Emergency Vehicle Car-Following Across Driving Regimes with Retrospective and Past-Only Regime Assignment

Abstract

This study examines whether regime-specific calibration improves the local acceleration representation of emergency-vehicle following episodes. Field trajectories of ambulances performing emergency-response tasks were obtained from two monitoring sites on Shaoshan East Road in Xiangtan, China, during February–December 2025 and screened into 499 valid following pairs. Standardised window features were reduced by principal component analysis and clustered using training pairs only. Three assignment protocols were compared: retrospective windows with acceleration, retrospective windows without acceleration, and acceleration-free windows completed strictly before the response time. Five car-following models were calibrated with fixed or regime-specific parameters and evaluated on identical observations. In the primary past-only analysis, the common test support comprised 1678 observations from 69 pairs. The Longitudinal Control Model (LCM) had the lowest test RMSE, 1.0478 m/s2, and reduced RMSE by 13.43% relative to its fixed-parameter counterpart; the paired pair-level bootstrap interval for the RMSE difference was [−0.2225, −0.0973] m/s2. Improvements for the other four models were smaller and their intervals included zero. LCM retained the lowest error with two and four clusters and under expanded parameter bounds, although its fitted parameters remained sensitive to constraints. These results support a dataset-specific benefit for LCM local acceleration representation.