Gaussian processes for learning-based indoor localization

The development of an efficient and accurate location sensing systems for indoor environments, based on the received signal strength (RSS) data, is usually a challenging task. In this paper, we discuss the feasibility of using Gaussian Processes (GPs) regression for learning based indoor localization algorithm. The GP is one of the machine learning algorithms that can be used to model a complete RSS map from few training data. We investigate the use of three different covariance functions, i.e. Squared Exponential (SE), Matérn, and Rational Quadratic (RQ), to find the suitable one for the indoor localization, and then compare their performance to the traditional weighted k-Nearest Neighbors (k-NN) algorithm. We show that GP regression can significantly outperform the k-NN, while keeping the training cost at a reasonable level. Furthermore, although, the smoothness property of SE covariance function, we demonstrate that GP-SE covariance provides better accuracy compared to GP-Matérn and GP-RQ, particularly, when a few training data are available.