Citation

Zhai B, Lu J, Wang Y, Wu B. Int. J. Transp. Sci. Technol. 2020; 9(4): 287-298.

Copyright

(Copyright © 2020, Elsevier Publishing)

DOI

10.1016/j.ijtst.2020.02.001

PMID

unavailable

Abstract

The study aimed to develop a real-time crash risk prediction model on freeways under fog conditions. The data used include traffic surveillance data, fog-related crashes data, road geometry data, and visibility data of Interstate-5 (I-5), Interstate-10 (I-10), Interstate-15 (I-15) and Interstate-405 (I-405) in California, United States. The random forests method was applied for variable selection to identify and rank the most important variables. And then, the Bayesian logistic regression model was employed to develop the real-time crash risk prediction model. The model estimation results show that the explanatory variables contributing to crash risk are different in different time slices before crashes. There are common features: (a) visibility is negatively-correlated with the real-time crash risk under fog condition, and (b) average and standard deviation of vehicle count at upstream station are positively-correlated with crash risk. The model of time slice 3 (interval between 10 to 15 min prior to a crash time) performs best with the lowest false alarm rate and the highest overall accuracy and the largest area under the receiver operating characteristic (ROC) curve. And it can identify over 72% of fog-related crashes with the pre-specified threshold of 0.2.

Language: en

Keywords

Bayesian logistic regression; Fog conditions; Freeway; Random forests method; Real-time crash risk