Graph Convolutional Network Recognizes Martial Arts Categories on Edge Hardware

Researchers have developed a Dynamic Multi-Stream Multi-Scale Graph Convolutional Network that correctly classified 73 out of 87 validation videos across 13 martial arts styles, achieving an 83.91% Top-1 accuracy and an 83.51% Macro-F1 score. The evaluation yielded a 95% Wilson confidence interval ranging from 74.78% to 90.17%, while the model identified the correct category among the top five candidates for 84 out of 87 videos, or 96.55%. Because model selection and performance evaluation relied on the same validation subset, these reported values describe development-set performance rather than an independent test set outcome.

The study addresses visual sensor technology and human pose estimation within sports analysis and training systems, where recognizing martial arts movements remains difficult because distinct fighting disciplines often share similar poses while differing in movement rhythm and coordination. To benchmark the task, the investigators constructed the Martial Arts Pose dataset, known as MA-Pose, containing 433 source videos spanning 13 martial arts categories. Rather than recording metric three-dimensional ground truth, the dataset provides 33-landmark sequences based on normalized image coordinates paired with monocular relative depth.

To process these skeletal trajectories, the proposed model, designated DMMS-GCN, integrates three separate information streams consisting of joint, bone, and motion features. The neural architecture combines physical, learnable, and input-adaptive graph topologies to capture skeletal connectivity, coupling these spatial representations with parallel temporal convolutions to track movement dynamics across sequential frames.

The experimental evaluation was conducted with a fixed random seed of 42, partitioning the MA-Pose dataset into 346 videos for model training and 87 videos for validation. During the validation phase, the evaluation protocol applied a standardized testing scheme that generated a single category prediction from one center crop per validation video, ensuring consistent input dimensions across the comparative evaluation.

Testing on a Raspberry Pi 4B demonstrated the feasibility of running the framework on resource-constrained edge hardware, though with clear operational constraints. On the single-board computer, pose extraction ran at approximately 10 frames per second, and classification of a prepared 150-frame tensor required approximately 120 milliseconds. Because collecting the initial 150 processed pose frames took approximately 15 seconds, the researchers emphasized that the deployment confirms edge-device feasibility rather than instantaneous real-time recognition. The authors have made the source code, model checkpoint, and MA-Pose dataset publicly available.

Subject of Research: Skeleton-based martial arts category recognition on edge computing devices

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

Article References: Wang, L., Zhang, R., Xu, H., & Guo, S. (2026). MA-Pose and DMMS-GCN: skeleton-based martial-arts category recognition with edge-device deployment. Frontiers in Neurorobotics, 20, Article 1967140. https://doi.org/10.3389/fnbot.2026.1967140

Image Credits: AI Generated

DOI: 10.3389/fnbot.2026.1967140

Keywords: action recognition, graph convolutional networks, martial arts, edge computing, pose estimation, sports analytics, computer vision, MA-Pose, DMMS-GCN, skeleton-based, martial-arts, category