Action recognition for Tai Chi with STGCN-MambaFormer
Kemin Chen, Xinyuan Tian, Xiangshuang Chen · 2025
Tai Chi is a traditional Chinese martial art. It is not only practiced globally but also widely recognized as a health-promoting physical activity. Accurate recognition of Tai Chi actions is crucial for teaching, performance evaluation, and the development of virtual coaching systems. However, the complexity of Tai Chi movements, their continuity, and individual differences among performers pose significant challenges to precise action recognition. To address these issues, a Spatio-Temporal Graph Convolutional Network (STGCN) and MambaFormer based Tai Chi action recognition method is proposed in this paper, termed STGCN-MambaFormer. Initially, we extract human keypoints sequences from videos with Openpose and input these sequences into the proposed STGCN-MambaFormer for recognition. In STGCN-MambaFormer, STGCN serves as an encoder, encoding and modeling the spatial and temporal correlations of human keypoints through spatial and temporal graph convolutions. Subsequently, the encoded spatio-temporal features are fed into MambaFormer to model the long-short term dependencies globally and obtain the predicted action output. We constructed an evaluation dataset from numerous online Tai Chi videos and performed extensive comparative experiments and ablation studies. The experimental results demonstrate that STGCN-MambaFormer outperforms all baseline models, validating the effectiveness of the proposed method.