per-satellite signals to adaptively set the EKF's observation noise and provide compensated innovations, while a dynamic hard example mining strategy strengthens training on difficult signals. It introduces a DOP-based feature (DPC) to quantify each satellite's impact on geometry, and uses multi-head attention to weigh satellite observations for improved measurement weighting. Experimental results on diverse public and private datasets show superior accuracy and stability compared with traditional and DL-based baselines, including tighter 3D RMSE and reduced outliers, and it outperforms IE-PPP in urban scenarios. The work includes comprehensive ablations confirming the value of DPC and HEM, and it publishes open-source code and urban GNSS datasets to accelerate AI-enabled GNSS research and deployment.
Abstract
Global Navigation Satellite System (GNSS) is essential for autonomous driving systems, unmanned vehicles, and various location-based technologies, as it provides the precise geospatial information necessary for navigation and situational awareness. However, its performance is often degraded by Non-Line-Of-Sight (NLOS) and multipath effects, especially in urban environments. Recently, Artificial Intelligence (AI) has been driving innovation across numerous industries, introducing novel solutions to mitigate the challenges in satellite positioning. This paper presents a learning-filtering deep fusion framework for satellite positioning, termed LF-GNSS. The framework utilizes deep learning networks to intelligently analyze the signal characteristics of satellite observations, enabling the adaptive construction of observation noise covariance matrices and compensated innovation vectors for Kalman filter input. A dynamic hard example mining technique is incorporated to enhance model robustness by prioritizing challenging satellite signals during training. Additionally, we introduce a novel feature representation based on Dilution of Precision (DOP) contributions, which helps to more effectively characterize the signal quality of individual satellites and improve measurement weighting. LF-GNSS has been validated on both public and private datasets, demonstrating superior positioning accuracy compared to traditional methods and other learning-based solutions. To encourage further integration of AI and GNSS research, we will open-source the code at https://github.com/GarlanLou/LF-GNSS, and release a collection of satellite positioning datasets for urban scenarios at https://github.com/GarlanLou/LF-GNSS-Dataset.