A Cycle-Level Distribution-Based Calibration of Microscopic Intersection Models Using UAV Trajectories

Abstract

Accurate calibration of microscopic traffic simulation is important at signalized intersections, where errors in queue formation and turning behavior can propagate as models scale to larger urban networks. This paper presents a UAV-driven, distribution-based calibration framework for SUMO that targets cycle-level intersection dynamics rather than aggregate link counts. Using pNEUMA UAV trajectory data from Athens, we jointly calibrate key SUMO components that govern intersection behavior, including car-following (Krauss), lane-changing, and junction interaction parameters.The observation horizon is segmented into signal cycles, and within each cycle vehicles are assigned to a signal phase and an exit direction. Cycle data are then aggregated by phase and direction to construct cumulative exit-count curves that act as distributional targets. Simulated outputs are processed in the same way, and calibration minimizes discrepancies between observed and simulated curves using the NGOpt optimizer from the Nevergrad library. To reflect UAV data gaps and avoid error accumulation across cycles, each cycle is reinitialized and simulated independently, enabling parallel evaluation of objective calls.Results show stable convergence of NGOpt. For the dominant phase--direction movement, the normalized objective deviation (nABC) is as low as 0.46%, while the maximum cumulative deviation does not exceed two vehicles per cycle in any phase-direction pair. Compared to SUMO default parameters, median cycle-level MAPE decreases by approximately 70% in the higher-volume movements, even when both scenarios are initialized from the same UAV-derived vehicle states.