Road State Inference via Channel State Information

TL;DR

This paper presents a novel traffic monitoring method using physical layer samples from vehicular communications and machine learning, achieving over 80% accuracy in predicting traffic conditions without extra infrastructure.

Contribution

The study introduces a new approach leveraging existing vehicular communication signals and machine learning for traffic state inference, validated through simulations and real-world experiments.

Findings

Achieved over 80% accuracy in traffic level prediction

Successfully estimated vehicle counts with low error

Validated approach in both simulated and real environments

Abstract

A wide variety of sensor technologies are recently being adopted for traffic monitoring applications. Since most of these technologies rely on wired infrastructure, the installation and maintenance costs limit the deployment of the traffic monitoring systems. In this paper, we introduce a traffic monitoring approach that exploits physical layer samples in vehicular communications processed by machine learning techniques. We verify the feasibility of our approach with extensive simulations and real-world experiments. First, we simulate wireless channels under realistic traffic conditions using a ray-tracing simulator and a traffic simulator. Next, we conduct experiments in a real-world environment and collect messages transmitted from a roadside unit (RSU). The results show that we are able to predict different levels of service with an accuracy above 80% both on the simulation and experimental data. Further, the proposed approach is capable of estimating the number of vehicles with a low mean absolute error on both data. Our approach is suitable to be deployed alongside the current monitoring systems. It doesn't require additional investment in infrastructure since it relies on existing vehicular networks.