Open Collections

Abstract

The increasing integration of autonomous vehicles (AVs) in urban traffic systems presents an opportunity for more advanced and proactive road safety management. However, the use of AV-generated data for real-time safety analysis presents several challenges due to the current low AV penetration rates, including data scarcity and spatial heterogeneity as well as non-stationarity of traffic conditions. This thesis addresses these challenges by developing a real-time road safety assessment framework using real-world AV data, with a focus on predicting crash risk and assessing vehicle exposure to hazardous conditions. To address the problems of data scarcity and imbalanced data, the framework employed an Extreme Value Theory (EVT) model with Bayesian hierarchical spatial random parameter (BHSRP) structure. The model estimates two key safety metrics, namely Risk of Crash (RC) and Return Level (RL), using Modified Time-to-Collision (MTTC) and Post-Encroachment Time (PET) as conflict indicators. A novel model comparison and validation criterion independent of crash records and sample size was introduced. The thesis also studied the risk of crash associated with different mobility Levels of service. The highest chance of a high-risk crash condition occurred during LOS D and LOS F for intersections and segments, respectively. To further quantify individual vehicle risk, a Risk Exposure (RE) index is introduced, reflecting the likelihood of encountering extreme conflict conditions based on the time spent under various levels of service (LOS). The results show that vehicles face the highest crash risk under LOS E and LOS F at intersections and segments, respectively. The thesis explored the transferability of real-time safety models to data-scarce regions. The transferred models performed reliably, demonstrating that safety models based on AV data can be adapted to other locations where data are limited. Through a multi-dimensional analysis incorporating various temporal aggregations and different EVT’s approaches, structures, and covariates, the study demonstrated that bivariate Peak over threshold (POT) models significantly outperform univariate models, particularly in scenarios requiring high temporal resolution, by achieving more accurate and reliable crash-risk estimates. These findings offer practical tools for enhancing real-time traffic safety assessment and contribute to the broader understanding of AV data utilization in dynamic urban environments.