A Comparative Study on Gaussian Process Regression-based Indoor Positioning Systems

Gaussian Process Regression (GPR) has been proved to be one of the most accurate ways of predicting online radio map for fingerprinting based localization, as it can better mimic the characteristics of wireless radio signals. However, the accuracy of the GPR model depends on the mean function used and most of the functions perform poorly while being used in localization. This paper presents a thorough comparative analysis on different Indoor Positioning Systems (IPS) exploiting GPR with different mean functions, among which zero mean and linear mean are the most commonly used ones. This paper also introduces two new mean functions-Single Hidden Layer Neural Network (NN) and Multiple Hidden Layer NN which outperforms traditional mean functions.