Impact of Feature Selection and CIR Window Length on NLoS Classification for UWB Systems.

Indoor localization systems based on UltraWideBand (UWB) technology can typically achieve cm-level accuracy, but their performance degrades in Non-Line-of-Sight (NLoS) conditions. To cope with this problem, Machine Learning (ML) techniques have been applied to detect such NLoS conditions and adapt the localization algorithm accordingly. However, such ML techniques are typically optimized for accuracy, resulting in computationally-complex models that cannot be run on resource-constrained UWB devices. In this paper, we study and propose methods to reduce the computational complexity of NLoS classification models by applying ML-based feature selection and by reducing the window length of the channel impulse response for feature extraction. Specifically, we consider 29 features and study the effect of feature selection across five different datasets to obtain generalizable results. We show that we can extract two sets of only 3 and 8 features, which result in tiny ML models (smaller than 1kB), and low computation times ($3.6 \mathrm{~ms}$ and $27.7 \mathrm{~ms}$ on a 80MHz ESP8266 microcontroller, respectively). This allows a reduction of the runtime by more than $90 \%$ compared to the state of the art, while still maintaining an average classification accuracy above $85 \%$ across all five datasets.