Modeling WLAN Received Signal Strengths Using Gaussian Process Regression on the Sodindoorloc Dataset

Abstract

While any wireless technology can be used for indoor localization purposes, WLANhas the advantage of having a huge existing infrastructure. A radio map that matches specific locations to received signal strength is needed, to enable most of these indoor localization methods. To create these radio maps, with enough detail to achieve sufficient localization accuracy, is expensive and time consuming. Therefore, methods to interpolate and extrapolate more detailed maps from sparse radio maps are being developed. One recent approach is to use Gaussian process regression. Even though some papers already studied Gaussian process regression, most studied only the basic model with zero mean and squared exponential kernel. In addition, when the model fit was evaluated in more detail, the experimental area was of limited complexity. Hence, this thesis evaluates the fit of Gaussian process regression, in a more complex indoor environment, based on adequate model metrics and analysis of the plots of the predicted mean and standard deviation functions. As a conclusion, the most suitable model is presented, as well as the reasoning why it was chosen.