Action recognition technology based on feature comparison in Wushu web teaching video

To improve students’ learning efficiency in online martial arts teaching and correct their erroneous movements, a Kinect device is selected to collect relevant data and construct a three-dimensional skeleton skin model of the character. The Dual Quaternion Linear Blending (DLB) algorithm is employed to enhance the realism of the model’s motion. The improved median-mean filtering algorithm and the improved Exponential Weighted Moving Average (EWMA) algorithm are utilized to detect and recognize human motion. These techniques are integrated into a martial arts learning assistance system. The results showed that compared with the standard median-mean filtering algorithm, the improved median-mean filtering algorithm had better filtering effect and more sensitive processing nodes in both martial arts motion forms. Its stability rate and real-time rate were 95.53% and 97.05%, respectively, with an improvement of 8.49% and 26.54%, indicating that the improved algorithm has better performance. In comparison to the standard EWMA algorithm, the enhanced EWMA algorithm demonstrated favorable performance in both scenarios involving angle alteration. This resulted in a reduction in computational error and delay, as well as an improvement in both the smoothing rate and real-time rate, with an increase of 53.02% and 70.42%, respectively. In the two-player mode, the CPU resource usage was reduced by 34.5% and the GPU resource usage was increased by 29.5% by updating the vertices and rendering through GPU. The online teaching method of martial arts adopted in this study can effectively help students learn martial arts and lay a solid foundation for technological innovation in the field of physical education teaching.