Learning crowd behavior from real data: A residual network method for crowd simulation

Traditional methods of crowd evacuation simulation reduce the visual realism of crowd motion modeling due to hypothetical scenarios and rules. Data-driven methods are an effective way to enhance the visual realism of crowd simulation. However, the existing work mainly focuses on training models for specific scenarios and applies them to the same scenario, so it lacks consideration of scene adaptability. To address this problem, we present a residual network based scene-independent crowd simulation (ResNet-SICS) framework to simulate crowd motion. First, a data-driven crowd properties quantization (DCPQ) method is proposed. This method divides crowd properties into physical properties and social properties: the physical properties (such as the position and the velocity) are extracted from a large number of real videos, and the social properties (such as local cohesiveness and global collectiveness) are then quantified according to the physical properties of crowds. Second, a residual network for crowd behavior properties learning (ResNet-CBPL) motion prediction model is established. This model takes the crowd properties as parameters to construct the residual network, and uses real data to learn the movement rules of the crowd. The ResNet-CBPL motion prediction model can fit the movement behavior of the crowd more accurately than can other models. Finally, we implement a crowd simulation system based on ResNet-CBPL to visualize the results of theoretical analysis in a graphical way. The experimental results show that the proposed method can simulate the crowd motion more realistically, and the trained crowd simulation framework can be applied to various scenarios.