[论文解读] Crash Themes in Automated Vehicles: A Topic Modeling Analysis of the California Department of Motor Vehicles Automated Vehicle Crash Database

Hananeh Alambeigi, Anthony D. McDonald+1 more|arXiv (Cornell University)|Jan 29, 2020

Computational and Text Analysis Methods参考文献 27被引用 36

一句话总结

本论文使用对加州 DMV 自动驾驶车辆事故叙述进行概率主题建模，以识别五个事故主题并强调驾驶员-自动化交互中的安全差距。

ABSTRACT

Automated vehicle technology promises to reduce the societal impact of traffic crashes. Early investigations of this technology suggest that significant safety issues remain during control transfers between the automation and human drivers and automation interactions with the transportation system. In order to address these issues, it is critical to understand both the behavior of human drivers during these events and the environments where they occur. This article analyzes automated vehicle crash narratives from the California Department of Motor Vehicles automated vehicle crash database to identify safety concerns and gaps between crash types and current areas of focus in the current research. The database was analyzed using probabilistic topic modeling of open-ended crash narratives. Topic modeling analysis identified five themes in the database: driver-initiated transition crashes, sideswipe crashes during left-side overtakes, and rear-end collisions while the vehicle was stopped at an intersection, in a turn lane, and when the crash involved oncoming traffic. Many crashes represented by the driver-initiated transitions topic were also associated with the side-swipe collisions. A substantial portion of the side-swipe collisions also involved motorcycles. These findings highlight previously raised safety concerns with transitions of control and interactions between vehicles in automated mode and the transportation social network. In response to these findings, future empirical work should focus on driver-initiated transitions, overtakes, silent failures, complex traffic situations, and adverse driving environments. Beyond this future work, the topic modeling analysis method may be used as a tool to monitor emergent safety issues.

研究动机与目标

• 促使理解现实世界情境中自动驾驶车辆事故如何发生。
• 从 DMV AV crash narratives 中识别常见的事故主题。
• 评估事故类型与当前 AV 安全研究重点的一致性。
• 为新兴安全问题提出经验研究和监测方法的方向。

提出的方法

• 将概率主题建模应用于来自 California DMV 数据库的开放式 AV 事故叙述。
• 识别代表自动驾驶车辆事故中安全关切的连贯主题（话题）。
• 分析主题之间的关联性（如驾驶员触发的转换与擦挂事件）。
• 突出复杂交通互动和不利环境在事故中的作用。
• 提出将主题建模方法作为对新兴 AV 安全问题的监测工具。

实验结果

研究问题

• RQ1 在 California DMV 自动驾驶车辆事故数据库中，哪些事故主题最为普遍？
• RQ2 驾驶员触发的转换与擦挂或追尾等其他事故类型之间有何关系？
• RQ3 在自动驾驶车辆事故中出现了哪些当前研究尚未充分解决的安全关切？

主要发现

• 在 DMV AV crash narratives 中识别出五个主题。
• 驾驶员触发转换的事故常与擦挂事故同时发生。
• 许多擦挂事故涉及摩托车。
• 在自动驾驶模式下，追尾事故发生在如在路口停驶、在转弯车道、或与迎面来车相遇等情景。
• 研究结果突出自动模式下的控制转移和交互的安全隐患，建议未来工作的重点领域。