A Driving Regime-Embedded Deep Learning Framework for Modeling Intradriver Heterogeneity in Multiscale Car-Following...

Area of Science:

• Traffic flow dynamics
• Behavioral modeling
• Deep learning applications in transportation

Background:

• Existing car-following models struggle with intradriver heterogeneity, failing to capture dynamic behavioral changes within a single driver.
• Current models often overemphasize interdriver differences or use simplified assumptions, limiting accuracy in complex traffic scenarios.

Purpose of the Study:

• To develop a novel data-driven car-following framework that explicitly models intradriver heterogeneity.
• To integrate discrete driving regimes into continuous vehicular motion predictions for enhanced accuracy.
• To comprehensively represent both interdriver and intradriver behavioral variations.

Main Methods:

• Proposed a hybrid deep learning architecture combining Gated Recurrent Units (GRUs) for regime classification and Long Short-Term Memory networks (LSTMs) for kinematic prediction.
• Utilized high-resolution traffic trajectory datasets for training and validation.
• Employed a bottom-up segmentation algorithm and Dynamic Time Warping (DTW) to identify distinct driving regimes.

Main Results:

• The proposed framework significantly reduced prediction errors across multiple metrics compared to existing models.
• Demonstrated the model's ability to accurately reproduce complex traffic phenomena like stop-and-go waves and oscillatory dynamics.
• Successfully unified discrete decision-making processes with continuous vehicular dynamics.

Conclusions:

• The novel framework effectively addresses the challenge of intradriver heterogeneity in car-following modeling.
• Integrating discrete driving regimes enhances the prediction of vehicular motion and captures dynamic driver behavior.
• This approach offers a more comprehensive representation of driving behavior for improved traffic flow analysis and simulation.