Autonomous vehicle sensor data and the estimation of network-wide spatiotemporal generalized extreme value models of rear-end injury-severity crash frequencies

Abstract

Existing traffic conflict-based extreme value modeling applications are primarily restricted to a few concentrated locations due to the scarcity of network-wide vehicular trajectory data and the constraints associated with traditional network-wide modeling techniques. As such, this study develops a network-wide bivariate spatiotemporal non-stationarity generalized extreme value model to estimate rear-end crash frequency by injury severity level using Argo AI autonomous vehicle sensor data. Fusing this dataset with road network data from the Florida Department of Transportation, this paper studies a road network of 57 intersections and mid-blocks in Miami, Florida. Modified time-to-collision and the expected post-collision velocity difference (Delta-V) are used to estimate severe and non-severe rear-end crashes. Road geometry, road classification, and traffic state variables are used as covariates to address spatiotemporal heterogeneity in the generalized extreme value model estimation. Results show the significant impact of spatiotemporal variables such as lane width, median width, dedicated street parking, dedicated bike lane, vehicle class, and road class on rear-end crash frequency by injury severity levels. It is found that the bivariate spatiotemporal generalized extreme value model outperforms the bivariate random intercept generalized extreme value model and the univariate generalized extreme value model with conditional severity probability when benchmarked against observed annual crash frequency using root mean square error and the coefficient of determination (R-squared). Additionally, the bivariate spatiotemporal generalized extreme value model provides the closest estimate of observed severe crashes by roadway segments in the study area. The findings of this study underscore the importance of proactive network-wide safety management using spatiotemporal heterogeneity and autonomous vehicle sensor data to estimate crash frequency by severity for real-time decision-making.