Driver Fatigue Detection Based on Convolutional Neural Networks Using EM-CNN.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Driver fatigue is a significant cause of road accidents.
• Existing fatigue detection methods often require specialized equipment or manual analysis.
• Automated, image-based detection offers a promising solution for real-time monitoring.

Purpose of the Study:

• To propose a fully automated algorithm for driver fatigue status detection using driving images.
• To develop a novel Convolutional Neural Network (CNN) for analyzing facial features related to fatigue.
• To evaluate the algorithm's performance against established CNN architectures.

Main Methods:

• Utilized the Multitask Cascaded Convolutional Network (MTCNN) for face and feature point detection.
• Extracted Regions of Interest (ROIs) based on facial feature points.
• Developed a custom CNN, EM-CNN, to analyze eye and mouth states within ROIs.
• Quantified fatigue using Percentage of Eyelid Closure over the Pupil over Time (PERCLOS) and Mouth Opening Degree (POM).

Main Results:

• The proposed EM-CNN algorithm achieved high accuracy (93.623%) and sensitivity (93.643%) in detecting driver fatigue.
• EM-CNN demonstrated superior performance compared to AlexNet, VGG-16, GoogLeNet, and ResNet50.
• The algorithm successfully identified fatigue-related indicators like PERCLOS and POM from driving images.

Conclusions:

• The developed EM-CNN algorithm provides an efficient and accurate method for automated driver fatigue detection.
• This approach holds potential for enhancing road safety by enabling real-time monitoring of driver alertness.
• The study validates the effectiveness of deep learning models in analyzing facial cues for fatigue assessment.