Abstract: In the context of intelligent sports and educational informatization,fine-grained human action recognition has become a key technology in physical education and training assessment.To address the limitations of traditional methods in utilizing multi-modal information and capturing spatio-temporal structures in complex motion scenarios,this paper proposes a multi-modal graph convolutional network model that fuses skeleton data and wearable sensor information.Firstly,it proposes a fusion method based on “virtual sensors,” which maps wearable sensor signals onto a spatio-temporal graph constructed from skeletal joints,enabling effective integration of multimodal information and enhancing fine-grained motion modeling and cross-modal semantic consistency.Secondly,it designs a multi-layer graph convolutional network tailored for complex sports movements,incorporating local body part segmentation to improve recognition performance in challenging scenarios.Thirdly,it constructs a high-quality multimodal dataset for fencing,covering various technical actions and skill levels,to support fine-grained action recognition and skill assessment.Experimental results on both this dataset and several public benchmarks demonstrate that the proposed method outperforms existing approaches in both action recognition accuracy and skill level classification.This work provides a novel mode-ling framework and technical support for intelligent recognition and evaluation in sports education.

Key words: Action recognition, Multimodal data fusion, Graph convolutional network, Physical education, Fencing dataset