A Driving Skill Evaluation Framework Based on Vehicle-State Deviations from Condition-Matched Model Responses

Objective and interpretable driving skill evaluation is important for vehicle-motion assessment, automated-driving-system evaluation, and driver monitoring. Existing approaches based on subjective ratings, fixed thresholds, or directly scored vehicle states can mix control–response differences with maneuver geometry and operating conditions. This paper proposes a vehicle-state framework that quantifies deviations between measured responses and condition-matched model responses. Measured trajectories are segmented into accelerated or decelerated lane-keeping and lane-changing units in a Frenet coordinate system. Unit boundary states define quintic reference trajectories, and a two-degree-of-freedom prediction model with MPC generates response signals through a detailed vehicle model. Deviations in motion accuracy, handling stability, and ride comfort are compressed, calibrated, and mapped into condition-wise and accumulated driving skill scores. The real-road case study used 20 Hz data from one general driver, one professional driver, and one production automated-driving-system record, each containing 43 identified units. Mean accumulated scores were 88.96, 87.48, and 85.11 for the professional driver, production system, and general driver, respectively. This ordering persisted across nine score-parameter combinations, criterion- and condition-weight perturbations, leave-one-expert recomputation, accumulation ablation, and leave-one-condition checks; the published empirical vehicle-motion-data index yielded the same descriptive record ordering. The framework provides a maneuver-specific and condition-consistent basis for evaluating driving skill manifested in human-driven and automated vehicle responses under the tested real-road conditions.