The Recognition of Aerobics Movements Using Bone Data Combined with ST-GCN

Abstract

To solve the aerobics action recognition and promote the gradual intelligence and standardization of aerobics teaching and evaluation, a network model based on spatial temporal graph convolution and combined with attention mechanism was proposed. This model improved the extraction efficiency of spatiotemporal features and channel features by introducing spatiotemporal graph and channel attention mechanisms, respectively, thereby improving the accuracy of action recognition. And a time extension module was introduced into each basic module, and additional features between adjacent vertices were extracted by extending the time graph between frames. These experiments confirmed that this model exhibited high accuracy in identifying aerobics movements. The recognition accuracy of basic actions was above 93.6%, and the recognition accuracy of two actions had reached 98.4% and 98.6% respectively. For advanced actions, the recognition accuracy of the model had slightly decreased, but the average value was still above 95%. The accuracy of difficult motion recognition had also achieved good results, reaching a maximum of 94.5%. These data indicate that this model can achieve high accuracy in handling action recognition tasks of different difficulty levels, and can identify aerobics movements of different difficulty levels.