Fitness Movement Types and Completeness Detection Using a Transfer-Learning-Based Deep Neural Network.

Area of Science:

• Computer Science
• Artificial Intelligence
• Sports Science

Background:

• Home fitness is popular for convenience and safety, but limited learning ability can lead to injuries.
• Existing fitness detection systems (wearable, body-node, image deep learning) have limitations in cost, accuracy, or user experience.
• There is a need for a cost-effective, accurate, and timely fitness detection system to enhance safety and awareness.

Purpose of the Study:

• To develop a deep transfer learning-based system for detecting fitness movement type and completeness.
• To establish a fitness database using Yolov4 and Mediapipe for real-time movement detection.
• To improve the accuracy and reduce the risk of injury in home fitness practices.

Main Methods:

• Utilized deep transfer learning to create a fitness database.
• Employed Yolov4 and Mediapipe for real-time fitness movement detection and 1D signal storage.
• Applied Multilayer Perceptron (MLP) for classifying 1D fitness signal waveforms.

Main Results:

• Achieved high performance in fitness movement type classification: mAP 99.71%, accuracy 98.56%, precision 97.9%, recall 98.56%, F1-score 98.23%.
• Demonstrated strong performance in fitness movement completeness classification: accuracy 92.84%, precision 92.85%, recall 92.84%, F1-score 92.83%.
• The system achieved an average detection frame rate of 17.5 FPS, outperforming existing methods.

Conclusions:

• The proposed deep transfer learning method offers a highly accurate and efficient solution for fitness movement detection.
• This system effectively addresses the limitations of previous methods, providing a valuable tool for home fitness.
• The findings suggest significant potential for AI in enhancing fitness safety, awareness, and user experience.