The widespread deployment of Connected Autonomous Vehicles (CAV) faces a critical challenge: public concerns regarding Autonomous Driving System (ADS) decision-making still impede adoption. While existing CAV intelligence assessment frameworks emphasize technical metrics, they often neglect passenger perceptions and responses that shape acceptance. This study proposes a passenger-driven evaluation framework integrating intervention behaviors and acceptance metrics across safety, smoothness, and autonomy. A three-vehicle CAV fleet was tested in the CARLA platform under six safety-critical scenarios, including emergency braking, pedestrian conflicts, and complex mergers. Human-in-the-Loop (HITL) experiments with 30 participants examined intervention behaviors under cautious, normal, and aggressive driving modes. To quantify CAV intelligence under varied configurations, Principal Component Analysis (PCA) and a weighted geometric mean method were applied. Results show intelligence scores decreased from cautious to aggressive modes, with the cautious mode achieving 90.7, approximately 27.6% higher than the aggressive mode’s lowest score of 71.1. Vehicle position and the system algorithm also influenced assessment outcomes: following vehicles demonstrated higher autonomy scores because of fewer interventions, whereas ego vehicles achieved greater smoothness through a more sophisticated control strategy. Across 27 valid experiments, 115 interventions were recorded, and collisions occurred in 33.3%. Notably, many interventions were unnecessary and counterproductive, as premature or insufficient braking often disrupted ADS response effectiveness. This research contributes a multi-dimensional intelligence assessment framework that complements existing technology-centered approaches with passenger-centric metrics. By bridging the perception gap between human passengers and ADSs, the proposed framework offers valuable insights for optimizing CAV driving behavior and enhancing public trust and acceptance.
Zhu et al. (Sun,) studied this question.