Previous Abstract Return to Session B3 Next Abstract

Session B3: AI and Machine Learning for GNSS Signal Processing, Error Modeling, and Threat Detection

Robust AI-Based Smartphone Positioning via Physics-Informed Feature Selection
Jingzhou Liu, Yang Gao, Department of Geomatics Engineering, University of Calgary
Location: Windsong 7-8
Date/Time: Thursday, Sep. 17, 9:43 a.m.

Great efforts have been made in the past few years to improve positioning accuracy using Global Navigation Satellite Systems (GNSS) measurements from smartphones. There are many potential high-performance location-based services (LBS) such as lane-level vehicle navigation and vehicle maneuver analysis. How to obtain reliable smartphone position solutions, however, remains a significant challenge, particularly in challenging environments due to large and frequent outliers in GNSS measurements from smartphones caused by significant multipath (MP) and non-line-of-sight signals (NLOS).
Currently three approaches are widely used to detect outlier and ensure the reliability of position solutions: robust estimation (Huber, 1964; Yang et al., 2002), solution separation test (SS-test) (Blanch et al., 2022) and random sample consensus (RANSAC) (Gao et al., 2023). A major deficiency, however, exists as they are all model-based methods. The hyperparameters in robust estimation, for example, are designed according to statistical assumptions or engineering heuristics, which therefore require real-time adjustment and tuning when applied to real-world applications. The robust estimation involves residual-based outlier detection but the estimation can be severely biased when there are frequent and large outliers in measurements, which will degrade the efficiency of the method (Martin & Xia, 2022; Wen et al., 2023). The RANSAC method considers only the self-consistency between measurements while the performance using the SS-test is low when there are multiple outliers (Liu et al., 2024).
In recent years, the use of machine-learning methods has been investigated for its strong data mining capabilities via training to discover complex, underlying data patterns, potentially enabling more accurate estimations in challenging environments. Hsu (2017) employed GNSS raw data, including received signal strength, change rate of received signal strength, and pseudorange residuals, to perform multipath discrimination using a Support Vector Machine (SVM). Jiang et al., (2025) improved SS-test using the long short-term memory (LSTM) method to ensure measurement subset free of outliers. Although it improves outlier discrimination efficiency, the selected observations may still contain outliers, resulting in significant position errors particularly in challenging environments. Considering this issue, Kanhere et al, (2022) used pseudorange residuals and light-of-sight (LOS) observations to conduct an end-to-end position estimation based on smartphones through Transformer. The method, however, is dependent on the quality of the initial position which limits its performance due to three fundamental factors. First, the local LOS vectors do not represent the global geometric topology, leaving the network blind to the spatial geometric strength and the ill-conditioned nature of the inverse problem. Second, without observational context, the network cannot decouple systematic geometric deviations from non-Gaussian environmental noise (e.g., NLOS) within the lumped pseudorange residuals. Third, relying solely on highly sensitive residuals amplifies the non-linear linearization truncation errors when the initial state deviation is large.
In this research, we propose a robust physics-informed end-to-end method to derive accurate smartphone position solutions with enhanced features and tolerance to initial position errors. To remove the limitations of existing end-to-end models, we have designed a comprehensive set of input features to incorporate elevation, azimuth, signal-to-noise ratio (SNR), the global covariance matrix (Cx), SNR-derived prior weights, and pseudorange residuals. By injecting the global covariance matrix, the network is built with explicit awareness of the global geometric topology, enabling adaptive scaling of position corrections based on structural reliability. Furthermore, the introduction of SNR and elevation as contextual anchors, the network can effectively isolate multipath noise from geometric errors. Crucially, features such as SNR, elevation, and azimuth remain virtually invariant to initial position deviations. By serving as stable contextual anchors, they enable the network to reliably evaluate satellite signal quality and decouple multipath noise from systematic geometric errors, even when the input residuals are severely corrupted by a degraded initial position.
To evaluate the effectiveness of the proposed method, comprehensive experiments are being conducted using the Google Smartphone Decimeter Challenge (GSDC) dataset, specifically targeting challenging urban canyon environments. The initial test results indicate that, driven by the robust signal-noise decoupling capability of the newly designed features, our framework significantly improves 3D positioning accuracy and outlier discrimination over conventional model-based methods. More importantly, compared to existing residual-LOS-based end-to-end solutions, our preliminary conclusion is that the proposed method can provide substantially higher robustness, maintaining stable meter-level positioning accuracy even with poor initial position. More evaluation results on precise metric gains will be conducted to establish a definitive link between physics-aware input feature selection and enhanced fault tolerance. These findings will offer fresh perspectives on the practical application of AI to enhance GNSS positioning reliability, particularly highlighting the critical role of physics-aware input feature selection.

Previous Abstract Return to Session B3 Next Abstract