Risky-Driving-Image Recognition Based on Visual Attention Mechanism and Deep Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Traffic Safety

Background:

• Risky driving behaviors significantly impair driver performance and are a leading cause of traffic accidents.
• Accurate identification of driver status is crucial for timely interventions to prevent accidents.

Purpose of the Study:

• To develop an advanced image-recognition system for identifying risky driving behaviors.
• To enhance the accuracy and efficiency of driver status identification using deep learning and visual attention.

Main Methods:

• Proposed a risky-driving image-recognition system integrating visual attention mechanisms with deep learning.
• Developed four ResNet-based deep learning models of varying depths.
• Embedded visual attention blocks into the image classification models.

Main Results:

• The integration of visual attention modules improved classification accuracy in lower-depth ResNet models.
• Lower-depth models with attention modules outperformed higher-depth models in accuracy.
• Model complexity remained largely unchanged, indicating improved recognition efficiency.

Conclusions:

• Visual attention mechanisms can enhance deep learning models for risky driving identification.
• Effective driver status recognition can be achieved without compromising computational efficiency.
• The proposed system offers a promising approach to improving road safety by detecting distracted driving.