Pedestrians represent a group of vulnerable road users who are at a higher risk of sustaining severe injuries than other road users. As such, proactively assessing pedestrian crash risks is of paramount importance. Recently, extreme value theory models have been employed for proactively assessing crash risks from traffic conflicts, whereby the underpinning of these models are two sampling approaches, namely block maxima and peak over threshold. Earlier studies reported poor accuracy and large uncertainty of these models, which has been largely attributed to limited sample size. Another fundamental reason for such poor performance could be the improper selection of traffic conflict extremes due to the lack of an efficient sampling mechanism. To test this hypothesis and demonstrate the effect of sampling technique on extreme value theory models, this study aims to develop hybrid models whereby unconventional sampling techniques were used to select the extreme vehicle–pedestrian conflicts that were then modelled using extreme value distributions to estimate the crash risk. Unconventional sampling techniques refer to unsupervised machine learning-based anomaly detection techniques. In particular, Isolation forest and minimum covariance determinant techniques were used to identify extreme vehicle–pedestrian conflicts characterised by post encroachment time as the traffic conflict measure. Video data was collected for four weekdays (6 am–6 pm) from three four-legged intersections in Brisbane, Australia and processed using artificial intelligence-based video analytics. Results indicate that mean crash estimates of hybrid models were much closer to observed crashes with narrower confidence intervals as compared with traditional extreme value models. The findings of this study demonstrate the suitability of machine learning-based anomaly detection techniques to augment the performance of existing extreme value models for estimating pedestrian crashes from traffic conflicts. These findings are envisaged to further explore the possibility of utilising more advanced machine learning models for traffic conflict techniques.

中文翻译：

---

重新审视异常检测和极值理论的混合方法，利用基于人工智能的视频分析获得的交通冲突来估计行人碰撞事故

行人是一群弱势道路使用者，他们比其他道路使用者遭受严重伤害的风险更高。因此，主动评估行人碰撞风险至关重要。最近，极值理论模型已被用于主动评估交通冲突造成的碰撞风险，这些模型的基础是两种采样方法，即块最大值和峰值超过阈值。早期的研究报告称这些模型的准确性较差且不确定性较大，这在很大程度上归因于样本量有限。造成如此差的性能的另一个根本原因可能是由于缺乏有效的采样机制而导致流量冲突极值选择不当。为了检验这一假设并证明采样技术对极值理论模型的影响，本研究旨在开发混合模型，使用非常规采样技术来选择极端的车辆与行人冲突，然后使用极值分布进行建模以估计碰撞事故风险。非常规采样技术是指基于无监督机器学习的异常检测技术。特别是，使用隔离森林和最小协方差行列式技术来识别以侵占后时间为特征的极端车辆与行人冲突作为交通冲突度量。我们从澳大利亚布里斯班的三个四足交叉路口收集了四个工作日（上午 6 点至下午 6 点）的视频数据，并使用基于人工智能的视频分析进行处理。结果表明，与传统极值模型相比，混合模型的平均碰撞估计更接近于观察到的碰撞，且置信区间更窄。这项研究的结果证明了基于机器学习的异常检测技术适用于增强现有极值模型的性能，以估计交通冲突造成的行人碰撞事故。这些发现旨在进一步探索利用更先进的机器学习模型来解决交通冲突技术的可能性。