Gaussian process based radio map construction for LTE localization

Location based services like localization in wireless network are drawing more and more attention in the recent years. According to published literatures, the fingerprint based method outperforms many other methods, where constructing an accurate fingerprint database is a new challenge. In this paper, we introduce a Bayesian regression model, Gaussian Process Regression(GPR) model to profile the signal strength values. The GPR is a nonparametric method which can be used to recover a complete radio map from a few fingerprint samples. We investigate the characteristics of different kernel functions and analyze the influence of applying them in radio map estimation. In order to find a suitable kernel for the Long Term Evolution(LTE) network, a kernel selection scheme is proposed based on construction of compositional kernels. Experiments are conducted based on data from telecommunication operators, demonstrating the feasibility of our proposed method.