A data-centric weak supervised learning for highway traffic incident detection.

Area of Science:

• Traffic engineering
• Machine learning
• Data science

Background:

• Near-real-time traffic incident detection is vital for mitigating highway congestion.
• Current supervised learning methods for incident detection suffer from high false alarm rates due to inconsistent human labeling.

Purpose of the Study:

• To develop a data-centric approach for improving the accuracy and reducing false alarms in highway traffic incident detection.
• To create a weak supervised learning workflow that generates high-quality training labels without ground truth data.

Main Methods:

• A data preprocessing and curation pipeline using labeling functions to generate training data.
• Evaluation of weak supervision-generated labels with Random Forest, k-NN, SVM ensemble, and LSTM classifiers.
• Development of an online real-time detection approach using model ensembles and uncertainty quantification.

Main Results:

• Significant improvement in the accuracy of all evaluated supervised learning models after using weak supervision-generated training data.
• Achieved a high incident detection rate of 0.90.
• Achieved a low false alarm rate of 0.08.

Conclusions:

• The proposed weak supervised learning workflow effectively enhances highway traffic incident detection.
• This data-centric approach offers a practical solution for reducing false alarms in real-world traffic management systems.