Author(s)

Jiawen Zhang

Corresponding Author

Jiawen Zhang

Abstract

Human pose estimation, as a significant research direction in computer vision, has broad applications in areas such as motion analysis, intelligent surveillance, and human-computer interaction. In particular, accurate pose estimation plays a crucial role in martial arts motion analysis, providing data support for technical evaluation and serving as a scientific basis for action correction during training and competitions. However, existing Convolutional Neural Network (CNN)-based methods often struggle in handling complex movements and fast actions due to challenges such as occlusion, motion blur, and pose similarity. On the other hand, Graph Neural Network (GNN)-based approaches lack the capability to model temporal dynamics in video sequences. To address these issues, this paper proposes a hybrid pose estimation method combining CNN and GNN for martial arts movements. The approach leverages CNNs to extract local spatial features, while integrating GNNs to model topological relationships among human keypoints, thereby improving the accuracy of pose prediction in complex action scenarios. We conduct experiments on a martial arts action dataset, and the results show that, compared with conventional CNN- or GNN-only methods, our approach improves keypoint localization accuracy by 6.4%, enhances robustness in complex pose detection by 18%, and achieves 22% faster inference speed than OpenPose, effectively improving the accuracy and real-time performance of martial arts pose estimation.

Keywords

Martial Arts Pose Estimation, Convolutional Neural Network (CNN), Graph Neural Network (GNN), Human Keypoint Detection, Real-Time Motion Analysis