Competitive physical action recognition via inertial sensing and graph convolutional transformer

Based on wearable-sensing internet of things technology, this paper proposes a complete solution encompassing data acquisition, processing, and intelligent recognition for the identification of athletic physical movements. A wireless inertial measurement unit sensor network is deployed across key body segments to capture motion data synchronously. After preprocessing, including low-pass filtering and sensor calibration, a spatio-temporal graph structure is constructed and fed into an innovative graph convolutional-transformer fusion model. This architecture fully leverages graph convolution to extract spatial correlation features from multiple sensors, while the transformer component effectively captures long-range temporal dependencies within movement sequences. Experimental results on the publicly available University of California, Irvine Human Activity Recognition dataset demonstrate that our method achieves 94.2% recognition accuracy. This performance confirms its practical value in wearable-sensing internet of things systems and provides an effective technological pathway for advancing smart sports training platforms.