Automated traffic incident detection with a smaller dataset based on generative adversarial networks

Area of Science:

• Intelligent Transportation Systems
• Machine Learning
• Data Science

Background:

• Traffic incident detection models often suffer from low detection rates and high false alarm rates due to imbalanced and small training datasets.
• Scarcity of incident samples presents a significant challenge for developing robust traffic monitoring systems.

Purpose of the Study:

• To propose a novel incident detection framework using generative adversarial networks (GANs) to overcome the limitations of small and imbalanced datasets.
• To enhance the performance of traffic incident detection by generating synthetic incident samples.

Main Methods:

• Extraction of spatial and temporal variables from traffic data.
• Utilizing the random forest algorithm for variable importance ranking.
• Generation of new incident samples using generative adversarial networks (GANs).
• Application of the support vector machine algorithm as the final incident detection model.

Main Results:

• The proposed GAN-based framework significantly improved the detection rate from 87.48% to 90.68%.
• The framework substantially reduced the false alarm rate from 12.76% to 7.11%.
• Experimental results validated the framework's effectiveness in addressing small sample size and data imbalance issues.

Conclusions:

• Combining GANs with machine learning models offers a promising approach to tackle data scarcity and imbalance problems in intelligent transportation systems.
• The developed framework demonstrates a considerable improvement in the accuracy and reliability of traffic incident detection systems.