Graph-Based Machine Learning for Practical Indoor Localization

Abstract

Machine learning approaches using channel state information (CSI) measurements can achieve accurate indoor localization. In this research, we propose a graph-based indoor localization system, that constructs a graph convolutional network (GCN) based on the geography of access points (APs). An AP pair with a high received signal strength indicator forms a link. Our GCN is efficient in finding relevant CSI measurements received from different APs. In a complex building environment, without any information on the building structure, our scalable system obtains an improved localization accuracy of 1.19 m. To the best of our knowledge, this is the first approach that introduces a graph-based network to find spectral features to improve indoor localization accuracy.