HMM-based CSI Embedding for Trajectory Recovery from RSS Measurements of Non-Cooperative Devices

Constructing channel state information (CSI) maps may help wireless communications and localization. However, CSI map construction requires up-to-date CSI measurement data with location labels, which induces a huge challenge in practice. Conventional CSI embedding methods project the CSI to a low dimensional latent space which may not have a clear physical meaning for localization purpose. This paper attempts to extract the user locations from CSI measurements and recover the trajectory of the user in an outdoor vehicular communication scenario. A graph-based hidden Markov model (HMM) is constructed, and an alternating algorithm is developed to learn the model parameters and recover the trajectory of the user. A proof-of-concept experiment is conducted using real measurement data from 5G network and demonstrates a localization accuracy of 23 meters only based on reference signal received power (RSRP) measurements from a few nearby base stations, which is a promising result for CSI map construction.