SFMamba: Social-Intention-Aware Human Trajectory Prediction with Social Force State Space Model

Human trajectory prediction is crucial for understanding human social behavior and motion intentions. Existing studies have often integrated social interaction and intent-driving features to predict changes in motion states, but overlook their varying impacts on specific motion states at different time nodes. In this paper, we propose a novel framework called Social Force Mamba (SFMamba) to tackle these issues for trajectory prediction. This framework constructs a social force state space model by integrating the motion differential equations and the social force model. It represents motion trajectories as transitions from hidden motion states to surface motion states and introduces the Social Mamaba Encoder to extract hidden motion states driven by social interaction and motion intentions. It projects surface motion states into coarse-grained future motion states and combines the Goal Mamba Decoder to refine motion transitions guided by goal intentions. By integrating the selection mechanism to adaptively regulate the injection of physical attributes at different time nodes, our approach models the social-intentionaware state transition process for motion trajectory prediction. Our approach achieves state-of-the-art (SOTA) results on ETHUCY and SportVU NBA datasets while effectively simulating real human movement with physical interpretability.