Exploring the Application of Neural Networks in the Learning and Optimization of Sports Skills Training

Sports skills training is a crucial component of sports education, significantly contributing to the development of athletic abilities and overall physical literacy. It is essential to utilized neural networks to optimize traditional training methods that are inefficient and rely on subjective assessments. This paper develops methods for sports action recognition and athlete pose estimation and prediction based on deep neural networks. Given the complexity and rapid changes in sports skills, we propose a multi-task framework-based HICNN-PSTA model for jointly recognizing sports actions and estimating human poses. This method leverages the advantages of Convolution and Involution operators in computing channel and spatial information to extract sports skill features and uses a decoupled multi-head attention mechanism to fully capture spatio-temporal information. Furthermore, to accurately predict human poses to avoid potential sports injuries, this paper introduces an MS-GCN prediction model based on the multi-scale graph. This method utilizes the constraints between human body key points and parts, dividing the 2D human pose into different levels, significantly enhancing the modeling capability of human pose sequences. The proposed algorithms have been thoroughly validated on a basketball skills dataset and compared with various advanced algorithms. Experimental results sufficiently demonstrate the effectiveness of the proposed methods in sports action recognition and human pose estimation and prediction. This research advances the application of deep neural networks in the field of sports training, providing significant reference value for related studies.