Deep Learning Improved Social Force Model to Simulate Crowd in Varying Scenes

Traditional social force models are capable of simulating various types of crowds, but due to their reliance on a single formula and simplified parameter control, they often lack realism when applied to crowd scenes. The integration of deep learning with social force models has already achieved success in pedestrian simulation tasks in specific scenarios, yet research in more generalized scenes is still lacking. Therefore, we have used deep learning in conjunction with social force models to construct crowd simulations in various scenarios. This model is based on the dynamics of crowd trajectories and refines key parameters of the social force model using deep learning methods. Experimental results prove that this model can effectively adapt to diverse environments different from the scenes on which the training data is based.

I. Introduction

Crowd simulation is dedicated to simulating and analyzing crowd behavior in various scenarios within virtual environments. This is crucial in fields such as computer games [1] and behavioral analysis [2]. The hybrid models [3], [4] combining deep learning and physics have achieved success in crowd simulations in specific scenarios, yet they still require further validation in more expansive environments. We propose a novel approach that combines deep learning technology with traditional social force model to simulated crowd in different scenes.