Abstract
Accurate and interpretable car-following models are essential for trafficsimulation and autonomous vehicle development. However, classical models likethe Intelligent Driver Model (IDM) are fundamentally limited by theirparsimonious and single-regime structure. They fail to capture the multi-modalnature of human driving, where a single driving state (e.g., speed, relativespeed, and gap) can elicit many different driver actions. This forces the modelto average across distinct behaviors, reducing its fidelity and making itsparameters difficult to interpret. To overcome this, we introduce aregime-switching framework that allows driving behavior to be governed bydifferent IDM parameter sets, each corresponding to an interpretable behavioralmode. This design enables the model to dynamically switch between interpretablebehavioral modes, rather than averaging across diverse driving contexts. Weinstantiate the framework using a Factorial Hidden Markov Model with IDMdynamics (FHMM-IDM), which explicitly separates intrinsic driving regimes(e.g., aggressive acceleration, steady-state following) from external trafficscenarios (e.g., free-flow, congestion, stop-and-go) through two independentlatent Markov processes. Bayesian inference via Markov chain Monte Carlo (MCMC)is used to jointly estimate the regime-specific parameters, transitiondynamics, and latent state trajectories. Experiments on the HighD datasetdemonstrate that FHMM-IDM uncovers interpretable structure in human driving,effectively disentangling internal driver actions from contextual trafficconditions and revealing dynamic regime-switching patterns. This frameworkprovides a tractable and principled solution to modeling context-dependentdriving behavior under uncertainty, offering improvements in the fidelity oftraffic simulations, the efficacy of safety analyses, and the development ofmore human-centric ADAS.