A Novel Neural Network Model Based on Real Mountain Road Data for Driver Fatigue Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Road Safety Engineering

Background:

• Mountainous roads present unique challenges for driver fatigue detection due to poor lighting and shadows.
• Existing driver fatigue recognition systems struggle with environmental variability and subtle facial cues.

Purpose of the Study:

• To develop an improved driver fatigue recognition method specifically for mountainous road conditions.
• To enhance the accuracy and robustness of fatigue detection using the YOLOv5 neural network.

Main Methods:

• Integration of Deformable Convolutional Networks (DCNs) into YOLOv5 for improved feature extraction and handling of posture variations.
• Incorporation of a Triplet Attention (TA) mechanism to suppress noise and increase recognition robustness.
• Implementation of the Wing loss function to improve sensitivity to micro-facial features.

Main Results:

• The modified YOLOv5 model demonstrated an average accuracy of 85% in recognizing driver fatigue states.
• The DCN modules enhanced flexibility in facial feature recognition under adverse conditions.
• The TA mechanism and Wing loss function improved the model's robustness and detail capture.

Conclusions:

• The proposed enhanced YOLOv5 method effectively addresses the challenges of driver fatigue detection on mountainous roads.
• The integration of DCNs, TA mechanism, and Wing loss significantly improves recognition accuracy and reliability.
• This advancement holds potential for enhancing road safety by providing a more dependable fatigue monitoring system.