Centimeter-Level Indoor Localization using Channel State Information with Recurrent Neural Networks

This paper aims at solving two major drawbacks of fingerprint-based localization methods: 1) existing fingerprint-based methods mainly rely on the received signal strength indicator (RSSI) as the input feature, which renders centimeter-level localization impossible; 2) existing deep learning methods that rely on channel state information (CSI) as a feature do not consider the user trajectory and/or the signal-to-noise-ratio (SNR) information. To address these issues, this paper introduces a recurrent neural network (RNN) for centimeter-level indoor localization. The proposed RNN takes into consideration the user trajectory, as well as, the SNR information. We show that when the training data set is small, our proposed network beats the state-of-the-art neural networks. Moreover, for the first time, we present an extensive comparison between neural-network-based and decision-tree-based localization methods. The simulation results show that neural networks have higher estimation accuracy than tree-based methods.