Machine Learning-Based Roadside Vehicular Traffic Localization via Opportunistic Wireless Sensing

Abstract

Comprehensive Situational Awareness (SA) in mixed traffic environments (i.e., both autonomous and human-operated platforms) is a critical requirement in addressing some of the challenges that hinder the deployment of autonomous vehicle (AV) systems onto roadways. In this paper, a novel framework that leverages machine learning techniques for utilizing Signals of Opportunity (SoO) for robust localization of all vehicles operating along a stretch of roadway is presented. By making use of ubiquitous wireless emissions from vehicles, the presented approach performs vehicle localization without any active participation/assistance from vehicles thus making it a suitable candidate for SA in mixed traffic environments. Our simulation results show that given the road shape and number of vehicles present, observed 2D localization estimates generated by an arbitrary algorithm whose error is described by a Gaussian bivariate distribution with 10 meters covariance yields unbiased vehicle centroid estimates with less than one meter mean squared error by eight Kalman Filter (KF) iterations. A set of KFs for each vehicle are used to leverage the filtering of multiple estimates per vehicle per filter step to reduce measurement noise by averaging, while a clustering algorithm performs the dual role of forming KF set priors and classifying location estimates to their correct vehicle.