An intelligent virtual coaching system for traditional martial arts based on 3D pose estimation and generative adversarial networks

Traditional martial arts are hard to hand on. The masters who hold the repertoire are ageing, instruction rarely travels beyond the training hall, and verbal correction remains stubbornly subjective. This paper presents an intelligent virtual coaching system that couples enhanced three-dimensional pose estimation with generative adversarial network refinement. A modified High-Resolution Network, augmented with attention and temporal convolution, extracts skeletons that survive fast strikes and self-occlusion, while a discriminator-guided sequence optimization module pulls noisy trajectories back onto the manifold of biomechanically plausible movement. On top of these components, a hierarchical feedback mechanism turns quantitative learner-reference comparisons into real-time, multimodal corrections. Evaluated on public benchmarks and on a purpose-built martial arts corpus, the pipeline reaches 38.6 mm Mean Per Joint Position Error while sustaining 28.4 frames per second, and its automated quality scores track expert ratings closely, with a correlation coefficient of 0.847. Taken together, the results describe a workable technological pathway for preserving intangible cultural heritage and for widening access to high-quality martial arts instruction.