Collision risk identification and prediction considering heterogeneous braking patterns using large-scale pre-collision trajectories

Rear-end collisions often occur in vehicles when successive braking events lead to insufficient deceleration by one or more vehicles. This paper proposed a novel method for collision risk identification and prediction by integrating the braking dynamics of both leading and following vehicles in pre-crash scenarios. Using a piecewise linear model of deceleration profiles, 45 collision risk moments were identified across 10 scenarios based on ten kinematic parameters. A novel critical time-to-collision metric was proposed to integrate both the timing and execution of braking behaviors into collision risk prediction. To account for driver heterogeneity in deceleration, deceleration rate, and reaction time, Gaussian mixture regression was used to perform conditional inference to estimate braking pattern parameters and generate interval-valued crash risk predictions. The performance and optimal threshold were validated using large-scale vehicle trajectories from collision and non-collision events. The results demonstrate that the collision risk moments vary with both the braking timing and execution of the leading and following vehicles. The proposed metric outperformed traditional surrogate safety measures in predicting collision risk, consistently yielding higher accuracy with less variability across pre-crash time intervals. These findings indicate that the estimated critical time-to-collision is a reliable and effective measure for collision avoidance systems and advanced driver assistance technologies.