TinyML‐Enhanced Cloud‐Edge Collaborative Framework for Real‐Time Sport Action Recognition

Advancements in wearable sensor networks and computer vision analytics necessitate real‐time kinematic recognition systems with ultra‐low latency and energy‐efficient edge computing architectures for athletic performance monitoring. However, conventional approaches face critical barriers of high computation and network dynamics, since dynamics naturally cause high computation. Hence, targeting addressing the critical challenges of high computational latency and resource constraints in traditional networks, this paper presents a cloud‐edge collaborative framework for real‐time sports action recognition using the Tiny Machine Learning (TinyML) method. First of all, a tiny machine learning architecture combining pruned MobileNetV3‐Sport with temporal motion encoding is proposed to optimize the cloud‐edge deployment through tensor decomposition. Secondly, the overall system integrates a WebRTC‐based streaming pipeline and accelerated model updates, enabling seamless coordination between cloud‐based federated training and edge inference. Lastly, comprehensive experiments are carried out on the Sports‐5K dataset containing 23 athletic disciplines; our solution achieves 91.5% recognition accuracy with around 30% reduction of latency compared to state‐of‐the‐art baselines. Tests also demonstrate robust performance of the proposed architecture and solution under dynamic network conditions, maintaining less accuracy degradation during abrupt motion transitions. The energy efficiency and response time validate its viability for next‐generation applications, including athlete biomechanical analysis and AI‐powered virtual coaching systems.