Hierarchical game interaction augmenter: Enhancing multi-agent trajectory prediction in competitive and cooperative...

Trajectory prediction in competitive ball games extends beyond geometrically structured environments, with future movements heavily influenced by ball-centric game dynamics. While recent methods enhance spatiotemporal feature extraction, they often treat players and the ball in the same manner, fail to explicitly model the recursive nature of opponent intent, and rely on inflexible group-relation structures that do not accommodate varying interaction strengths or sizes. To address these limitations, this paper introduces the Hierarchical Game Interaction Augmenter (HGIA), a plug-and-play framework that enhances standard spatiotemporal models with three distinct modules. The Ball Possession Module captures control shifts as ball-centric indicators. The Team Intention Module utilizes dual-path recursive reasoning to model strategic exchanges at the team level. The Tactical Cluster Module constructs a learnable, prototype-based soft hypergraph with smooth, adaptive association strengths to represent high-order group relations, avoiding brittle heuristic assignments. Extensive experiments on Basketball-U, Football-U, and Soccer-U datasets demonstrate HGIA's state-of-the-art accuracy and favorable efficiency. The code is available at https://github.com/wsj-neu/HGIA-sports.