Safety evaluation for heavy vehicle drivers using extreme value model based on the multi-source sensing data

Abstract

Increasingly, the collection of multi-source sensing data from heavy vehicles through intelligent networked platforms has become prevalent for safety management and supervision. However, practical approaches for crash risk management and safety evaluation have not been fully developed to capitalize on this high-value driving data. This study proposed a safety evaluation framework for heavy vehicles using the extreme value modeling approach. First, univariate extreme value models were developed to determine the thresholds of crash risk for two kinematic indicators under different loading conditions. Then, various bivariate logistic-based extreme value models were developed to analyze the dependence structure between the two kinematic indicators, construct probability-based crash risks, and estimate them according to the thresholds. The univariate and bivariate block maxima models were applied to the dataset containing 3,452 trips from 64 heavy vehicles recorded in Hangzhou, China. The results show that the Speed time-varying stochastic volatility (Speed-$V_f$) and Jerk are effective indicators for assessing the driving risks of heavy vehicles. Meanwhile, unloaded conditions and extremely high distraction and fatigue warning frequencies are identified as trip-level factors contributing to the crash risk of heavy vehicles. The optimal thresholds of 1.49 and 1.47 for Speed-$V_f$ and 1.23 ${m/s}^3$ and 1.25 ${m/s}^3$ for Jerk under two loading conditions, respectively, were identified for crash estimation. Additionally, the bivariate logistic-based models can effectively capture dependency structures and provide robust crash risk estimations, outperforming their univariate counterparts. Overall, this study demonstrates a safety evaluation framework for heavy vehicles that includes determining crash estimation thresholds under different driving tasks, analyzing the joint probabilities of crashes to model dependence between indicators, and selecting the best safety evaluation model.