Ashraf Md Tanvir, Dey Kakan
Department of Civil and Environmental Engineering, Michigan State University, Lansing, MI 48910, USA.
Accid Anal Prev. 2025 Mar;211:107897. doi: 10.1016/j.aap.2024.107897. Epub 2024 Dec 18.
Navigating intersections is a major challenge for autonomous vehicles (AVs) because of the complex interactions between different roadway user types, conflicting movements, and diverse operational and geometric features. This study investigated intersection-related AV-involved traffic conflicts by analyzing the Arogoverse-2 motion forecasting dataset to understand the driving behavior of AVs at intersections. The conflict scenarios were categorized into AV-involved and no AV conflict scenarios. Depending on whether AVs passed the conflict region first or second in AV-involved scenarios, AV-involved scenarios were further classified into AV-first and AV-second scenarios. An agglomerative hierarchical clustering with t-SNE dimension reduction technique was applied to categorize the driving styles, and a three-layer Bayesian hierarchical model was applied to analyze the effect of driving volatility measures and traffic characteristics on relative crash risks. The clustering result showed that about 29% of the conflict events in the AV-first scenario (human-driven vehicle (HDV) was the following vehicle in passing the conflict region) exhibited high-risk of conflicts. In contrast, all conflicts events in the AV-second category were either low-risk or medium-risk conflicts. Parameter estimates showed that AVs had safer interactions with the other roadway users (i.e., HDVs, pedestrians/cyclists) while maintaining higher speeds and uniform driving profiles. AV's interaction with vulnerable road users (i.e., pedestrians and cyclists) showed lower crash risk compared to HDVs, indicating AV's safer driving behavior. AVs also demonstrated safer conflict resolution behavior in performing unprotected left turns compared to HDVs. This study discovered some unique insights into the challenges of introducing AVs in diverse intersection types (i.e., signalized, unsignalized, stop-controlled), which can be used to identify AV technology's improvement need to better adapt to the mixed traffic driving environment.
由于不同道路使用者类型之间复杂的相互作用、冲突的交通流以及多样的运营和几何特征，自动驾驶车辆（AV）在交叉路口行驶是一项重大挑战。本研究通过分析Arogoverse - 2运动预测数据集来调查与交叉路口相关的自动驾驶车辆参与的交通冲突，以了解自动驾驶车辆在交叉路口的驾驶行为。冲突场景被分为有自动驾驶车辆参与的和无自动驾驶车辆冲突的场景。在有自动驾驶车辆参与的场景中，根据自动驾驶车辆是先通过还是后通过冲突区域，进一步将其分为自动驾驶车辆先通过和自动驾驶车辆后通过的场景。应用具有t - SNE降维技术的凝聚层次聚类来对驾驶风格进行分类，并应用三层贝叶斯层次模型来分析驾驶波动性指标和交通特征对相对碰撞风险的影响。聚类结果表明，在自动驾驶车辆先通过的场景（人类驾驶车辆（HDV）是通过冲突区域的后随车辆）中，约29%的冲突事件呈现出高冲突风险。相比之下，自动驾驶车辆后通过类别中的所有冲突事件均为低风险或中等风险冲突。参数估计表明，自动驾驶车辆在保持较高速度和均匀驾驶模式的同时，与其他道路使用者（即HDV、行人/骑自行车的人）有更安全的交互。与HDV相比，自动驾驶车辆与弱势道路使用者（即行人和骑自行车的人）的交互显示出较低的碰撞风险，表明自动驾驶车辆的驾驶行为更安全。与HDV相比，自动驾驶车辆在进行无保护左转弯时也表现出更安全的冲突解决行为。本研究发现了一些关于在不同类型交叉路口（即信号控制、无信号控制、停车控制）引入自动驾驶车辆所面临挑战的独特见解，可用于识别自动驾驶技术需要改进的方面，以更好地适应混合交通驾驶环境。
