Gaussian processes for real-time urban traffic estimation using a fleet of drones

Abstract

Urban traffic monitoring requires accurate network-wide estimates even though only a limited portion of the road network can be observed in real time. At the same time, the growing deployment of Unmanned Aerial Vehicles (UAVs) in everyday applications is creating new opportunities for high-resolution, real-time traffic monitoring. Compared with traditional fixed loop detectors, fleets of UAVs acting as mobile sensing networks offer substantially greater flexibility, improved accuracy, and broader spatial coverage. In this paper, we propose a Gaussian-process-based framework for real-time estimation and short-term prediction of urban traffic variables from partial drone observations. The area of interest is represented as a weighted graph, and the proposed covariance structure combines a graph Matérn kernel, which captures the topology of the road network, with a temporal kernel that models the evolution of traffic over time. In addition to providing networkwide estimates and predictions, the proposed framework also yields a principled quantification of predictive uncertainty at unobserved locations. The model is trained and evaluated using microsimulations of Barcelona’s urban road network under multiple demand scenarios, demonstrating the effectiveness of Gaussian processes for traffic parameter estimation and prediction.