Dense Crowd Flow Simulation Considering Pedestrian-Venue Relationships

We propose a dense crowd simulation framework that reflects the influence of venue structures on pedestrian flows in crowded environments. Existing studies on crowd simulation and multi-agent trajectory prediction have mainly focused on interactions among pedestrians. However, as crowd density increases, the inherent structure and characteristics of each venue have a significant impact on pedestrian movement directions and changes in crowd density. In this study, such distinctive spatial elements are defined as venue nodes, and semantic information is assigned to each node to model not only pedestrian-pedestrian relationships but also pedestrian-venue and venue-venue relationships. In addition, a graph neural network is used to reflect interaction weights according to the types and characteristics of relationships, enabling the model to effectively learn changes in crowd flow caused by venue structures. The proposed method predicts pedestrian trajectories based on vectorized trajectory representations and analyzes not only position prediction errors but also simulation-oriented metrics such as collision rate and crowd density around venue nodes. Through this approach, this study aims to demonstrate the potential of dense crowd flow modeling that incorporates the semantic meaning and structure of venues.