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

An imbalanced and small training sample can cause an incident detection model to have a low detection rate and a high false alarm rate. To solve the scarcity of incident samples, a novel incident detection framework is proposed based on generative adversarial networks (GANs). First, spatial and temporal rules are presented to extract variables from traffic data, which is followed by the random forest algorithm to rank the importance of variables. Then, some new incident samples are generated using GANs. Finally, the support vector machine algorithm is applied as the incident detection model. Real traffic data, which were collected from a 69.5-mile section of the I-80 highway, are used to validate the proposed approach. A total of 140 detectors are installed on the section enabling traffic flow to be measured every 30s. During 14 days, 139 incident samples and 946 nonincident samples were extracted from the raw data. Five categories of experiments are designed to evaluate whether the proposed framework can solve the small sample size problem, imbalanced sample problem, and timeliness problem in the current incident detection system. The experimental results show that our proposed framework can considerably improve the detection rate and reduce the false alarm rate of traffic incident detection. The balance of the dataset can improve the detection rate from 87.48% to 90.68% and reduce the false alarm rate from 12.76% to 7.11%. This paper lends support to further studies on combining GANs with the machine learning model to address the imbalance and small sample size problems related to intelligent transportation systems.