Crash frequency prediction based on extreme value theory using roadside lidar-based vehicle trajectory data

Crash prediction models (CPMs) are mostly developed using statistical or data-driven methods that rely on observed crashes. However, the historical crash records can be unreliable due to availability and data quality issues. Near-crashes based CPMs offer a proactive approach to predict crash frequencies prior to the occurrence of crashes. Surrogate safety measures can be used to identify near-crashes from road user trajectories. Roadside LiDAR offers an innovative approach to collect vehicle trajectory data at a microscopic resolution with high accuracy providing detailed information of all road user movements. This study presents a methodology to identify near-crashes from Roadside LiDAR based vehicle trajectory data using the surrogate indicators: TTC (Time to Collision), PET (Post Encroachment Time), ACT (Anticipated Collision Time) and MaxD (Maximum Deceleration). Additionally, time-based, and evasive-action-based surrogate measures are combined as different pairs to obtain crash probabilities using extreme value theory (EVT). The study results show that the bivariate EVT model displays a better fit to conflict extremes, predicting crash frequencies better than the univariate model. Likewise, while the bivariate model with ACT and MaxD pair performed the best in terms of accuracy, the TTC and MaxD pair was able to reflect the relative threat levels at the study intersections. Overall, the methodology lays ground for using roadside lidar based trajectory data for proactive safety analysis of signalized intersections.