Abstract Manager

Exploring the dynamic landscape of autonomous vehicles (AVs), this study investigates their transformative impact on real-time safety assessment in urban settings. AVs offer accurate, real-time information within a wide detection range, unlocking new possibilities for safety applications. As AV market penetration rises, their pivotal role in assessing and enhancing overall network safety emerges as a key contributor to the cost-effectiveness of AV deployment. Additionally, the realization of temporal fluctuations in safety levels in real-time presents great potential to reduce the overall risk of crash by mitigating periodic hazardous situations. The effectiveness of real-time safety models is intricately linked to the chosen time interval, influenced by factors such as data quality, continuity, and granularity, making a deeper understanding of the time intervals’ effects crucial for better applicability.

Addressing a significant gap in the literature, this research centers on the real-world applicability of AV data in urban safety assessment. An often-overlooked aspect in existing case studies is understanding the effect of the time interval on safety model performance. Our key objective is to investigate the impact of the time interval, as a hyperparameter, on model accuracy and precision through a multi-metric model comparison criterion.

Utilizing an Extreme Value Theory (EVT) model with a Block Maxima approach within a Bayesian hierarchical structure, the study facilitates multi-site analysis to overcome data scarcity in underrepresented sites. A diverse range of block sizes were employed to perform time interval sensitivity analysis, providing insights into the optimal choice for model robustness. For model evaluation, comprehensive metrics, including Deviance Information Criterion (DIC), Watanabe-Akaike Information Criterion (WAIC), and out-of-sample extreme estimates, are employed.

In conclusion, despite the challenges associated with the current real-world AVs, the proposed framework effectively informs real-time safety assessments, enhancing the reliability of safety models in urban settings. The introduction of a multi-metric model comparison criterion not only provides actionable insights into the interplay between time intervals and model performance but also offers a roadmap for practitioners and consultants navigating safety applications.