A Tai Chi movement recognition framework integrating local–global temporal dynamics and spatial structure features

As an important representative of traditional Chinese martial arts, Taijiquan (Tai Chi) contains continuous transitions, subtle inter-class differences, and coordinated whole-body movements, which make skeleton-based movement recognition challenging. In this paper, we propose TaichiNet, a skeleton-based Taijiquan movement recognition framework that integrates local spatio-temporal modeling and global temporal dependency learning. Specifically, we design a Time-Space Graph Unit (TS-GU), a residual recurrent graph block that extracts frame-level spatial dependencies through graph convolution, reweights the resulting node features through feature-sensor attention, and carries local motion information across adjacent frames. The distinguishing design choice is that TaichiNet explicitly forms a graph-conditioned, node-wise recurrent state across adjacent frames before the sequence enters the Transformer. The TS-GU output sequence is then fed into a Transformer encoder to capture dependencies over the whole movement segment. Extensive evaluations on UMONS-TAICHI, including random-split and participant-disjoint settings, show that TaichiNet consistently outperforms the comparison methods across the reported classification criteria. An adapted segment-level evaluation on NTU RGB+D 120 tests whether this recurrent local-state-before-global-attention interface remains effective on a larger and more diverse action benchmark. The evaluation also includes controlled architectural comparisons, component ablations, motion-pattern analysis, and reproducibility checks.