Research on Real-Time Recognition and Posture Assessment of Tai Chi Cloud Hands Movement Based on Improved YOLOv7

Traditional Tai Chi Cloud Hands teaching is facing the challenges of frequent arm-crossing and large rhythmic variations, which make it hard for traditional methods to stably track joints and assess posture in real time. In this paper, we design an improved YOLOv7 dual task framework with spatiotemporal attention and human prior knowledge. Firstly, a lightweight coordinate attention mechanism is introduced to improve the perception of occlusion on joint. Secondly, a temporally aligned neck design is implemented. Optical flow is used to fuse inter-frame features and suppress jitter. Then, template geometric constraints are introduced. Sag and arc deviation angles are introduced into loss function for quantitative scoring. Finally, reparameterization and simplification of detection head guarantee real-time application. The experimental results show that our model can achieve an average recall rate of 91.7% for occluded joints, the improvement of 19.6% over the baseline in fully occluded situation, parameter amount of 38.3M, running time of 14. 2ms and correlation coefficient between score and judge of 0.882. This model effectively balances computational efficiency, joint localization accuracy, and professional biomechanical evaluation, providing a robust and feasible framework for the digitized instruction of traditional martial arts. The real-time posture framework can be combined with wearable motion sensors in traditional martial arts training.