Design of Sports Training Method Based on Multilayer Feature Fusion and Deep Neural Network

Abstract

To enhance the effectiveness of conventional sports training techniques and efficiently evaluate and apply the vast quantity of feature data generated during athlete sports, this paper proposes a method based on multi-level fusion of features and deep learning. Firstly, posture feature data during human motion is collected. Then, a shallow visual feature extractor embedded in the backbone network extracts shallow visual details. Next, deep semantic features from the backbone network and shallow visual information from the branch network are aggregated into a mixed representation through a fusion layer for motion feature recognition. Additionally, a loss function is introduced to address the issue of unbalanced data samples and improve the model. The extracted motion features are then transformed into a two-dimensional image dataset, and a Visual Geometry Group neural network with an attention mechanism is used to train and recognize this dataset. The results demonstrate that the sports training detection model constructed using the proposed method achieves high accuracy, achieving an accuracy rate of over 93%.

Keywords

• Multi-level fusion of features and deep learning methods
• loss function
• attention mechanism
• visual geometry group neural network