Roadway Risk Map Establishment Based on Features of Roadway Design and Traffic Situation

With the increase of vehicle ownership and complication of road systems, road traffic safety has become a serious threat to social development and property of people. Studying road traffic risks is of great significance to meet the needs of road traffic safety management. Traditional road traffic safety analysis uses count regression models to estimate the expected number of traffic accidents on road segments. However, prediction accuracy for the crash number is not high for its uncertainty. The objective of this paper is to explore a method to assess the risk level of road traffic safety instead of studying the exact crash number. Firstly, the K-Medoids clustering algorithm is used to assess the risk level of road traffic safety and to construct an urban roadway risk map based on the accidents data of Shanghai expressway. This paper also analyzes the update frequency of and the time period segmentation of the risk map. Then four supervised machine learning algorithms, including random forest, SVM, kNN and multiple ordered logit regression, are respectively applied in predicting the risk level using historical accident data (or not), road geometry factors, traffic flow data, and weather data. The result shows that the random forest has the best performance at an accuracy of 80.67%. Random forest is also used to rank the significance of variables. Finally, this paper discusses how to complete the construction of the road traffic safety risk map when historical data of traffic accidents is not available which often happens. Experiments shows that the risk level prediction accuracy can reach 78.56% which is much better than 47.9% of the traditional crash number prediction method.