Dynamic Group-Aware Graph Generative Model for Socially Coherent Pedestrian Trajectory Prediction

Accurately forecasting pedestrian trajectories in crowded environments requires understanding not only individual motion patterns but also the complex group and social interactions that shape collective behavior. In this article, we introduce the group-aware Spatial-Temporal Graph and Generative Model, designed to predict socially coherent and multimodal future pedestrian paths. Our approach begins by dynamically detecting pedestrian groups using an interpretable threshold-based mechanism that jointly considers spatial proximity, velocity similarity, and directional alignment. This enables real-time identification of cohesive groups based on observed dynamics rather than heuristic clustering. We construct a unified hybrid graph that incorporates both pedestrian and group nodes, where edges encode intra-group consistency and inter-group conflicts, continuously updated via a dynamic adjacency matrix. An enriched spatial-temporal attention mechanism further enhances interaction modelling by explicitly embedding behavioral cues such as speed differences, orientation similarity, and conflict scores. For trajectory prediction, we propose a lightweight group-aware generative decoder based on a conditional variational autoencoder with Gaussian mixture model output. This design enables diverse yet interpretable trajectory generation, conditioned on both individual history and rich group-level contextual features. Extensive experiments conducted on standard pedestrian benchmarks (ETH and UCY) demonstrate that our model achieves state-of-the-art performance in both average displacement error and final displacement error. Ablation studies confirm the effectiveness of each model component, while qualitative analyses highlight the model’s ability to capture group behavior, social cohesion, and realistic multimodal futures. Our results validate the importance of dynamic group-awareness and structured interaction modelling in advancing the state of pedestrian trajectory forecasting.