Research On Expressway Visual Environment Information Extraction And Accident Impact Based On Machine Learning

As an important component of China’s road traffic,the expressway has great significance for the development of the transportation industry.At the same time,expressway traffic accidents have the characteristics of high frequency and severity,which seriously affect the safety operation of road.How to figure out the key factors that affect expressway traffic accidents is of great significance to ensure the safety of expressway operation and drivers.In this study,some sections of Qing-Yin Expressway are selected as the research object.Through the statistical analysis of the basic data and traffic accident data of this section of expressway,characteristics such as the time distribution of expressway traffic accidents,the distribution of horizontal and vertical sections of roads,and the distribution characteristics of different road types are obtained.By analyzing existing data and considering the feasibility of data collection and processing,the study road is divided by the fixed length method,the geometric linear data of the expressway and the structure data of bridge and tunnel are processed accordingly.And through the real vehicle experiments,videos are collected using a camera as driver visual environment data.Using Python as the development tool,we build the Detectron2 platform and select a deep learning algorithm suitable for research data.Then,panoramic segmentation of the driver’s visual environment image is carried out,and roads,trees,hills,buildings,etc.,are identified and marked.The mean value,standard deviation and coefficient of variation of the proportion of roads,trees and other objects in the driver’s visual environment are obtained by calculating the proportion of pixels of various objects.The number of signboards in the unit research section is counted and the index to measure the complexity of the driver’s visual environment is introduced.On this basis,the accident feature variables are encoded according to road,driver’s visual environment and time.In order to obtain the best classification results,data preprocessing based on the SMOTE algorithm and feature selection are conducted.By using the selected features,this article compares and analyzes relevant ensemble learning algorithms.Finally,the classification results of model are visually interpreted based on the SHAP value,and the important factors selected are analyzed based on the global,single factor and double factor interaction.The results show that the performance of the accident frequency classification model based on XGBoost is the best,with an accuracy of 87.8%.And for the unit study section,the complexity of drivers’ visual environment,the average longitudinal slope of road,the number of signboards,the number of route type changes and the average value of the road proportion,the average value of the tree proportion,the standard deviation of the hill proportion,and the coefficient of variation of the building proportion in the driver’s visual environment all have important effects on the occurrence of accidents.The relevant conclusions can provide some theoretical reference for the safe operation of expressway.