Search

Digital maps of the transportation network are the foundation of future mobility solutions. Autonomous and connected vehicles rely on real-time, at-scale updating of the environment in which they operate. Successful operation in a hybrid environment, where human and machine intelligence coexist, requires explicit knowledge of the traffic regulator infrastructure. Future generation traffic management strategies and path planning systems must be tightly integrated with the regulator infrastructure in order to improve traffic dynamics and reduce congestion in urban environments. In this article, we present Intersense, a regulator identification and categorization system that leverages raw, naturalistic GPS trajectory data to infer the existence and type of regulators present at network intersections. A supervised learning approach, based on the eXtreme Gradient Boosting (XGBoost) algorithm, combines infrastructure and vehicle telemetry data features to analyze movement patterns vehicles follow when approaching intersections. For the most widely used regulator types (traffic light, stop sign and right of way), the achieved accuracy surpasses 96%. Permutation feature importance is used to evaluate the relative importance of the generated features, providing a detailed view of the classification model and insights for future data collection and processing stages. We showcase the Intersense's adaptability to various GPS data sources with uneven penetration and sampling rates and we illustrate the system's transferability across very diverse cities by performing experiments in San Jose and San Francisco. Finally, the relationship between the number of available trajectories and classification accuracy is derived to determine the necessary data collection investment. [Display omitted] • Intersense enables wide-scale regulator identification via GPS data with 96% accuracy. • The system is transferable to cities with different relief and regulator densities. • Intersense is robust to GPS data sources with uneven penetration and sampling rates. • Data size requirements and feature importance are evaluated. [ABSTRACT FROM AUTHOR]