Filter-Pruned 3D Convolutional Neural Network for Drowsiness Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Human drowsiness during operation of vehicles or machinery poses significant safety risks.
• Existing systems may lack real-time efficiency and computational lightness for practical applications.

Purpose of the Study:

• To develop a visual-based drowsiness detection system for real-time attention status prediction.
• To create a lightweight and computationally efficient system suitable for integration into assistance systems.

Main Methods:

• Utilized a 3D convolutional neural network (CNN) for spatio-temporal feature extraction from video frames.
• Implemented temporal smoothing to refine prediction accuracy by reducing noise.
• Introduced a novel Scale Module for estimating filter importance within convolutional layers.

Main Results:

• The Scale Module effectively identified filters crucial for model performance.
• Filters with low scale values were successfully pruned with minimal impact on accuracy.
• The developed system demonstrated potential for real-time, efficient drowsiness detection.

Conclusions:

• The proposed Scale Module facilitates model optimization for real-time applications.
• Visual-based drowsiness detection systems can be made more efficient without sacrificing performance.
• This technology can contribute to enhanced safety in transportation and machinery operation.