Segment-then-refine: a general calibration framework incorporating intra-driver heterogeneity into Car-Following Models

Physics-based Car-Following Models (CFM) are fundamental to microscopic traffic flow research owing to their simplistic parametric structures, offering valuable insights into longitudinal driving behaviors among human drivers. However, the standard parameter calibration settings are limited in both calibration accuracy and generalization capability. Firstly, significant influential factors, such as intra-driver heterogeneity, are not effectively captured by calibrated CFMs, resulting in substantial calibration errors. Secondly, the estimated CFM parameters often fail to generalize to unseen driving scenarios. These problems cast doubt on the reliability of such CFMs. To address these challenges, this paper introduces a general segment-then-refine calibration framework designed to improve the calibration accuracy and generalization capability of CFMs. The segmentation stage incorporates intra-driver heterogeneity into CFMs by optimizing trajectory segments, each representing distinct driving behavioral pattern and their corresponding CFM parameters, thereby improving calibration accuracy. In the subsequent refinement stage, the CFM parameters of each trajectory segment are refined to enhance generalization capability using a feature-sharing approach based on an auto-regressive neural network (ARNN) model. The effectiveness of this framework is validated using simulation data to assess segmentation accuracy and real-world traffic data to evaluate trajectory simulation accuracy. Experimental results demonstrate that this framework outperforms standard calibration settings in both calibration accuracy and generalization capability, and is more effective in capturing distinct driving styles.