SocialMOIF: Multi-Order Intention Fusion for Pedestrian Trajectory Prediction

TL;DR

The paper tackles uncertainty in agent intentions and complex social interactions for pedestrian trajectory prediction. It introduces SocialMOIF, a two-layer multi-order intention fusion framework that models both direct (first-order) and indirect (higher-order) neighbor influences, augmented by a trajectory distribution approximator and a global trajectory optimizer for parallel predictions. A distance-direction fused loss and a KL-based regularization jointly supervise dynamic states, yielding state-of-the-art results across ETH/UCY, NBA, SDD, and NuScenes with enhanced interpretability. The approach offers practical impact for autonomous systems by improving prediction accuracy and enabling efficient real-time trajectory planning while clarifying the role of latent distributions in decision-making.

Abstract

The analysis and prediction of agent trajectories are crucial for decision-making processes in intelligent systems, with precise short-term trajectory forecasting being highly significant across a range of applications. Agents and their social interactions have been quantified and modeled by researchers from various perspectives; however, substantial limitations exist in the current work due to the inherent high uncertainty of agent intentions and the complex higher-order influences among neighboring groups. SocialMOIF is proposed to tackle these challenges, concentrating on the higher-order intention interactions among neighboring groups while reinforcing the primary role of first-order intention interactions between neighbors and the target agent. This method develops a multi-order intention fusion model to achieve a more comprehensive understanding of both direct and indirect intention information. Within SocialMOIF, a trajectory distribution approximator is designed to guide the trajectories toward values that align more closely with the actual data, thereby enhancing model interpretability. Furthermore, a global trajectory optimizer is introduced to enable more accurate and efficient parallel predictions. By incorporating a novel loss function that accounts for distance and direction during training, experimental results demonstrate that the model outperforms previous state-of-the-art baselines across multiple metrics in both dynamic and static datasets.

Figures (6)

• Figure 1: The process of SocialMOIF includes: (1) extracting the position, velocity and trajectory information of agents, (2) achieving multi-order fused intention modeling, (3) guiding the distribution of trajectories and performing global optimization. Finally we obtain accurate parallel predictions.
• Figure 2: Computational pipeline of SocialMOIF. Firstly, the first-order interaction layer focuses on direct interactions between the target agent and its neighbors. The higher-order interaction layer addresses the interactions among neighbors. Multi-order intention is obtained by fusing the information from both layers. This multi-order intention is then combined with the actual future trajectory, with the distribution of the trajectory guided by an approximator. An optimizer is subsequently employed to enhance the decoded trajectory. Additionally, a fused distance-direction loss function supervises the training process. Finally, efficient parallel prediction is achieved.
• Figure 3: Representative cases. Row 1 represented complex historical trajectory cases. Row 2 represented state maintenance cases. Row 3 represented numerous neighbors cases.
• Figure 4: Predictions comparison of predictions in ETH/UCY. The comparison included heatmaps of the overall distribution of predictions and best-of-20 predictions of the related methods.
• Figure 5: Distribution heatmaps. Every heatmap of future trajectory distribution was shown above the real scene image. (a) represented the actual distribution. (b) presented the experimental group containing only I component. (c) showed the group with both I and B components.