Deep learning for sports motion recognition with a high-precision framework for performance enhancement.

Area of Science:

• Sports Science
• Biomechanical Engineering
• Artificial Intelligence

Background:

• Traditional deep learning models like LSTM and Transformers face challenges in accurately recognizing sports motion due to long-term dependencies and input noise.
• Effective sports motion recognition is crucial for athlete performance analysis, injury prevention, and monitoring.

Purpose of the Study:

• To develop a novel deep learning framework, the Evolved Parallel Recurrent Network (EPRN), integrated with wavelet transforms for high-precision sports motion recognition.
• To address the limitations of existing models in capturing complex motion dynamics and noise.

Main Methods:

• Proposed an Evolved Parallel Recurrent Network (EPRN) architecture featuring parallel recurrent pathways for enhanced temporal modeling.
• Implemented wavelet-based feature extraction to preserve fine-grained motion details across multiple resolutions.
• Evaluated the EPRN model on benchmark sports motion datasets, comparing its performance against LSTM, GRU, and CNN models.

Main Results:

• The EPRN model demonstrated superior performance, reducing Root Mean Squared Error (RMSE) by 23.5% and increasing Structural Similarity Index (SSIM) by 12.7% compared to other architectures.
• Residual analysis indicated that EPRN exhibits lower error variability and reduced sensitivity to abrupt motion transitions, signifying enhanced robustness.
• The combination of wavelet-transform-based feature extraction and recurrent deep learning significantly boosted motion recognition accuracy.

Conclusions:

• The Evolved Parallel Recurrent Network (EPRN) offers a more accurate and robust solution for sports motion recognition compared to traditional deep learning models.
• This approach holds significant potential for real-life applications including sports performance analysis, real-time motion tracking, and rehabilitation systems.
• Future research directions include multimodal data fusion and the development of lightweight EPRN variants for real-time applications.