Beyond Safety: An Attention-Based Subjective--Objective Mapping Model for Autonomous Driving Intelligence Evaluation

Abstract

As autonomous driving systems advance beyond Level 3, traditional unidimensional evaluation methodologies are proving insufficient. A critical need exists for a reliable and trustworthy evaluation framework that aligns with human subjective judgment. Current approaches, however, often lack consensus on the relative importance of key performance dimensions and fail to capture how these evaluation priorities shift under varying scenario complexities. To address these gaps, a novel human-centered autonomous driving intelligence evaluation model, termed the Subjective-Objective Mapping Model (SOMM), is proposed. An evaluation index system is first established through the Delphi method, comprising five performance dimensions (safety, efficiency, comfort, regulatory compliance, traffic coordination) and one scenario dimension (scenario complexity). The SOMM is developed to map objective vehicle metrics onto subjective evaluation scores, demonstrating high predictive accuracy. Weight analysis quantifies the relative importance of performance dimensions, identifying safety and efficiency as the dominant factors in overall intelligence. Furthermore, a sensitivity analysis confirms that evaluation priorities are highly dynamic and scenario-dependent, shifting from traffic coordination and compliance in simple scenarios to safety and efficiency in complex environments. The proposed SOMM provides an accurate, interpretable, and dynamic evaluation framework that effectively quantifies autonomous driving intelligence by aligning it with human-centered criteria.