A crash prediction method based on bivariate extreme value theory and video-based vehicle trajectory data

Traditional statistical crash prediction models oftentimes suffer from poor data quality and require large amount of historical data. In this paper, we propose a crash prediction method based on a bivariate extreme value theory (EVT) framework, considering both drivers' perception-reaction failure and failure to proper evasive actions. An unmanned aerial vehicle (UAV) was utilized to collect videos of ten intersections in Fengxian, China, at representative time periods. High-resolution vehicle trajectory data were extracted by a Kanade-Lucas-Tomasi (KLT) technique, based on four detailed metrics were derived including Time-to-accident (TA), Post-encroachment Time (PET), minimum Time-to-collision (mTTC), and Maximum Deceleration Rate (MaxD). TA was expected to capture the chance of perception-reaction failure, while other three metrics were used to measure the probability of failure to proper evasive actions. Univariate EVT models were applied to obtain marginal crash probability based on each metric. Bivariate EVT models were developed to obtain joint crash probability based on three pairs: TA and mTTC, TA and PET, and TA and MaxD. Thus, union crash probability within observation periods can be derived and the annual crash frequency of each intersection was predicted. The predictions were compared to actual annual crash frequencies, using multiple tests. The findings are three-folds: 1. The best conflict metrics for angle and rear-end crash predictions were different; 2. Bivariate EVT models were found to be superior to univariate models, regarding both angle and rear-end crash predictions; 3. TA appeared to be an important conflict metric that should be considered in a bivariate EVT model framework. In general, the proposed method can be considered as a promising tool for safety evaluation, when crash data are limited.