This is the authors' abstract. We don't add key points for this paper.
To address the lack of objective real-time evaluation and corrective feedback in traditional martial arts training, this study designs a real-time movement evaluation and error correction system based on a lightweight convolutional neural network and human keypoint detection. The system uses monocular RGB video as input, extracts human keypoints through a compact lightweight network, and constructs normalized kinematic descriptors including joint angles, limb-length ratios, centroid stability, terminal acceleration, and rhythm features. Standard martial arts templates are built from expert-labeled demonstrations, and dynamic time warping is used on compact kinematic sequences to locate posture deviation and timing errors. Experiments on a self-built 17-class traditional martial arts dataset show that the system achieves 96.2% PCK@0.05 keypoint detection accuracy, 94.8% action recognition accuracy, and 35 ms average single-frame processing time on ordinary hardware. A feedback validation with 24 learners shows that the average expert-rated movement quality score increases from 72.4 to 84.6 after using the correction module. The proposed system provides a feasible technical tool for low-cost, real-time, and interpretable martial arts training assistance.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Ru Li (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: