EdgeFuser: A tightly coupled adaptive framework for real-time athlete group analytics

This paper presents an integrated framework for real-time athlete group movement analysis in competitive sports, addressing heterogeneous sensor noise, complex multi-agent interactions, and edge-device constraints. The core innovation is a unified state-space model with factor-graph optimization that tightly fuses raw IMU, GPS, and vision data, achieving a mean positioning error of 0.18 m-a 42% improvement over loosely coupled baselines. To overcome computational limitations, we introduce a resource-aware adaptive inference mechanism that dynamically adjusts model complexity based on scene dynamics, reducing latency to 7.8 ms while maintaining over 91% accuracy. For group analysis, a spatiotemporal graph neural network models collaborative and adversarial relationships, attaining 87.4% tactical pattern recognition. Beyond empirical validation, we distill three generalizable design principles: cross-layer Pareto optimality for resource-accuracy trade-offs, context-aware computation frameworks, and semantic graph construction via domain priors. These contributions advance edge-based multi-agent perception systems, extending applicability to autonomous driving and robotic coordination.