Probabilistic Ultra-Wideband TDoA Localization with Bias Correction

Ultra-wideband (UWB) radio localization is a popular solution for indoor navigation. The time delay of radio signals between agents and anchors enables the inference of the agents' positions. The measurement of the time difference of arrival (TDoA) of these radio signals provides a scalable way to achieve localization. Due to factors like the antenna and room geometry TDoA measurements tend to contain a bias error. We present a probabilistic model-based approach to solve the TDoA localization problem with bias correction. By using stochastic variational Gaussian process (SVGP) regression with a tailored kernel we can exploit the problem structure and efficiently predict the measurement bias. Then we correct this bias by incorporating the Gaussian process (GP) predictions to a factor graph based localization scheme. The method is tested on data recorded from a quadrocopter and validated against an optical marker-based tracking. The framework manages to infer the location of the drone accurately and the proposed bias correction reduces localization errors significantly.