Influential Factors Analysis And Prediction Of Vehicle-Vehicle Crash Severity Based On Latent Class Clustering Analysis

Urban road vehicle-vehicle crash is the most important type of traffic accident in the field of urban traffic safety.This type of traffic accident has the characteristics of high frequency,wide impact area,large casualty degree and serious property loss.Scholars at home and abroad have achieved very rich research results in this area,but the theory and method of influential factors analysis of the vehicle-vehicle crash severity still need to be refined and deepened.With the accumulation and improvement of traffic accident data,using appropriate model and methods to improve the efficiency and accuracy of data analysis is a hot research topic in the field of road traffic safety.Based on that,this paper proposes a two-stage influential factors analysis of crash severity method based on latent class clustering analysis.And using the convolutional neural network to predict the crash severity.The main work of this paper is as follows:(1)Taking the vehicle-vehicle accident data of Houston,Austin,Dallas and other cities in Texas,USA from 2017 to 2019 as the research object,this paper conducts descriptive statistical analysis from five aspects of people,vehicles,roads,environment and officially recognized causes.And then it further defines and adjusts characteristic variables and severity levels based on the results of statistical analysis.In order to avoid multicollinearity between characteristic variables,which affects the output of influential factors analysis,this paper uses the variance inflation factor for verification.It determines 20 characteristic variables and 4 severity levels in the end.(2)A two-stage method based on latent class clustering analysis is proposed to model and analyze the influential factors of crash severity.In the first stage,latent class clustering analysis is used to cluster the accident data.By comparing the evaluation indicators of the model output corresponding to different number of clusters,the accident data is finally selected to be divided into 4 clusters.This method can maximize both intra-cluster homogeneity and inter-cluster heterogeneity.In the second stage,the random forest model is used to train and fit different clusters and overall data.Based on the algorithm of permutation feature importance,it can output the important features of different clusters and overall data.Based on the algorithm of accumulated local effects,it can explore the way that the important features affect crash severity.The results show that the cluster-based model excavates five features that are masked by the overall data,namely "airbag deployment position","position of severely damaged vehicle","speed limit","gender","the relative direction of travel between the vehicles before the collision".At the same time,the study found that there are similarities and differences in important features in different clusters,and there are similarities and differences in the way that the same feature affects crash severity in different clusters.In addition,different clusters have different crash features,among which the crash location is the main feature to divide the clusters.Therefore,according to the features of different clusters,targeted measure and suggestions are put forward to help prevent and control the occurrence of malignant traffic accidents in different traffic scenarios.(3)Drawing on the existing literature,a prediction model of crash severity is constructed.By reviewing the relevant literature,the limitation that most of the current crash severity prediction models only consider the one-dimensionality of the influencing factors is pointed out.In recent years,with the development of deep learning theory,many scholars have proposed innovative methods based on deep learning to transform accident data into multi-dimensional structures for exploration.In this paper,based on the result of important features output by influential factors analysis of crash severity,the one-dimensional accident data is converted into a two-dimensional grayscale image.By training convolutional neural networks with different structures,adjusting the relevant learning parameters,it selects the model with the best training effect on the validation set to predict the severity of the test set data.The results show that the accuracy of the method on the validation set is higher than that of other comparison models.The accuracy of the test set is 3% lower than that of the validation set.The ROC curve and AUC area are better than other comparison models,indicating that the method is reasonable and feasible to predict the severity of crash data.