Evaluating the performance of traffic conflict measures in real-time crash risk prediction using pre-crash vehicle trajectories.

The primary objective of this study was to evaluate the performance of traffic conflict measures for real-time crash risk prediction. Drone recordings were collected from a freeway section in Nanjing, China, over a year. Twenty rear-end crashes and their associated trajectories were obtained. Vehicle trajectories preceding the crash were segmented based on different time periods to represent varying crash conditions. The Extreme Value Theory (EVT) approach combined with a block maxima sampling method was then employed to investigate the generalized extreme value (GEV) distributions of extremely risky events under non-crash and crash conditions. The prediction performance was demonstrated by the differences in GEV distributions under these two conditions. Within the proposed modeling framework, the performances of Time-to-Collision (TTC), Deceleration Rate to Avoid a Crash (DRAC), and Absolute value of Derivative of Instantaneous Acceleration (ADIA) were examined and compared. The results revealed a decreasing trend in the prediction performances as the preceding time window before a crash increased. For any given length of crash conditions, TTC consistently outperformed DRAC and ADIA. Notably, TTC's reliability in crash risk prediction became more uncertain when forecasting crashes more than 2 s in advance. This study provided the optimal thresholds for TTC and ADIA for practical application in crash early warning. The methods and results in this study have the potential to be used for crash risk assessments in autonomous vehicles.