Real-time crash prediction for a long low-traffic volume corridor using corrected-impurity importance and semi-parametric generalized additive model
Arash Khoda Bakhshi, Mohamed M. Ahmed
Abstract
Real-time risk assessment studies have investigated a limited length of corridors. However, the necessity of assessing the safety performance of Connected Vehicles (CVs) requires looking into an entire corridor. Aligned with the CV Pilot Program on 402-miles Interstate-80 in Wyoming, this study serves as a baseline to quantify the safety performance of the corridor during CV pre-deployment. Real-time traffic-related predictors were characterized to capture the spatial variation in traffic characteristics, both longitudinally and laterally. Nine Crash Prediction Models (CPMs) were conducted following the matched-case control design within two main parts. First, important predictors were detected using three feature selection techniques; Corrected-Impurity Importance (CII), Mean Decrease Impurity, and Mean Decrease Accuracy. Secondly, for each of the three sets of selected features, three different Logistic Regression models were developed; the Generalized Additive Model (GAM), Generalized Linear Model, and Generalized Nonlinear Model. The combined GAM and CII outperformed other CPMs by obtaining minimum error, maximum prediction performance, and detecting a larger number of significant predictors, which would enhance the safety performance measurement of the few numbers of CVs by comparing CVs pre- to post-deployment. Findings showed that investigating individual lanes is beneficial to comprehend crash patterns on corridors with comparatively less traffic volume.