Simultaneous probabilistic localization and learning: A new algorithm for online learning

Abstract

Indoor localization systems based on existent radio communication networks often make use of received signal strength (RSS) as measured feature. In order to achieve a good accuracy such systems have a huge payload in the called calibration phase, where many labeled measurements are collected and used to build a representative feature map. This paper presents a new algorithm based on previous works from the same authors, where explicit calibration efforts are avoided by unsupervised online learning, while the system is already operational. Using probabilistic localization and non-parametric density estimation, the new approach uses unlabeled measurements to learn a feature map with the probabilistic mass function of the measurements, having as start only a rough initial model based on plausible physical properties. Simulations with artificial generated data and with real measurements validate the introduced algorithm, covering discontinuities on the feature map and multimodal distributions, imposed by a structured indoor environment.