Ultra-Wideband (UWB) enables high-precision distance measurements owing to a bandwidth over 500 MHz. Its pulse-like signals can be converted into Channel Impulse Response (CIR), revealing detailed channel characteristics. However, real-world conditions such as Non-Line-of-Sight (NLOS) path, multipaths, and hardware noises degrade measurement accuracy. To address these challenges, we propose a machine learning-based ranging correction method using CIR-based features. We conducted experiments with four UWB anchors and 24 ground-truth points in an indoor parking lot. Localization error was calculated as the Euclidean distance between estimated and actual positions. Seven CIR-based features were extracted from the measured channel impulse responses, such as Root Mean Square delay spread. These features were used as the input for three regression models: Random Forest, Gradient Boosting Regression, and Support Vector Regression. An exhaustive feature selection identified four optimal inputs, reducing average localization error from 0.227 m (without ranging correction) to 0.127 m for the proposed method, which is a 44.1% improvement. To investigate the regression model performance, each feature’ s capability to distinguish Line-of-Sight (LOS) and NLOS was quantified using Effect Size and compared with Permutation Feature Importance. A strong correlation (r = 0.863) was observed, confirming the high LOS/NLOS separability features were more important in model learning.
Kim et al. (Fri,) studied this question.