Deep Learning-Driven Adaptive-Weight Kalman Filtering for Low-Cost GNSS in Challenging Environments

Area of Science:

• Geomatics Engineering
• Signal Processing
• Machine Learning

Background:

• Smartphone Global Navigation Satellite System (GNSS) observations suffer from multipath and non-line-of-sight (NLOS) errors in urban environments.
• Traditional stochastic models (e.g., elevation-dependent, SNR-based) are inadequate for capturing complex, variable observation errors in dense urban canyons.
• Accurate parameter estimation requires robust stochastic models to handle GNSS measurement uncertainties.

Purpose of the Study:

• To develop an adaptive GNSS stochastic model using a deep neural network (DNN) to improve positioning accuracy in challenging urban environments.
• To integrate signal-to-noise ratio (SNR), satellite elevation angle, and post-fit pseudorange residuals as input features for the DNN.
• To adaptively weight GNSS measurements epoch-wise within a Kalman filtering process.

Main Methods:

• A fully connected DNN was designed to learn the nonlinear relationship between input features (SNR, elevation angle, residuals) and measurement uncertainty.
• The DNN output was used to adaptively update the measurement noise covariance matrix for real-time weighting.
• The proposed model was evaluated using data from a u-blox ZED-F9P receiver and a Samsung Galaxy S21+ smartphone during vehicle experiments in urban canyons.

Main Results:

• The DNN-based stochastic model significantly outperformed conventional models in single point positioning (SPP) accuracy for both smartphone and receiver data.
• Smartphone 3D RMSE decreased from ~13-14m to 8.94m (approx. 35% improvement) in obstructed urban environments.
• Receiver 3D RMSE improved from ~4-5m to 3.10m, demonstrating consistent performance gains.

Conclusions:

• The proposed DNN-based stochastic model effectively mitigates complex GNSS observation errors in urban settings.
• This approach offers a promising solution for reliable and accurate positioning using low-cost GNSS receivers, especially on smartphones.
• Adaptive weighting based on DNNs enhances robustness and accuracy in challenging geomatics applications.