Related Experiment Videos
Athletes' action recognition and performance prediction based on TS-GCN-SAM
Shuangfei Liu1,2, Zhengbo Chen2, Biao Chen3
1School of Public Teaching, Ningbo Polytechnic University, Ningbo, 315800, Zhejiang, China.
Scientific Reports
|July 16, 2026
Summary
This study introduces a new Two-Stream Graph Convolutional Network (TS-GCN) with a Spatial Attention Module (SAM) for accurate athlete movement prediction. The model enhances performance analysis, aiding coaches in real-time training adjustments.
Area of Science:
- Sports Science
- Computer Vision
- Machine Learning
Background:
- Competitive sports require precise athlete movement analysis for performance optimization.
- Existing methods for athlete performance prediction often lack efficiency and accuracy in capturing dynamic movements.
Purpose of the Study:
- To develop an advanced athlete performance prediction model.
- To improve the accuracy and efficiency of human movement recognition in sports.
Main Methods:
- Utilized OpenPose technology to extract skeletal data from videos.
- Developed a Two-Stream Graph Convolutional Network (TS-GCN) integrating spatial and temporal streams.
- Incorporated a Spatial Attention Module (SAM) to focus on critical joint features.
Main Results:
- The TS-GCN-SAM model achieved Top-1 (33.6%) and Top-5 (56.1%) accuracies on the Kinetics-700 dataset.
- Demonstrated superior performance compared to traditional athlete movement analysis methods.
- Showcased model stability and efficiency, with a learning rate of 0.001 yielding faster convergence and 88.2% accuracy.
Conclusions:
- The TS-GCN-SAM model significantly enhances human movement recognition accuracy.
- Provides coaches with real-time feedback for dynamic training plan adjustments.
- Offers a valuable tool for optimizing training content and athlete performance.