Improving the Correction of NLoS-Induced Ranging Errors in UWB Systems through Enhanced Labeling

Ultra-wideband (UWB) ranging estimates taken in Line-of-Sight (LoS) conditions are typically centimeter accurate, but large errors may occur when devices are in Non-Line-of-Sight (NLoS). Several techniques, often based on machine learning, have been proposed to correct NLoS-induced ranging errors. These techniques are typically designed to effectively correct ranging estimates obtained in NLoS conditions, but significantly worsen the ranging estimates taken in LoS. We show that this problem stems from the method commonly used to label LoS/NLoS conditions and to train classification and regression models, which results, among others, in error correction techniques being applied to LoS samples that are erroneously classified as NLoS. To tackle this problem, we propose ranging error labeling (RELa); an approach that defines more suitable labels, based on the severity of the ranging error, to guide the decision of whether applying ranging error correction and the selection of which data should be used to train the error correction model. The effectiveness of RELa hinges on the careful selection of two
thresholds: one identifying which ranging estimates should be corrected (α), and one defining the training set used for the error correction model (β). After deriving suitable values for these
two thresholds empirically and showing their applicability to different settings and UWB platforms, we evaluate and prove the effectiveness of RELa. Our experimental results show that RELa reduces the error introduced on ranging estimates taken in LoS by up to 99%, without significantly impacting the effectiveness of ranging error correction in NLoS conditions.