Edge‐Assisted Multimodal Athletic Performance Analysis Optimization for Smart Sports

ABSTRACT The rapid integration of the Internet of Things (IoT) and Artificial Intelligence (AI) into competitive sports has catalyzed a paradigm shift towards data‐driven athletic training. However, existing solutions for athletic performance analysis face a critical Accuracy‐Latency dilemma. Cloud‐centric approaches suffer from transmission delays that prohibit real‐time feedback, while lightweight edge models often lack the capacity to model complex, fine‐grained kinematic chains. Furthermore, traditional vision‐only systems fail to capture the internal physiological states (e.g., fatigue) that precipitate injury. To bridge these gaps, this paper proposes an Edge‐Assisted Spatiotemporal Graph Fusion (E‐STGF) framework optimized for smart sports environments. We introduce a novel Skeleton‐Adaptive Graph Convolutional Layer (SA‐GCL) that dynamically learns the functional topology of the human body, surpassing the limitations of fixed‐graph structures. To capture multi‐scale movement patterns, from explosive impulses to gradual fatigue, we employ a Multi‐Scale Temporal Convolutional Network (MS‐TCN) fused with physiological signals via a cross‐modal attention mechanism. Crucially, to enable real‐time deployment on resource‐constrained edge devices, we implement a Knowledge Distillation (KD) strategy that transfers the dark knowledge of a complex teacher network to a streamlined student model. Experimental results on the UTD‐MHAD and a self‐collected badminton dataset demonstrate that E‐STGF achieves 98.1% recognition accuracy with an end‐to‐end latency of 41.5 ms, significantly outperforming state‐of‐the‐art methods. This framework provides a scalable, low‐latency solution for proactive injury prevention and precision coaching in the next generation of smart stadiums.