The development of a holistic approach to modeling driver behavior : accounting for driver heterogeneity in car-following models
Car-following behavior has been studied since the 1940s. However, complex calibration requirements and challenges with collecting high-resolution data have stunted advancements in this domain. Thus, methodologies to adequately capture naturalistic behavioral heterogeneity are largely missing from the literature.
For this dissertation, a sample from the second Strategic Highway Research Program Naturalistic Driving Study was analyzed. This sample contains 665 trips completed on freeways in clear weather conditions. Driver demographics, vehicle CAN bus, and external sensor data are available for each trip. The trajectories in this sample were processed and used to calibrate the Gipps, Intelligent Driver Model, and Wiedemann 99 car-following models.
This dissertation seeks to improve how inter-driver heterogeneity in car-following behavior is accounted for in microsimulation models. This dissertation has three primary objectives. Objective 1 identifies which driver attributes are sources of inter-driver heterogeneity. Objective 2 explores the viability of using census-level data to calibrate microsimulation models. Objective 3 develops and evaluates a new mechanism for properly capturing inter-driver heterogeneity in microsimulation: an ensemble car-following model.
To achieve these objectives, first, Kruskal-Wallis one-way analysis of variance tests were applied to show statistically significant differences in both the estimated car-following model calibration coefficients and the overall model performance across groups of drivers categorized by commonalities in their driver attributes.
Next, the Expectation Maximization clustering algorithm was applied to show that, despite differences in driver behavior, homogeneous driver groups, or groups of drivers that behave similarly, exist in the dataset. Moreover, this dissertation shows that drivers can be classified into their proper homogeneous driver group only knowing their driver specific attributes.
Finally, VISSIM was used to implement the homogeneous driver groups in microsimulation. This case study illustrated that when inter-driver differences in driving behavior are explicitly modeled, there are notable impacts on the performance metrics collected from the microsimulation models. These performance metrics are ultimately used by decision makers to evaluate alternatives for transportation funding. Thus, this dissertation provides evidence of the importance of appropriately modeling inter-driver differences to improve the quality of the microsimulation model results and inform better funding allocation decisions.