Research has shown that a driver’s emotional state is closely related to the driver’s driving behavior. However, most studies have focused on the impact of static and discrete emotions on driver behavior, neglecting the emotion fluctuations that can occur in real-world contexts. To address this gap, we analyzed emotion change patterns of drivers engaged in aggressive driving versus normal driving. We used online car-hailing data to identify aggressive driving, which is defined by a dynamic acceleration threshold determined from vehicle kinematics data. We used two dimensions, valence and arousal, to describe different levels of emotions by processing videos of drivers' facial expressions using FaceReader software. By analyzing the characteristics of drivers’ emotional changes, we determined a six-second time window as the length of an emotion fluctuation segment. Then, we applied dynamic time warping -means clustering as a time-series clustering method to divide the emotion fluctuations into several categories. We categorized the valence fluctuations into negative, calm, and positive and the arousal fluctuations into high and low. The clustering analysis revealed a marked fluctuation in the emotional valence of drivers during aggressive driving as opposed to the relatively smooth patterns observed during normal driving. Moreover, during aggressive driving, the driver’s arousal can continuously reach higher values compared to during normal driving. This study presents a novel approach to investigating the relationship between emotions and driving behavior and provides important reference and guidance for drivers’ emotional perceptions and the detection of emotion fluctuations for advanced assisted driving systems to improve driving safety.

中文翻译：

---

利用自然驾驶数据研究网约车司机驾驶行为的情绪波动


研究表明，驾驶员的情绪状态与驾驶员的驾驶行为密切相关。然而，大多数研究都集中在静态和离散情绪对驾驶员行为的影响上，忽略了现实世界中可能发生的情绪波动。为了解决这一差距，我们分析了激进驾驶与正常驾驶的驾驶员的情绪变化模式。我们使用网约车数据来识别攻击性驾驶，这是通过车辆运动学数据确定的动态加速度阈值来定义的。我们使用 FaceReader 软件处理驾驶员面部表情视频，使用效价和唤醒度两个维度来描述不同程度的情绪。通过分析驾驶员情绪变化的特点，确定6秒的时间窗口作为情绪波动段的长度。然后，我们应用动态时间扭曲均值聚类作为时间序列聚类方法，将情绪波动分为几类。我们将价态波动分为负、平静和正，将唤醒波动分为高和低。聚类分析显示，与正常驾驶期间观察到的相对平稳的模式相比，激进驾驶期间驾驶员的情绪效价存在显着波动。此外，在激进驾驶过程中，与正常驾驶期间相比，驾驶员的兴奋度可以持续达到更高的值。该研究提出了一种研究情绪与驾驶行为之间关系的新颖方法，为驾驶员情绪感知以及先进辅助驾驶系统情绪波动检测以提高驾驶安全提供重要参考和指导。