A Spatiotemporal Graph Transformer Network for real-time ball trajectory monitoring and prediction in dynamic sports environments

Ball trajectory prediction in sports is a crucial task for real-time applications such as performance analytics and coaching systems. Existing methods often struggle to simultaneously capture the spatial relationships between players and the ball while modeling long-term temporal dependencies. To address these limitations, we propose the Spatiotemporal Graph Transformer Network (SGTN), which integrates a Spatiotemporal Graph Convolutional Network (ST-GCN) with a Transformer architecture. This combination allows for precise modeling of dynamic sports scenarios by effectively handling both spatial and temporal data. Experimental results on the SoccerNet-v2 dataset demonstrate that SGTN significantly outperforms recent models in terms of prediction accuracy, trajectory coverage, and robustness, while maintaining competitive inference times suitable for real-time applications. SGTN achieves a significant 25% reduction in prediction error compared to existing models. This improvement underscores its enhanced accuracy, robustness, and real-time feasibility, making it highly suitable for applications such as live sports analysis, performance monitoring, and decision-making. The model’s ability to generalize across multi-scenario environments further demonstrates its potential applicability to other fields beyond sports, including autonomous systems and robotics. Our future work aims to enhance memory efficiency and optimize multi-scenario inference speed to broaden the model’s deployment in edge computing environments.