A bivariate, non-stationary extreme value model for estimating opposing-through crash frequency by severity by applying artificial intelligence-based video analytics

Multivariate extreme value modelling techniques are widely applied to estimate crash risks from traffic conflicts, with a predominant focus on rear-end crashes. In contrast, the suitability of conflict measures within a multivariate framework for estimating opposing-through crash risks has received less attention. This study proposes a non-stationary bivariate extreme value model to identify a suitable set of traffic conflict measures for estimating opposing-through crashes (i.e., right-turn crashes for left-hand driving conditions and vice versa) by severity levels. In the proposed Generalised Extreme Value model, three crossing course conflict measures were considered, including post encroachment time (PET), gap time (GT), and supplementary time-tocollision (T2min). Artificial intelligence-based video analytics were employed to extract these opposing-through conflict measures from a total of 144 h of video recordings of four permissible right-turn approaches for three signalised intersections in Brisbane, Australia. The models included exposure variables such as conflicting volume, right-turning volume and through volume, and evasive action-based variables like deceleration and relative velocities measured at the signal cycle level to account for non-stationarity in the extreme value models. Results suggested that a bivariate model with PET and GT as the traffic conflict measures performs better than a univariate model or other combinations of traffic conflict measures in the bivariate models. This PET-GT combination of conflict measures also showed better accuracy in estimating opposingthrough crash frequencies by severity levels when combined with the (Delta-V) based severity measure. This study demonstrated the importance of accounting for various stages of opposingthrough conflicts within a bivariate extreme value model to predict crash risks.