Deep learning-based Ultra-Wideband indoor positioning for integration of real and virtual spaces

Abstract

Recently, the Ultra-Wideband (UWB) has attracted attention for its ability to accurately track and localize objects and people in indoor environments. Many researchers use a deep learning approach for processing UWB signal features and obtaining precise indoor positions. However, in Non-Line of Sight (NLOS) conditions, UWB position estimation deteriorates, revealing significant positioning errors. In this paper, we propose the identification of the operating conditions of UWB, and the design of a deep learning-based positioning model tailored to the identified condition. We initially collected features of UWB Channel Impulse Response (CIR) under both Line of Sight (LOS) and NLOS conditions. Using this dataset and machine learning techniques, we identified operating conditions as LOS, Hard NLOS, and Soft NLOS. Subsequently, we designed deep learning models to minimize positioning errors occurring in Soft NLOS and Hard NLOS scenarios. The entire process takes place within the Head Mounted Display (HMD), which displays real-space positions in the virtual space. Through this process, synchronization of positions becomes possible, enabling the integration of real and virtual spaces.