The Study Of The Crash Risk Prediction And Crash Causation Analysis For Freeways Based On The Real-time Traffic Flow

Relying on various types of high-tech,China’s freeway system has constantly been developing in informatization and intelligence.However,most of the corresponding traffic safety control still uses the traditional management methods,which mainly reflects in the more on enforcement of traffic law and much less on traffic management.Which obviously cannot adapt to or meet the current and future needs of freeway traffic safety management and sustainable development.To help freeway traffic managers formulate reasonable traffic active safety management strategies and develop efficient crash prevention and control techniques,the author has taken the real-time traffic flow as an entry point and conducted a series of related research.By mining crash precursor characteristics,constructing crash risk prediction models,and analyzing crash causation under different traffic states,the author aims to realize the active control of crash risk and improve traffic safety management.The specific research contents could be summarized as follows:(1)Research on the crash prediction model for freeways based on the real-time traffic flow dataFirstly,the thesis extracted more than 800 crash records and related traffic flow data,and meteorological data and built a crash risk candidate variables set consisting of traffic flow and external environmental variables.Then the paired Case-Control Study method was used to match the non-crash corresponding data for each crash record.Next,the Random Forest algorithm was used to select important variables that significantly affected crash risk and regarded the selected eight variables as crash precursors to build several experiment sample sets.Finally,the Support Vector Machine algorithm,which’s collaborated with some most used kernel functions,was used to develop the crash risk prediction model.The experimental results have proved the model’s availability and shown that the Gaussian kernel and Sigmoid kernel collaborated with the model could get a higher AUC value(reaches 0.721 5 and 0.740 9 respectively),which are the ideal kernel functions for the model.When designing the experiment sample structure,increasing the crash and non-crash matching ratio could enhance the model’s performance to a certain extent;thus,the ratio should be defined according to the actual situations.(2)Analysis of the crash causation in different traffic states on the freewayFirstly,the thesis,from classification purpose,basis to standard,analyzed the basic principles of the most commonly used theories or methods for traffic state classification;at the same time,an unsupervised clustering algorithm was introduced to divide the traffic states of the experiment sample sets as well.Then,conditional Logistic Regression modeling was used to analyze crash risk levels under different traffic conditions and the distribution differences of crash-related traffic flow variables under each state.Again the Random Forest algorithm was used to select the traffic flow variables that significantly affect crash risk in each traffic state.Then,for each state,the conditional Logistic Regression was used to establish the statistical relationship between the crash-related traffic flow variables and the crash risk.By analyzing the various traffic flow characters in each state,the thesis has found the crash’s causation in different traffic states.The experimental results have shown that the distribution of crash-related traffic flow variables is different under different traffic conditions;conclusions could provide a theoretical reference for the subsequent establishment of crash risk prediction models based on various traffic conditions.