Methods to Obtain Representative Car-Following Model Parameters from Trajectory-Level Data for Use in Microsimulation

Microsimulation models help agencies obtain robust estimates of project benefits and spend their resources effectively. The realism of these models depends on the quality of input data and the realism of the sub-models controlling driver behavior. The availability of trajectory-level driving data provides new opportunities to improve car-following models and their application in practice. Procedures for calibrating car-following models using a single driving trajectory are well documented in the literature. However, methods for identifying a representative parameter set to describe a collection of observed driving behavior must be developed and tested before trajectory-level data can be applied in practice. This paper describes eight methods for obtaining representative sets of calibration parameters to describe a group of drivers or a specific driving condition. The methods are tested using a 100-trip sample from the SHRP2 Naturalistic Driving Study and validated with a 10-fold cross-validation procedure. The method capturing the average behavior while preserving underlying correlations between the calibrated model parameters performed the best across all four models. Methods that adequately captured the average behavior while relaxing the assumption of underlying parameter correlations performed better than all other tested methods. Therefore, simply taking the mean or median of the distribution of observed parameter values offers a practical approach for generating a representative parameter set, significantly outperforming default parameter values.