RNN-based Robust Smartphone Indoor Localization on Ultra-wideband DL-TDOA

Indoor localization in the presence of permanent obstructions or temporary/partial obstacles is a unique challenge, due to the lack of practical channel models applicable in realistic scenarios. Ultra-wideband (UWB) signal-based localization, with its cm level accuracy, has gained a lot of research interest in recent years. Among the various UWB ranging methods, the downlink time difference of arrival (DL-TDOA) based localization, which does not require uplink transmissions or central computation, is the most practically applicable in large-scale industrial scenarios. However, the performance of DL-TDOA localization, based on the traditional least squares estimation method, can be severely affected by non-line-of-sight (NLOS) and multipath effects. Prior work investigates using Kalman filter-based methods on UWB and IMU sensor data for localization inaccuracy mitigation. This approach suffers from high localization error during UWB signal outage events, owing to its reliance on IMU data and the gaussian assumption of noise. In this paper, we propose a novel recurrent neural network (RNN)-based localization algorithm. The proposed algorithm utilizes the past UWB DL-TDOA values and the IMU sensor data to predict localization coordinates into the future and uses the model-generated augmented TDOA values during UWB signal outage. We also demonstrate the real-time performance of the RNN-based localizer, using NXP(SR100T) anchors and Samsung Galaxy S20 Ultra user devices. As shown in the experimental results, the proposed algorithm achieves 31% lower localization error compared to the Kalman filter-based baseline methods, while also being robust to signal outage.