Lightweight Machine Learning Models for UWB Localization and Gesture Recognition

Cristian-Alexandru Tanase, Anamaria Dumitrescu, Alin Banel Dumitru Trasca, Bogdan Mocanu

Abstract: Ultra-Wideband (UWB) technology enables high-precision indoor localization, but its accuracy is degraded by multipath propagation, non-line-of-sight conditions, and intrinsic measurement errors. This work investigates the use of lightweight machine learning models to enhance UWB performance in two directions: localization accuracy and gesture recognition. For localization, we integrate distance and angle-of-arrival measurements from a single-anchor Two-Way Ranging (TWR) setup into a Convolutional Neural Network (CNN), achieving a root mean squared error (RMSE) of 10 cm, compared to 20 cm for a moving-average baseline. For gesture recognition, we construct a dataset of digits (0–9) drawn in free space with a UWB tag and train a long short-term memory (LSTM) model, reaching a test accuracy of 92% with real-time inference of ~24 ms. The proposed framework demonstrates that UWB systems, when coupled with efficient deep learning models, can deliver more reliable localization while enabling novel, privacy-preserving interaction modalities such as smartwatch-based gesture control.

External IDs:dblp:conf/wpmc/TanaseDTM25