CSI-Based Proximity Estimation: Data-Driven and Model-Based Approaches

Abstract

Proximity estimation forms the backbone of contact tracing solutions, as quarantining potentially infected individuals is essential for controlling the spread of epidemics. It is also essential for device discovery in Device-to-Device (D2D) and Vehicle-to-Vehicle (V2V) communications, which will be critical in future 6G networks. Despite the widespread coverage of cellular networks, no previous work has evaluated cellular-based proximity estimation in an experimental setting. In this paper, we collect Channel State Information (CSI) measurements from an actual cellular network and utilize them to train and evaluate two proposed solutions. Capitalizing on CSI spatial correlation, we propose a data-driven method to classify devices based on their respective proximity. The proposed neural network has an accuracy of 91.18% when classifying devices as within 5 meters from one another or not, from only 10 seconds of CSI measurements. This method is complemented by a model-based approach that provides a solid theoretical model for estimating proximity. The model-based approach uses Bayesian inference of the conditional distribution of power correlation across devices, assuming spatially-correlated shadowing and stochastic geometry tools. The proposed Bayesian regressor fits our dataset better than the standard exponentially-decaying correlation model, reaching a $2.8 \times $ lower RMSE while requiring fine-tuning of only two more parameters. A Bayesian classifier is also proposed and reached a 91.67% accuracy on the binary case while also outperforming the data-driven model significantly at higher distances.

Keywords

• Bayesian methods
• cellular networks
• channel models
• channel state information
• machine learning
• proximity detection