Unsupervised Learning of Propagation Time for Indoor Localization

Many of today´s localization systems for indoor and outdoor positioning are based on propagation time measurements of radio signals. In order to achieve high positioning accuracy in presence of non line of sight (NLOS) propagation, these systems require either an expensive manual calibration or additional information. In this paper, we present a novel approach for a channel impulse response (CIR) based fingerprint system, which reduces the calibration and measurement effort and simultaneously improves localization results. The basic idea is to initialize a simple model, which is improved by an online learning procedure using unlabeled measurements. This unsupervised learning algorithm is composed of two independent components, which exploit the similarity of neighboring CIRs as well as the temporal relation of measurements. Our tests indicate a significant improvement compared to traditional methods in case of time difference of arrival (TDoA) measurements. The algorithm can straightforwardly be adapted to arbitrary propagation time measurements.