New Framework Improves Trajectory Prediction with Staged Social Modeling

The INTraJ framework addresses a limitation in existing trajectory prediction models by explicitly separating how social influence impacts an agent's motion. It posits that agents first form a general plan based on anticipating others' future actions, and then make local, reactive adjustments. This two-stage decomposition, where planning precedes reaction, allows for a more nuanced understanding of social interactions in dynamic environments. The framework constructs initial reference trajectories using future social information in a planning stage. Subsequently, a reaction stage refines these trajectories by accounting for local adjustments, derived from the difference between a full-context prediction and the initial reference. This structured approach is applicable to both single and multi-target prediction scenarios. Evaluations across four major benchmarks, including Argoverse 2 and ETH/UCY, show that INTraJ consistently improves performance, particularly in metrics like Final Displacement Error (FDE) and maintaining consistency over longer prediction horizons. The findings suggest that explicitly modeling social interactions in a staged manner is crucial for achieving stable and accurate trajectory predictions.