Real-time tracking and analysis of martial arts movements based on adaptive 3D detection and quaternions

With the widespread application of artificial intelligence technology, accurate digital analysis of martial arts movements has become one of the efficient development directions for promoting Chinese traditional culture. Since existing attitude estimation models have problems such as detection jitter, depth error, and inaccurate joint rotation expression in complex scenes, this study proposes a real-time tracking analysis method that integrates adaptive 3D detection and quaternion local coordinate system transformation. This method first enhances the stability of MediaPipe through four strategies: speed threshold correction, human body tilt statistics, limb proportion adjustment, and hybrid filtering, and builds an adaptive 3D detection module. Afterwards, this study designs a local coordinate system canonical transformation model based on quaternions, which combined with the physiological constraints of human joint rotation to achieve high-precision motion capture. Experiments showed that the average accuracy of the proposed hybrid filtering method reaches 88%, and the average joint error is 64.4 px. The proposed technology has an overall recognition rate of 95% for the five basic martial arts stances. The position error of the motion capture method is reduced to 49.5 px, the motion similarity is increased to 93.17%, and the overall accuracy rate in 7 leg method recognition reaches 89.3%. This method can not only ensure high-precision martial arts movement recognition, but also provide better stability. This provides a better technical tool for future martial arts teaching and professional training. Trail registration 4102026-0363