Study on Normativity Evaluation of 24-Form Tai Chi Movements Based on Joint Point Spatiotemporal Features and Lightweight Machine Learning

To address the issues of low efficiency and high subjectivity in the normativity evaluation of 24 -form Tai Chi movements, this study proposes an automated evaluation method that integrates computer vision and lightweight machine learning. The Open Pose algorithm is used to extract 18 skeletal joint coordinates from standardized 24-form Tai Chi instructional videos, constructing a multidimensional spatiotemporal descriptor that includes joint angles, movement speed, and temporal transition features. An improved random forest classifier is employed to achieve movement quality grading. The experimental data covers standard videos of “24-form Tai Chi” and synthesized error samples with three types of deviations: translation, angle, and rhythm. The results show that the model's recognition accuracy for standard movements reaches 91.2 % (synthesized data) and 85.3 % (real videos), with an average processing time of 120 ms per Single Window, significantly faster than traditional LSTM models. The system can real-time annotate typical errors such as joint hyper extension and arm angle deviation, with an error tolerance controlled within $\pm 30^{\circ}$, meeting the needs of actual teaching scenarios. This study provides a low-cost, scalable technical solution for Tai Chi teaching and exercise, supports realtime feedback on mobile devices, and has high application value.