Abstract: Based on the freeway accident data, traffic data, and weather data, a model is built to predict the real-time accident risk by taking traffic flow as the main influencing factor of accidents. Firstly, regarding each accident in the records as a case and the experimental case group is obtained, and then by the case-control study method, the matched corresponding control group set up. Then the random forest algorithm is used to screen out the 10 most important variables impacting the accident risk most. Finally, a model is built to predict real-time accident risk based on the support vector machine. Experiments show the SVM model we built works when predicting the accident risk. At the same time, the SVM model performance with the Gaussian and Sigmoid kernel is better than with the linear and polynomial kernel. Experiments showed the SVM model we built works when predicting the accident risk. At the same time, the SVM model performance with the Gaussian or Sigmoid kernel is better than with the linear or polynomial kernel. Especially the accuracy of the SVM model with the Gaussian kernel reaches 73.20% and 91.44% respectively when predicting the accident and non-accident situation.

Key words: traffic engineering, accident risk, support vector machine, case-control study, random forest