A Deep Learning-Based Intelligent Detection Algorithm for Athlete Movement Recognition
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
This paper proposes a convolutional recurrent neural network (ST-ACGRU) based on a spatiotemporal fusion attention mechanism to improve the accuracy and efficiency of athlete movement recognition. First, the characteristics of athlete movement data are analyzed. Multi-dimensional metrics such as sit-and-reach and grip strength from uploaded data are combined to provide a basis for algorithm design. The algorithm utilizes a convolutional neural network to extract spatial features of the movements, capturing their spatial structure. A gated recurrent unit (GRU) is used to extract temporal features of the movements, processing their sequential variations. Furthermore, an attention mechanism is introduced to adaptively focus on key movement features. In the experimental simulation, a dataset is constructed using athlete test data. After normalization and other preprocessing, it is divided into training, validation, and test sets. The experiments are conducted in a specific hardware and software environment, with appropriate training parameters. The ST-ACGRU algorithm is compared with classic CNN and RNN algorithms, using precision, recall, and F1 score as evaluation metrics. The experiment constructed a dataset based on 48 samples from four scenes featuring 12 athletes (reduced to 102 training samples after Min-Max normalization and data augmentation). The performance of the four-class classification task was compared with that of simple CNNs and RNNs on an Intel i7-12700K and RTX 3090. The results showed that the ST-ACGRU achieved 100% accuracy on the test set, a 25 percentage point improvement over the CNN (75%) and a 12.5 percentage point improvement over the RNN (87.5%). Both the macro recall and macro F1 values were 100%, and the validation set loss remained stable during training (less than 0.1 difference from the training set loss). The model's generalization ability was superior to that of the competing models, providing technical support for accurate athlete movement recognition.
Fields of Study & Subjects
Context-Aware Activity Recognition Systems
Human Pose and Action Recognition
Educational Technology and Pedagogy
Publication details
- Journal / source
- IEEE International Conference on Power, Electronics and Green Energy (ICPEGE)
- Pages
- 902–906
- Publication date
- 28 January 2026
- Language
- EN