From Driving Simulator Experiments to Field-Traffic Application: Improving the Transferability of Car-Following Models

Abstract

Over the last few decades, there have been two main streams of data used for driving behavior research: trajectory data collected from the field [such as using video recordings and global positioning systems (GPS)] and experimental data from driving simulators (where the behaviors of the drivers are recorded in controlled laboratory conditions). Previous research has shown that the parameters of car-following models developed using simulator data are not directly transferable to the field. In this research, we investigate the differences in detail and compare alternative methods to overcome the problem. Two types of approaches are tested in this regard: (1) econometric approaches for increasing model transferability—Bayesian updating and combined transfer estimation—and (2) joint estimation using both data sources simultaneously. Car-following models based on a stimulus-response framework are developed in this regard, using experimental data collected at the University of Leeds Driving Simulator (UoLDS) and detailed trajectory data collected at California Interstate 80 (I-80), in the US, and the UK Motorway 1 (M1). The estimation results of the initial models show that car-following models using driving-simulator data are closer to the UK (M1) data than the I-80 data but not directly transferable. Performances of the proposed approaches for improving transferability are evaluated using t-tests for individual parameter equivalence and transferability test statistics (TTS). The results indicate that the transferability can be improved after parameter updating, and the combined transfer estimation is found to outperform the other approaches. The findings of this study will enable a more effective usage of the driving simulator data for the estimation of mainstream mathematical models of driving behavior while the techniques used can be applied to other types of econometric models.