Beyond Convolutions: Transformer Networks for Improved UWB CIR-based Fingerprinting

Abstract

Indoor positioning using UWB has gained popularity due to its low cost while still providing centimeter-level accuracy. Currently, Convolutional Neural Network (CNN)based approaches are often proposed for NLOS detection, error correction, etc. to make these UWB positioning systems more accurate. Transformer (TF) networks have shown to be a more capable alternative in several other domains, but have not been used for UWB fingerprinting. We present two novel TF-based approaches: one that processes channel impulse responses (CIR) directly, and a second one that uses cross-attention to incorporate anchor position information to improve geometric understanding. Moreover, we propose a second innovation by combining fingerprinting with a time calibration method that synchronizes the CIR data using a TDOA-based setup. This second innovation can be used with our novel approach, or with other state-of-the-art fingerprinting methods. The proposed models are evaluated in an industrial environment and outperform previous state-of-the-art CNNs in both LOS and NLOS situations, reaching accuracies with errors as low as 3cm in real-life conditions, while having lower complexity and requiring fewer samples.