An extreme value theory based approach for calibration of microsimulation models for safety analysis

There is an increasing interest in using microsimulation models for traffic safety evaluations and several calibration procedures were proposed for the simulation models to obtain a high correlation between the simulated and field-measured conflicts. However, since the ultimate goal of safety studies is to reduce crashes, the simulation model needs to be calibrated to ensure an accurate prediction of crashes. The objective of this study is to propose an innovative calibration approach for microsimulation models using the extreme value theory (EVT) approach. The EVT approach was shown to connect non-crash events to crashes, providing reasonable crash estimates from traffic conflicts. The goal of the proposed calibration process is to produce an extreme value distribution of simulated conflicts that matches the one of field-measured conflicts. A Genetic algorithm is utilized to obtain VISSIM model parameters that can achieve this goal. Traffic video data collected from two approaches at a signalized intersection in Surrey, Canada were used as case studies. Automated traffic conflicts analysis techniques were used to extract field-measured conflicts. Simulated conflicts were extracted from vehicle trajectories from VISSIM using the surrogate safety assessment model (SSAM) tool. EVT models were developed to estimate the generalized Pareto (GP) distributions of both field-measured conflicts and simulated conflicts of different scenarios. The calibration results show considerable matching of the estimated GP distributions between simulated conflicts and field-measured conflicts. As such, the proposed calibration procedure for microsimulation model is recommended as a promising approach in simulation-based safety evaluation of signalized intersections.