An Improved Social Force Model‐Driven Multi‐Agent Generative Adversarial Imitation Learning Framework for Pedestrian Trajectory Prediction

• Published in: Computer animation and virtual worlds Vol. 36; no. 3
• Main Authors: Zhou, Wen, Shen, Wangyu, Meng, Xinyi
• Format: Journal Article
• Language: English
• Published: Hoboken, USA John Wiley & Sons, Inc 01.05.2025; Wiley Subscription Services, Inc
• Subjects: group‐enabled; Heterogeneity; individual heterogeneity; Learning; multi‐agent generative adversarial imitation learning; Pedestrians; social force model; trajectory prediction
• ISSN: 1546-4261; 1546-427X
• DOI: 10.1002/cav.70058

ABSTRACT Recently, crowd trajectory prediction has attracted increasing attention. In particular, the simulation of pedestrian movement in scenarios such as crowd evacuation has gained increasing focus. The social force model is a promising and effective method for predicting the stochastic movement of pedestrians. However, individual heterogeneity, group‐driven cooperation, and poor self‐adaptive environmental interactive capabilities have not been comprehensively considered. This often makes it difficult to reproduce real scenarios. Therefore, a group‐enabled social force model‐driven multi‐agent generative adversarial imitation learning framework, namely, SFMAGAIL, is proposed. Specifically, (1) a group‐enabled individual heterogeneity schema is utilized to obtain related expert trajectories, which are fully incorporated into the desire force and group‐enabled paradigms; (2) A joint policy is used to exploit the connection between the agents and the environment; and (3) To explore the intrinsic features of expert trajectories, an actor–critic‐based multi‐agent adversarial imitation learning framework is presented to generate effective trajectories. Finally, extensive experiments based on 2D and 3D virtual scenarios are conducted to validate our method. The results show that our proposed method is superior to the compared methods. Overview of the proposed framework, which consists of training and test stages. In the training stage, the improved group‐enabled social force model, which incorporates group and individual heterogeneous behavior to describe the overall behavior, is used; subsequently, a joint policy‐based multi‐agent generative imitation method is used to study the intrinsic properties of the proposed social force model. In the test stage, the policy network is used to predict the multi‐agent trajectories by exploring existing learning models.