Real-time safety analysis using autonomous vehicle data: a Bayesian hierarchical extreme value model

This study proposes an approach for real-time road network safety analysis using autonomous vehicles (AVs) generated data. The approach utilises a Bayesian hierarchical spatial random parameter extreme value model (BHSRP). The model simultaneously addresses the scarcity and non-stationarity of conflict extremes and unobserved spatial heterogeneity. Two real-time safety metrics are estimated: the risk of crash (RC) and return level (RL). The RC and RL were applied to three months AVs data for evaluating the real-time safety level of an urban corridor in Palo Alto, California. The indicator time to collision (TTC) was used to characterise traffic conflicts. The conflict extreme was defined as the maxima of negated TTC in a 20-min interval (block). The results show that RC can differentiate the block-level risk level, while RL can reflect safety levels among blocks. For the RC, the hot (crash risk prone) segments and intersections are associated with more severe conflict frequency.

Automated summary

This study proposes an approach for real-time road network safety analysis using data generated by autonomous vehicles (AVs). The approach utilizes a Bayesian hierarchical spatial random parameter extreme value model (BHSRP) to estimate the risk of crash (RC) and return level (RL), providing real-time safety measurements for urban corridors. The results highlight the association between high crash risk areas and severe conflict frequency.