Empowering 6G Positioning and Tracking with Bayesian Neural Networks

In the rapidly evolving domain of forthcoming 6th generation (6G) networks, achieving precise dynamic positioning down to the centimeter becomes critical, particularly in complex urban scenarios as those envisioned for cooperative intelligent transport systems (C-ITSs). To face the challenges introduced by severe path loss and blockages in new 6G frequency bands, machine learning (ML) provides innovative strategies to extract locational intelligence from wide-band space-time radio signals. This paper proposes the integration of Bayesian neural networks (BNNs) into cellular multi-base station (BS) tracking systems, where uncertainties of BNNs account for finite training sets and measurement errors. Our approach utilizes a deep learning (DL)-based autoencoder (AE) structure that exploits the full channel impulse response (CIR) to infer location-centric attributes in both line-of-sight (LoS) and non-LoS (NLoS) conditions. Validations in a 3rd Generation Partnership Project (3GPP) compliant urban micro (UMi) setting, simulated with ray-tracing and traffic simulations, demonstrate the superior performances of BNN-based tracking with respect to both traditional geometric-based tracking methods and state-of-the-art DL models.