Improving GNSS Localization in IoT Applications with Machine Learning and 3D Map Integration

In Internet of Things (IoT) applications, ultra-low-power GNSS receivers combined with cloud-based positioning engines often rely on a single-epoch positioning framework with a limited set of raw measurements. As redundancy and prior estimates cannot be leveraged for signal selection, these approaches pose significant challenges when it comes to GNSS degraded environments. To address these limitations, we propose a novel approach that integrates Machine Learning (ML) techniques with 3D map data to enhance the accuracy and robustness of GNSS localization in complex urban environments. Our method leverages a Long Short-Term Memory Neural Network (LSTM-NN) to predict pseudorange error magnitudes, while a simplified ray-tracing algorithm is used to derive map-based features. The predicted error magnitudes are utilized to weight measurements within the positioning engine, enabling reliable single-epoch positioning even when working with a limited set of signals. Real-world experiments demonstrate the effectiveness of our approach, showcasing strong performance in low-connectivity contexts. Furthermore, the methodology shows promise for extension to alternative signals such as Narrowband-IoT (NB-IoT) and 5G.