Prediction of Vehicular Yielding Intention While Approaching a Pedestrian Crosswalk

Abstract

This study proposes a novel approach for predicting vehicular yielding intention at pedestrian crosswalks, leveraging transformer-based deep-learning and multilayer perceptron for time-series data representation and classification, respectively. The model was trained using parameters obtained from real-world video feeds, which encompassed vehicle dynamics, pedestrian factors, and traffic context. The study compared the transformer–multilayer perceptron model with other models, such as recurrent neural network, long short-term memory, and gated recurrent unit, in predicting vehicular yielding intention. The transformer model achieved a superior performance with an accuracy of 94.47% compared with other models. This result was statistically validated using the Friedman test. The study used a 2-s prediction horizon, which improved its practical usefulness by giving pedestrians enough time to react to avoid dangerous situations. The study utilized attention scores from the transformer model to identify significant factors influencing yielding decisions. The analysis revealed that the number of visible pedestrians, time remaining to reach the crosswalk, and the behaviors of the preceding vehicle were key determinants of yielding decisions. This model has potential applications in developing advanced warning systems, integrating vehicle-to-pedestrian technology, and providing insights for traffic management policies and road infrastructure design. Future research could explore the integration of human factors and demographic information into the current model.