Quantitative Assessment of Lane-Changing Risk for Nonmotorized Vehicles Based on YOLOv8s and Nonstationary Extreme Value Theory

Abstract

To evaluate the dynamic risks of nonmotorized vehicle lane-changing, this study develops an analytical framework based on extreme value theory. Vehicle trajectories were extracted from unmanned aerial vehicle footage using the YOLOv8s (You Only Look Once) object detection model and the ByteTrack multiobject tracking algorithm. The minimum time-to-collision ( TTC min ) and postencroachment time ( PET min ) were modeled as response variables, with key riding behavior variables serving as covariates. Through a threshold selection method integrating mean residual life plots, threshold stability plots, and Akaike information criterion (AIC) minimization, data-driven safety thresholds were established as 0.73 s for TTC min and 0.54 s for PET min . To capture risk dynamics, a nonstationary generalized Pareto distribution (GPD) model was developed, with its scale parameter linked to the covariates. The results indicate that the nonstationary GPD models with covariates significantly outperform the stationary baseline. The optimal model identifies maximum yaw rate and minimum longitudinal distance as key covariates for TTC min , while minimum longitudinal distance is also significant for PET min . These findings provide safety thresholds and a dynamic modeling method for risk assessment, offering an analytical basis to support safety-related applications in intelligent transportation systems, such as the calibration of collision avoidance systems and real-time risk monitoring.