Application of PoseConv3D algorithm in cheerleading training action recognition.

To address the low recognition accuracy and high misclassification rate of existing models caused by the highly dynamic, strongly coordinated, and rhythm-bound characteristics of cheerleading movements, this study improves a Pose-based Convolutional Neural Network 3D (PoseConv3D). By embedding three core modules-dynamic skeletal correction, multi-scale spatiotemporal encoding, and motion rhythm perception integrating a dual-path classification layer, the proposed method achieves joint output of action categories and difficulty coefficients. Experiments conducted on a public benchmark dataset demonstrate that the proposed model attains an accuracy of 93.2%, representing improvements of 11.5% and 8.5% over the Spatial-Temporal Graph Convolutional Network (ST-GCN) and the original PoseConv3D, respectively. The dynamic correction module reduces joint drift error by 42.3%, while the rhythm perception module lowers the misclassification rate of “wave/kick” movements by 12.7%. Additionally, an inference speed of 88.2 frames per second meets real-time application requirements. The main contributions of this work are threefold. First, a novel dynamic-rhythm fusion recognition framework is proposed. By incorporating a rhythm-aware temporal attention mechanism and a confidence-guided skeletal correction strategy, the framework effectively addresses two key limitations of existing approaches: the insufficient modeling of temporal rhythm in graph neural networks (e.g., ST-GCN) and the poor robustness of Transformer-based methods (e.g., PoseFormer) to joint drift under high-dynamic conditions. Second, a closed-loop “recognition-difficulty quantification” system tailored to cheerleading is constructed. Finally, the proposed framework offers a transferable solution for the recognition of high-dynamic sports movements, providing methodological support for broader applications in complex action analysis.