Intelligent recognition of fundamental physical training actions based on skeleton-rendered pose images

Abstract

Fundamental physical training action recognition plays an important role in intelligent sports analysis, physical education, and motion monitoring. However, RGB-video-based recognition methods are sensitive to background variation, clothing, illumination, and privacy-related constraints, while graph-based skeleton models and pose-heatmap pipelines may introduce additional modeling or preprocessing complexity. To address these issues, this study systematically investigates a lightweight skeleton-rendered pose image pipeline for six fundamental physical training actions. OpenPose was used to extract human keypoints, normalized skeletons were rendered as grayscale pose images, and two-dimensional CNN backbones with temporal aggregation were evaluated under five-fold subject-wise cross-validation. Representative RGB-based, graph-based skeleton, and pose-heatmap baselines were also added for comparison, including R3D-18, Video Swin-T, ST-GCN, CTR-GCN, and PoseC3D. Experimental results show that ResNet18-Max achieved 92.30 ± 0.57% accuracy and 91.88 ± 0.60% macro F1, with 11.69 M parameters, 29.20G FLOPs per 16-frame clip, 18.72 ms inference time, and 53.42 FPS under the implemented setting. The results indicate that skeleton-rendered pose images provide a practical accuracy-efficiency trade-off for structured physical training action recognition.