We present a novel UWB-based positioning Dining system that incorporates a deep learning-based non-line-of-sight (NLOS) classifier and a switching tracking filter.Unlike conventional methods, the proposed classifier determines whether a measured distance was obtained under LOS or NLOS conditions by using not only the channel impulse response (CIR) but also the distance measurement and a previously estimated distance sequence.This combination enhances the detection of abrupt changes in distance, thereby improving classification accuracy.
Additionally, we propose switching filters that operate using different noise statistics based on the classification result.Specifically, we explore combinations of a particle filter and a robust Kalman filter to mitigate the impact of significant errors in NLOS signals.In experiments, the proposed classifier significantly outperformed conventional classifiers.
Furthermore, the switching filter-based methods demonstrated substantial improvements in localization performance compared to traditional approaches.Among the switching filters, the switching particle filter provided the best localization performance, though at the cost of higher computational complexity, whereas the switching robust Kalman Wash Basin filter was the fastest, albeit with slightly lower localization performance.