A Comparative Study Of Prediction Models On The Duration Of Traffic Incidents

Traffic incident is a big challenge for traffic safety management. The duration of traffic incidents is critical to measure the severity of the incident and essential for improving traffic management abilities. We collected the traffic incident data of major urban roads in Tianjin, China, for consecutive 2 years. Our study conducts a comprehensive analysis of the main factors affecting the duration of the traffic incidents, such as the occurrence time, the road type, the vehicle type, and the number of injured. By using the classical statistical regression model and the several popular machine learning techniques, we try to predict the duration of the traffic incidents and compare the results generated by these models. The results show that the prediction errors of the model are distributed between 30% and 40%. The model established by the machine learning techniques has higher precision than simple linear regression models. Notably, support vector regression performs best among these models. Moreover, the tree-based can be applied for a fast estimation of the traffic incident duration, which deserves application in real-life scenarios.