Research On Vehicle Collision Risk Prediction Method Under The Influence Of Multiple Factors

In 2021,WHO launched the Decade of Action for Road Safety 2021-2030,which aims to reduce road traffic casualties by at least 50 per cent by 2030.In order to achieve this goal,it is urgent to do a good job in the research of vehicle collision prediction,so as to help vehicles better advance warning and avoid the occurrence of traffic accidents.The occurrence of vehicle collision accidents is affected by people,vehicles,roads and environment,especially the driving behavior represented by the driving secondary task.Based on this,this paper will study the vehicle collision risk prediction method under the influence of multiple factors on the basis of the discussion on the risk level of driving secondary task.The main work of this paper is as follows:(1)Data preprocessing.In this paper,the 100-CAR NDS data set is taken as the research object and combined with the overall research objective,two kinds of data sets are constructed: " secondary task data set" and "SCE data set".For the " secondary task data set",the data extraction and merging,the data preprocessing of the screening and retention of the driving secondary task were carried out,and 780 positive example(SCE)samples and 4420 negative example(non-SCE)samples were obtained.For "SCE data set",abnormal trip elimination,missing and abnormal data filling,jerk feature generation,normalization processing,sample data window processing and marked data preprocessing were carried out,and 1225 positive samples(collision event class)and16,050 negative samples(near collision event class)were obtained.(2)Research on risk level division of driving secondary task.Based on the "secondary task data set",this paper uses Chi-square independence test and Cramer’s V coefficient to clarify the correlation between driving secondary task and SCEs.A comprehensive case-control study and meta-analysis were used to calculate the risk degree OR value and 95% confidence interval for 39 driving secondary tasks.The Kmeans clustering algorithm was used to complete the clustering of all the driving secondary tasks with significant influence,and the driving secondary tasks were further divided into five levels.(3)Research on collision risk prediction methods.This paper carries out research based on "SCE data set".First,select the static influencing factors by using the random forest feature selection method,reduce 11400 dimension SCE sequential data to 20 dimension by kernel PCA method,use SMOTE oversampling method to process the unbalanced dimensionality reduction data.Secondly,considering the time series data such as vehicle kinematic parameters and vehicle-vehicle relationship,this paper constructs a vehicle collision risk prediction model by using seven kinds of machine learning methods,and draws the conclusion that the Random Forest,Ada Boost and XGBoost models are better.On the basis of time series,the results of risk level division and undivided risk level of driving secondary tasks were introduced,and the above three models were used to build a vehicle collision risk prediction model.The results showed that the overall performance of the model was significantly improved after the risk level division results of driving secondary tasks were introduced.On the basis of time series and risk level division of driving secondary tasks,12 static factors including weather conditions,lighting conditions and so on were introduced.The above three models were used to build a vehicle collision risk prediction model under the influence of multiple factors.It was found that compared with the improved model,the overall performance of the XGBoost model was further improved,and the comprehensive performance of the XGBoost model was better.