Crash risk assessment at unsignalized intersections using vehicle trajectory data

Abstract

Crash prediction models (CPMs) typically use statistical or data-driven approaches derived from observed crash data, but these can be limited by unreliable historical data. Near-crash-based CPMs provide a proactive alternative, predicting crash frequencies before actual crashes occur. Surrogate Safety Measures (SSMs) examine potentially hazardous traffic events to improve the understanding of traffic safety dynamics. These events serve as proxies for crashes, enabling proactive and timely safety assessments. This study proposes a methodological framework for evaluating crash risk at unsignalized intersections using UAV-acquired vehicle trajectory data and applies Extreme Value Theory (EVT) to statistically model the tail behavior of a time-based SSM—Post Encroachment Time (PET). High-resolution (4 K) video data were acquired at six different unsignalized intersections to capture morning rush hour traffic (8 to 9 a.m.). Vehicle trajectories and surrogate measures such as Post Encroachment Time (PET) were extracted using advanced AI-driven video analysis via the DataFromSky (DFS) platform. The analysis employed the Peak Over Threshold (POT) method. The threshold was determined to be −1.25 s using the Mean Residual Life (MRL) plot, as well as the scale and shape parameter stability plots of the Generalized Pareto Distribution (GPD). The results show that traffic volume and crash frequency have a significant impact on collision risk. As traffic volume increases, PET decreases, leading to a higher likelihood of conflicts and crashes. Additionally, mean speed shows an inverse relationship with both crash frequency and collision risk. Overall, traffic volume and conflict frequency emerge as key predictors of crash risk occurrences. This study establishes a foundation for leveraging UAV-based vehicle trajectory data in conducting proactive safety assessments at unsignalized intersections.