Deep Learning with Attention Mechanisms for Road Weather Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Transportation Safety

Background:

• Automatic road weather detection is crucial for transport network safety and maintenance.
• Deep learning models struggle with multi-label weather, data imbalance, and road-specific features.

Purpose of the Study:

• To enhance deep learning models for accurate road weather prediction.
• To address challenges of multi-label weather classification and data imbalance.
• To improve road safety systems through better weather condition recognition.

Main Methods:

• Utilized a focal loss function to manage imbalanced weather data.
• Employed attention mechanisms within vision transformer models for feature weighting.
• Developed and experimented with a novel multi-label road weather dataset.

Main Results:

• Focal loss significantly improved accuracy for imbalanced weather conditions.
• Vision transformers outperformed convolutional neural networks.
• Achieved 92% validation accuracy and 81.22% F1-score on the challenging dataset.

Conclusions:

• The proposed methods effectively handle multi-label and imbalanced road weather data.
• Vision transformers show superior performance in road weather prediction.
• This research contributes to safer driving and efficient road maintenance.