An ensemble time-embedded transformer model for traffic conflict prediction at RRFB pedestrian crossings

• Analyzed vehicle-to-vehicle (v2v) rear-end conflicts at RRFB signals. • Collected video data and utilized computer vision algorithms for analysis. • Applied the Time-Embedded Transformer model for conflict prediction. • Combined models using a voting classifier, improving recall and reducing FAR. • Insights aim to optimize RRFB performance, enhance driver alerts, and improve traffic flow reliability. Previous studies have addressed the pedestrian-vehicle conflicts at Rectangular Rapid Flashing Beacons (RRFB) signals. However, none have explored vehicle-to-vehicle (v2v) conflicts, leaving a significant gap in the current research. This study fills this gap by analyzing v2v rear-end conflicts at RRFB signals. Fifty-two hours of video data were collected using portable CCTV cameras and analyzed using computer vision algorithms. A bounding box system was employed to predict vehicle conflict points and collision pairs. Modified Time to Collision (MTTC) and Deceleration Rate to Avoid Collision (DRAC) were taken individually as conflict indicators, and peak-over-threshold method was employed to determine the threshold ranges for conflict analysis. The comprehensive feature selection process ensures the identification of the most significant predictors. The average speed of vehicles, standard deviation of speed, and traffic volume in each lane, in addition to time of day, signal activation, pedestrian count, and pedestrian signal compliance rate were identified as important features. Initially, the Time-Embedded Transformer model was applied separately to MTTC and DRAC. The thresholds of ≤1.5 seconds as conflict for MTTC and ≥4 m/s² as conflict for DRAC produced the best results, achieving recall 82% and 83%, and False Alarm Rate (FAR) 27% and 38%, respectively. For improved performance, these models were then combined using a voting classifier technique. This combination led to significant improvements, achieving a recall of 86% and FAR of 23%. Insights from the study can optimize RRFB performance by identifying conflict times and locations, improving driver alerts during peak hours, and ensuring smoother traffic flow and reliability.