The car-following model (CFM) is essential for characterizing microscopic vehicular behavior, where accurate model calibration critically influences trajectory reproduction fidelity and driving behavior interpretation. Traditional approaches decouple calibration and behavioral analysis as calibration process of optimization to minimize trajectory deviations, and behavior analysis. This separation introduces two limitations: (1) conventional calibration methods prioritize numerical optimization over parameter interpretability, potentially distorting physical meaning; (2) reliance on manually engineered features fails to capture fine-grained behavioral patterns in high-dimensional trajectory data, constraining model generalizability. To address these challenges, we propose a dual framework of rule- and data-driven calibration, Multi Channels Variational Auto-Encoder – Car Following Model (MCVae-CFM), integrating behavioral semantics into time-varying parameter estimation. The framework employs: (1) a Multi Channels Variational Auto-Encoder to extract hierarchical behavioral representations from raw trajectories; (2) a Parameter Generator (PG) that bijectively maps latent representations to physically interpretable CFM parameters; and (3) a differentiable CFM computational graph ensuring dynamical consistency between generated parameters and real-world driving physics. Experimental validation demonstrates that MCVae-CFM achieves superior calibration efficiency (faster) while maintaining or exceeding accuracy across traditional methods. The case studies also reveal its capability to discern nuanced driving strategies (e.g., conservative vs. aggressive following) through parameter dynamics.
Yu et al. (Wed,) studied this question.