What Determines a Step: Observations from Pedestrian Movement Evaluation at Intersections with Machine Learning Methods

Pedestrian flows at multi-directional crossings have become a focal point of research due to their implications for traffic efficiency and public safety in increasingly crowded urban areas. However, accurately predicting pedestrian movement in these scenarios remains challenging due to inherent complexities. This study addresses this issue by developing interpretable machine learning models to predict pedestrian movements in multi-directional flows, using data from experimental intersection studies at varying densities. Our results indicate that adopting a destination-oriented perspective significantly enhances prediction accuracy, yielding lower displacement errors in the Cartesian coordinate system compared to the polar coordinate system. The analysis reveals that forward displacement is primarily influenced by a pedestrian's current state, while lateral displacement is more affected by nearby pedestrians. Furthermore, our findings emphasize the importance of considering both the relative positions and velocities of surrounding pedestrians in different directions, rather than the Social Forces they exert. These insights suggest that our data-driven models not only advance the theoretical understanding of pedestrian movement dynamics but also can more accurately capture pedestrian dynamics in predictions. This improved accuracy, whether as the core of pedestrian simulation software or integrated with intelligent detection systems, can inform the design of safer public spaces, optimize traffic management, and enhance the planning of large events, thereby contributing to more efficient and resilient urban transportation networks.