A rigorous evaluation of Gaussian process models for WLAN fingerprinting

Abstract

Location based services require accurate and seamless positioning in large urban areas. In contrast to GNSS, WLAN fingerprinting positioning offers seamless localization in these areas. Though, it requires a huge effort to create the radio maps. Interpolating radio maps is a viable solution; in particular Gaussian process (GP) regression is very effective for this task. Based on a thorough evaluation of different Gaussian process models we appoint the best suited model for spatial signal strength interpolation. We pursue the model evaluation by establishing GP maximum likelihood (ML) estimators and assess their effects on the positioning accuracy in a realistic WLAN indoor/outdoor localization scenario. Insights on the spatial density of fingerprints are included in our study. We found that the commonly used GP model, with zero mean and squared exponential covariance function, is not the best suited model and propose a better and more robust alternative. Moreover, this study demonstrates that a low amount of fingerprints not necessarily impairs, but potentially improves the accuracy of the ML estimators.