Curling Video Action Recognition and Instructional Design Based on the MSTF-GCN Model

Abstract

With the deepening of smart sports construction and the popularization and development of ice and snow sports, the refined teaching and standardized training of curling have put forward higher requirements for intelligent action recognition technology. In response to industry pain points such as strong temporal correlation of curling movements, difficulty in capturing fine-grained features of the force chain, and insufficient spatiotemporal feature fusion, this study constructs a multi-scale spatiotemporal fusion graph convolutional network (MSTF-GCN) action recognition model for curling motion videos, achieving high-precision classification, quality evaluation, and error diagnosis of the two core actions of curling throwing and ice brushing. The model solves the problem of insufficient modeling of long-term curling action context and loss of fine-grained action features in traditional action recognition models by optimizing human joint structure, designing multi-scale spatiotemporal feature fusion units, and introducing adaptive gating attention mechanism. The experimental results show that the MSTF-GCN model proposed in this study achieved a Top-1 recognition accuracy of 97.26% on a self-built curling action dataset, with an average action quality evaluation error as low as 2.18 points. The recognition accuracy on the NTU RGB + D 120 public dataset was 4.37% points higher than that of the mainstream ST-GCN model, and it has significant performance advantages in curling-specific action recognition compared with mainstream action recognition models, while maintaining good generalization performance on public datasets. In practical teaching applications, the intelligent teaching scheme designed based on this model can improve the average movement standard of subjects at different levels by 42.21%, and the diagnostic accuracy of common incorrect movements in curling can reach 93.15%. This study provides a practical technical path for the intelligent and standardized teaching of curling, and also provides theoretical reference and methodological support for computer vision assisted training in ice and snow sports.