HFGCN-Based Action Recognition System for Figure Skating

With the rising popularity of winter sports, there is a growing interest in figure skating, and an action recognition algorithm for figure skating is increasingly needed. The figure skating action recognition algorithm is not only good for intelligent event understanding but also for objective competition judging. However, figure skating actions are difficult to distinguish by one frame, and the errors caused by the keypoint detection may be propagated in the action classification, and affect the performance of the figure skating action recognition algorithm. In this paper, a figure skating hierarchical dataset FSHD-10 and a Figure Skating Action Recognition System (FSARS) are established. The FSARS adopts a multi-stream structure and a decision fusion module to learn features at different dimensions. A Hierarchical Fine-Grained Graph Convolutional Neural Network (HFGCN) is also proposed in this paper. The HFGCN extracts the temporal features by a temporal modeling module and an attention module. The hierarchical classification adopted in HFGCN takes advantage of the hierarchy of figure skating actions and improves the precision of the action classification. The experimental results show that these improvements contribute to FSARS and make it achieve a final accuracy of 93.70% on the FSHD-10 dataset.