Accurate Indoor Positioning for UWB-Based Personal Devices Using Deep Learning

Recently, there has been a rapidly emerging demand for localization technologies to provide various location-based services in indoor environments, such as smart buildings, smart factories, and parking lots, as well as outdoor environments. Ultra-wideband (UWB), an emerging wireless technology, provides opportunities for precise indoor positioning with sub-meter accuracy, much more accurate than WiFi or BLE-based techniques, thanks to its signal and communication characteristics. UWB technology has recently begun to be applied to personal devices such as smartphones and is expected to be used for indoor localization of personal devices. However, personal devices often cause signal problems because they are worn on human hands or bodies and move dynamically or are in a non-line-of-sight (NLoS) condition, such as pockets or bags. Therefore, these challenges in the dynamic environment of personal devices must be addressed to enable accurate indoor positioning services based on UWB. In this paper, we propose a novel UWB-based indoor positioning approach that significantly improves localization accuracy under dynamic personal device environments. Our proposed approach detects various NLoS conditions of personal environments by leveraging channel impulse response (CIR) and deep learning. Based on the detected NLoS conditions, the proposed approach adjusts the Kalman filter to adaptively estimate the position of the target UWB-based personal device. Specifically, the distance measurement errors by NLoS are minimized by applying weights determined by deep learning to the Kalman filter. Through experiments conducted in practical indoor environments, we have shown that our proposed approach considerably improves the accuracy compared to the traditional Kalman filter-based and trilateration approaches. According to our method of counting the number of points relative to ground truth, the proposed positioning system improved positioning accuracy significantly by 20.84% to 27.22% with an error tolerance of +/- 25 cm and by 7.78% to 22.78% with an error tolerance of +/- 50 cm, compared to the traditional approaches.