Modeling of pedestrian crash frequency at unsignalized intersections using traffic conflict indicators

This study develops a predictive model to estimate pedestrian crash frequency at unsignalized intersections by analyzing pedestrian-vehicle conflicts using video data from Cheema Chowk and Duke Chowk in Ludhiana, Punjab. Utilizing AI techniques like YOLO and Deep SORT, key conflict indicators such as post-encroachment time (PET) and time to collision (TTC) were extracted. These indicators informed the construction of extreme value theory (EVT) models, with bivariate Copula models outperforming univariate ones in predictive accuracy, evidenced by low mean absolute percentage error (MAPE) values of 2.164% and 0.60%, respectively. The study demonstrates the efficacy of bivariate Copula models in forecasting pedestrian crashes, providing a proactive tool for enhancing road safety at unsignalized intersections by identifying high-risk areas and facilitating targeted safety interventions. This approach highlights the importance of incorporating multiple conflict indicators to improve predictive accuracy and reduce pedestrian casualties.