MA-Pose and DMMS-GCN: skeleton-based martial-arts category recognition with edge-device deployment

With the development of visual sensor technology and human pose estimation algorithms, skeleton-based action recognition has received extensive attention in sports analysis and training systems. The recognition of martial-arts actions is difficult because different styles can share similar poses while differing in rhythm and coordination. To support this task, we constructed the Martial Arts Pose (MA-Pose) dataset, which contains 433 source videos across 13 martial arts categories. The 33-landmark sequences include normalized image coordinates and monocular relative depth rather than metric 3D ground truth. We further propose a Dynamic Multi-Stream Multi-Scale Graph Convolutional Network (DMMS-GCN), which integrates joint, bone, and motion features. The model combines physical, learnable, and input-adaptive graph topology with parallel temporal convolutions. With a fixed seed of 42, 346 videos are used for training and 87 for validation, with one center crop and one prediction per validation video. DMMS-GCN correctly classifies 73/87 videos (83.91% Top-1; 83.51% Macro-F1; 95% Wilson confidence interval, 74.78%–90.17%) and retrieves the correct class among the top five for 84/87 videos (96.55%). These values describe development-set performance because model selection and evaluation use the same validation subset. On a Raspberry Pi 4B, pose extraction runs at approximately 10 frames/s and classification of a prepared 150-frame tensor takes approximately 120 ms. Collecting the first 150 processed pose frames takes about 15 s, so the deployment demonstrates edge-device feasibility rather than instantaneous recognition. The source code, model checkpoint, and MA-Pose dataset are publicly available.