A deep learning based traffic crash severity prediction framework.

Area of Science:

• Transportation Engineering
• Traffic Safety
• Machine Learning Applications

Background:

• Highway work zones are high-risk areas for traffic congestion and collisions.
• Accurate prediction of crash severity is crucial for emergency response and traffic management.
• Existing statistical and machine learning models show limitations, particularly for severe crashes.

Purpose of the Study:

• To propose a novel deep learning approach for predicting traffic crash severity at work zones.
• To compare the deep learning model's performance against traditional machine learning models.
• To optimize prediction for precision and recall, focusing on severe crash outcomes.

Main Methods:

• Utilized crash data from Louisiana work zones (2014-2018) including road, vehicle, and human factors.
• Transformed crash data features into images using t-SNE and convex hull algorithms.
• Developed a Convolutional Neural Network (CNN) with a customized f1-loss function to optimize precision and recall.

Main Results:

• The deep learning approach demonstrated improved performance in predicting fatal and injury crash severity.
• The customized loss function effectively optimized the model for precision and recall.
• The methodology shows potential for enhancing traffic safety and mitigating congestion.

Conclusions:

• Deep learning with a customized f1-loss function offers a superior method for predicting traffic crash severity in work zones.
• This approach can significantly contribute to improving road safety and reducing traffic congestion.
• The image transformation technique combined with CNNs provides a promising direction for future traffic safety research.