ABSTRACT
Road traffic incident detection has long been a well-attended research domain of intelligent transportation systems. However, the problems of sample imbalance and small sample size have hindered the progress of such studies. The study proposes a hybrid model (i.e., ADASYN-XGBoost) consisting of an adaptive synthetic over-sampling method (ADASYN) and XGBoost for automatic and efficient traffic incident detection. The ADASYN is employed to enhance the quality of traffic accident samples, and then XGBoost is applied to filter the features of the intensified samples. For the built ADASYN-XGBoost model, we conducted many experiments on a publicly available Portland Highway dataset. Experimental results revealed that our proposed ADASYN-XGBoost excelled in the advanced benchmark models with higher performance indexes and strong adaptability.
