Ensemble learning based approach for traffic incident detection and multi-category classification

Abstract

Traffic incident is one of the important causes of road congestion. Traffic incident detection plays a crucial role in the safety application of intelligent transportation systems, which provides timely information for traffic management departments and reducing losses. Despite many researches on incident detection approach, the identification of different incident categories is not enough. In addition, traffic incident detection is still a challenging task due to the problem of data imbalance and feature selection. In this study, we propose a two-stage traffic incident detection framework based on ensemble learning. In the first stage, a binary classification algorithm based on XGBoost (eXtreme Gradient Boosting) is established to detect whether there is a traffic incident, and 24 feature variables are determined by model feature selection. In the second stage, three resampling algorithms are utilized to reconstruct and balance the dataset. Through comparative analysis, SMOTE (Synthetic Minority Over-sampling Technique)-XGBoost is the best method for incident multi-category classification with precision of 87.27%, 78.52% and 92.54%, respectively. Moreover, the baseline comparison experiments are conducted to evaluate our model performance with real-word datasets. The proposed model achieves the highest average accuracy of 93.45% in the first stage and the macro-precision of 86.11% in the second stage. The results indicate that the proposed method outperforms baselines and the two-stage framework can accurately realize the incident detection and multi-category classification.