UWB and GNSS Sensor Fusion Using ML-Based Positioning Uncertainty Estimation

Abstract

This article presents a novel machine learning (ML) augmented sensor fusion scheme for seamless indoor-outdoor positioning using ultra-wideband (UWB) and global navigation satellite system (GNSS) sensors. Dilution of precision (DOP) is a common metric that describes the level of geometrical uncertainty and is often applied in sensor fusion schemes. However, it does not fully reflect other factors that may influence positioning integrity, such as signal quality, pseudorange error, or the number of servicing nodes. An incorrect estimation of sensor uncertainty may significantly affect the precision and robustness of individual sensors and their fused positioning solution, especially in indoor-outdoor transition zones, where positioning is most challenging. Therefore, this article proposes a sensor fusion scheme augmented with two distinct extreme gradient boosted (XGBoost) ML models for estimating UWB and GNSS positioning uncertainties. Trained using real-life datasets, these models have the advantage of considering an ensemble of features rather than a single parameter to estimate the uncertainty of the current coordinate. In contrast to sensor fusion solutions, which implement only highly accurate RTK fixed solution mode, the GNSS model can operate with different correction qualities as well. The proposed scheme was applied on a moving testbed with UWB and GNSS sensors while relying on the ML models to enhance the coordinate filtering. The results show that the ML-based approach can improve seamless transition between indoor and outdoor areas with almost no sensor dropouts with a mean positioning error of 0.16 m and a maximum error of approximately 0.5 m.