Xu Zheng, Wang Xiaomeng, Wang Xuesong, Zheng Nan
Department of Civil Engineering, Monash University, Melbourne, VIC 3800, Australia.
College of Transport and Communications, Shanghai Maritime University, Shanghai, China.
Accid Anal Prev. 2025 Jun;215:108011. doi: 10.1016/j.aap.2025.108011. Epub 2025 Mar 18.
Public concern over the implementation of Connect Autonomous Vehicles (CAVs) remains a significant issue, and safety validation for CAVs remains a critical challenge due to the limitations of existing testing methods. While real-world testing is crucial, it can be expensive, time-consuming, and potentially impractical for evaluating the operation of CAV fleets. This paper presents a comprehensive co-simulation framework integrating the fully compiled CARLA with traffic microsimulation to establish a large-scale (20 × 20 km) testing environment for systematic CAV safety validation. The framework encompasses three key components: 1) a high-fidelity testing environment featuring diverse road geometries and dynamic conditions including weather variations and realistic traffic flows; 2) an intelligent CAV function developed through deep reinforcement learning and enhanced with utility-based connectivity strategies; 3) A sophisticated safety measurement metric that utilizes surrogate safety assessments, integrating a multi-type Bayesian hierarchical model to comprehensively evaluate risk factors and incident probabilities. The case study assessed CAV penetration rates ranging from 0 % to 100 %, identifying an optimal safety performance at a 70 % penetration rate, which resulted in an 86.05 % reduction in accident rates compared to conventional driving scenarios. This optimal safety level was effectively achieved in rural and suburban areas, where the average conflict probability was 0.4. However, in transition zones that connect high-, medium-, and low-density areas, significant traffic conflicts persisted even at this optimal penetration rate, with a conflict probability exceeding 0.7. Key results highlight critical safety patterns under optimal conditions, revealing that roundabouts and signalized intersections account for over 70 % of conflicts involving CAVs. This work advances CAV safety validation by providing a more realistic, large-scale testing environment that compensates for real-world testing limitations and allows for comprehensive safety evaluations across diverse scenarios.
公众对联网自动驾驶汽车（CAV）实施的关注仍然是一个重大问题，由于现有测试方法的局限性，CAV的安全验证仍然是一项严峻挑战。虽然实际测试至关重要，但对于评估CAV车队的运行而言，它可能成本高昂、耗时且可能不切实际。本文提出了一个综合协同仿真框架，将完全编译的CARLA与交通微观仿真相结合，以建立一个大规模（20×20公里）的测试环境，用于系统的CAV安全验证。该框架包含三个关键组件：1）一个高保真测试环境，具有多样的道路几何形状和动态条件，包括天气变化和现实的交通流量；2）通过深度强化学习开发并采用基于效用的连接策略增强的智能CAV功能；3）一种复杂的安全测量指标，利用替代安全评估，集成多类型贝叶斯层次模型，以全面评估风险因素和事故概率。案例研究评估了从0%到100%的CAV渗透率，发现在70%的渗透率下具有最佳安全性能，与传统驾驶场景相比，事故率降低了86.05%。在农村和郊区有效地实现了这一最佳安全水平，那里的平均冲突概率为0.4。然而，在连接高密度、中密度和低密度区域的过渡地带，即使在这个最佳渗透率下，仍存在重大交通冲突，冲突概率超过0.7。关键结果突出了最佳条件下的关键安全模式，表明环形交叉路口和信号控制交叉口占涉及CAV冲突的70%以上。这项工作通过提供一个更现实的大规模测试环境推进了CAV安全验证，该环境弥补了实际测试的局限性，并允许对不同场景进行全面的安全评估。
