Automatic action recognition technology combining pyramid convolution and attention mechanism in gymnastics training

This study proposes an action recognition method that combines pyramid convolution and attention mechanism to address the problem of low efficiency and inability to effectively capture the detailed features of movements in traditional gymnastics training. The aim is to achieve the level of intelligence in gymnastics training and improve the precision of action recognition. This method improves the efficiency and accuracy of finite element by constructing a multi-scale feature detection model based on a pyramid convolutional kernel dual stream neural network. To further enhance the performance of action analysis, an improved selective kernel attention mechanism is introduced, and a method for automatic motion analysis that incorporates attention mechanism alongside multi-level features is proposed. The outcomes indicate that in comparison with conventional dual stream neural networks, the proposed method improves the accuracy of spatial flow and event flow by 5.00 % and 3.12 %, respectively. In comparison with the original attention mechanism, the recall rate of the proposed method increases by 9.73 %, accuracy increases by 5.79 %, and the average accuracy of motion analysis for spatial and temporal streams increases by 1.99 % and 0.83 %. The outcomes reveal that the proposed action recognition method can efficiently extract key features and has excellent accuracy in recognizing gymnastics training actions. This study introduces an innovative technological approach in the realm of sports science, which has the potential to enhance the intelligence quotient of athlete training and maximize training efficiency.