Motion Fatigue State Detection Based on Neural Networks.

Area of Science:

• Sports Science
• Computer Vision
• Machine Learning

Background:

• Athlete fatigue poses risks in sports.
• Accurate fatigue detection is crucial for performance and safety.
• Existing methods may lack real-time capabilities or robustness.

Purpose of the Study:

• To develop a deep learning system for detecting athlete fatigue.
• To propose a novel convolutional neural network (CNN) fatigue detection model.
• To enhance fatigue detection accuracy and real-time performance.

Main Methods:

• A cascade deep learning system utilizing MTCNN for face detection and RESNET-based multiscale pooling (MSP) for eye and mouth state analysis.
• Training the CNN model with specific parameters (batch_size=100, initial learning rate=0.01, reduced to 0.001, 50 epochs).
• Determining fatigue using PERCLOS and a proposed mouth opening and closing frequency (FOM).

Main Results:

• The system achieved high precision and recall in fatigue detection.
• RGB images in daytime conditions yielded better results than simulated night (infrared) conditions.
• The deep learning model demonstrated high detection accuracy, met real-time requirements, and showed robustness in complex environments.

Conclusions:

• The proposed CNN fatigue detection model is effective and accurate.
• The system is suitable for real-time fatigue monitoring in athletes.
• The approach offers a robust solution for fatigue detection across various conditions.