Empirical Error Modeling of Android GNSS Using Machine Learning for PVT Improvement

Accurate modeling of GNSS measurement errors is imperative for an effective computation of GNSS position and velocity using raw measurements. This is because the model allows us to quantitatively define the accuracy of each measurement. There are multiple different ways to model the errors including the theoretical approaches using elevation angles and signals’ Carrierto-Noise-Ratio-Densities (C/No). Also, in the case of Android devices, the embedded Application-Specific Integrated Circuit (ASIC) chipsets also provide the expected measurement noises. Furthermore, there are empirical methods where the errors can be modeled using training data. The questions that may be raised include which theoretical model will provide the most accurate position, velocity, and time (PVT) solution, whether the chipset reported noises can be trusted, and how the empirical models should be trained. The novelty of the paper lies in addressing all these pending questions. The paper assesses the positional accuracies from each model and shows that the theoretical model, incorporating both elevation angles and C/No features, achieves the highest accuracy. Also, for the chipsets, Samsung and Qualcomm provide noise values that are representative of the expected measurement residuals, while MediaTek and Broadcom provide over-bounding estimates. Finally, the paper discusses the empirical modeling approach, detailing how an Extended Kalman Filter (EKF) can be applied to estimate and eliminate receiver clock bias and drift terms from the residuals before using them as model labels. For the model features, elevation angles, C/No, and user heading are utilized. A tree-based machine learning (ML) regression model is employed. The results indicate that when test environments closely resemble the training environments, the empirical models outperform the theoretical ones; however, when the environments differ, there is a decline in performance. All data for this study comes from the publicly available Google Smartphone Decimeter Challenge (GSDC) 2023 dataset, making the proposed methodologies and analysis easily replicable.