LSTM-Assisted UWB Indoor Positioning Using CIR-Based NLoS Probability and Adaptive Kalman Filtering

Accurate indoor positioning using ultra-wideband (UWB) technology is essential for location-based services in personal devices, yet its performance degrades significantly under non-line-of-sight (NLoS) conditions commonly encountered in indoor environments. Factors such as human body shadowing, device orientation, and surrounding obstacles introduce biased ranging measurements, leading to poor localization accuracy. To address this challenge, this paper proposes an LSTM-based UWB indoor positioning framework that exploits the temporal characteristics of channel impulse response (CIR) signals. An LSTM network is employed to estimate the probability of NLoS propagation from raw CIR sequences. This probability is then integrated into a weighted adaptive Kalman filter to dynamically correct unreliable distance measurements. The corrected distances are subsequently used in a trilateration-based positioning algorithm to estimate the device location. Experimental evaluation using a publicly available UWB CIR dataset demonstrates that the proposed approach consistently outperforms conventional trilateration and standard Kalman filter methods. Specifically, the proposed framework improves positioning accuracy by approximately 17% at a ±25 cm error tolerance and 16% at a ±50 cm tolerance, while achieving nearly 99% accuracy within a ±100 cm error bound. These results validate the effectiveness of combining temporal deep learning with adaptive filtering for robust indoor UWB localization.