A novel approach for real time crash prediction at signalized intersections

• Published in: Transportation research. Part C, Emerging technologies Vol. 117; p. 102683
• Main Authors: Zheng, Lai, Sayed, Tarek
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.08.2020
• Subjects: Bayesian hierarchical model; Extreme value theory; Real time crash prediction; Signalized intersection; Signalized intersection; Bayesian hierarchical model; Extreme value theory; Real time crash prediction
• ISSN: 0968-090X; 1879-2359
• DOI: 10.1016/j.trc.2020.102683

•A Bayesian hierarchical extreme value model is developed for intersection safety prediction at the signal cycle level.•Traffic conflicts automated extracted from informative vehicle trajectories are used for model development.•Non-stationarity and unobserved heterogeneity in conflict extremes are accounted for to improve safety estimates.•Risk of crash (RoC) and return level of a cycle (RLC) are developed as real-time safety indicators.•RoC and RLC quantitatively show how risky a crash prone traffic condition is. This study proposes a novel approach to predict real time crash risk at signalized intersections at the signal cycle level. The approach uses traffic conflicts extracted from informative vehicle trajectories as an intermediate for crash prediction and develops generalized extreme value (GEV) models based on conflict extremes. Moreover, a Bayesian hierarchical structure is developed for the GEV model to combine conflict extremes of different intersections, and the aim is to further improve safety estimates through borrowing strength from different intersections and accounting for non-stationarity and unobserved heterogeneity in conflict extremes. The proposed approach was applied to four signalized intersections in City of Surrey, British Columbia. Traffic conflicts measured by modified time to collision and three cycle-level traffic parameters (traffic volume, shock wave area, and platoon ratio) were extracted from collected video data using computer vision techniques, and a best fitted model was then developed. Two safety indices, risk of crash (RC) and return level of a cycle (RLC), were derived from the GEV model to quantitatively measure the safety cycle-by-cycle. The results show that the non-negative RC can directly point out cycles with crash prone traffic conditions with RC > 0, and RLC is a more flexible safety index which can differentiate between safety levels even for “safe” cycles with RC = 0. The real time crash prediction results are validated at an aggregate level by comparing to observed crashes.