Data-Driven Approach for Prediction of Drivers' Decision in Type-II Dilemma at Signalized Intersection

During the amber phase at signalized intersections, many drivers often face a dilemma when deciding whether to stop or go. This indecisiveness has significant safety issues; hence, predicting a driver's decision, evaluating policies that supplement the driver's decision-making process, and mitigating the dilemma zones at signal-controlled intersections are crucial. The present study develops multiple dilemma decision prediction models using statistical and data-driven machine learning (ML) approaches. Also, the models are developed while focusing on the internal validation and external transferability for in-field application to aid real-time performance. Several vehicular, geometric, and signal operational parameters derived from vehicular trajectory data from four study locations are used for the analysis. The modeling approach yields good performance results in predicting a driver's decision during the amber phase. ML-based models are observed to yield better performance. Further, the statistical method provides statistically significant coefficients and corresponding elasticity values for several parameters used for evaluating and visualizing the effect of varying the parameter value on the location of the dilemma zone. These models are beneficial in assessing the effect of changes in operational policies. Whereas ML-based models yield the advantage of higher prediction accuracy, faster and robust predictions, and inherent quality of better understanding of the complexity in data, these models are more suited for real-time operation, driver assistance, and signal optimizations. The SHapley Additive exPlanations (SHAP) values that support measuring the effect of an individual parameter on prediction performance of the ML model are also studied.