Observation from Crossing Pedestrian Flow with Data-Driven Movement Prediction

Data-driven models are considered a promising solution for predicting pedestrian movements in complex, multi-directional flow scenarios. However, the question of what feature extraction methods, particularly in terms of observation viewpoints and coordinate systems, are critical to training accurate models remains a critical yet underexplored issue. This study aims to enhance data-driven pedestrian movement prediction using experimental trajectory data from a controlled 90-degree crossing flow experiment from the Pedestrian Dynamics Data Archive. We constructed multiple prediction forms with different datasets consisting of input features and output variables extracted under combinations of different observation viewpoints, coordinate systems, and the inclusion or exclusion of Social Force (SF). By training LightGBMs and comparing the prediction displacement errors from models trained with features, we found that the form with the destination-oriented viewpoint, which uses the pedestrian’s intended destination as the reference axis for feature extraction, under the Cartesian coordinate system was proved to outperform other forms, while SF-related features minimally impacted accuracy. The findings offer practical guidance for developing accurate crowd movement prediction models which can further assist architectural designer, event organizers, and urban planners in making more informed decisions to enhance the safety and efficiency of complex pedestrian flows.