Analysis Of The Causes And Prediction Of Road Traffic Accident Severity

With the increase in urbanization and motorization,the conflict between traffic supply and demand is becoming more and more apparent,and the number of traffic accidents is gradually increasing,putting people’s lives and property safety under serious threat.In order to reduce the likelihood of accidents and the adverse consequences they bring,this paper uses historical traffic accident data to analyze the factors that cause accidents and investigate accident severity prediction models to build an accident prediction system that can provide comprehensive assessment,early warning and timely intervention.The specific work of this paper is as follows:Firstly,statistical analysis of historical traffic accident data is carried out to obtain a preliminary accident prediction feature set;the data is pre-processed to improve the data quality,and on this basis the filtering method and Boruta packing method are applied to feature selection,and finally 19 features are screened to constitute the accident severity prediction feature set.Secondly,the accident severity prediction model was studied using the processed dataset,and the traditional machine learning algorithms of Random Forest,Adaboost,XGBoost and deep learning network models CNN and LSTM were selected for experiments respectively,while to combine the advantages of the two models,a CNN-LSTM model was constructed in this paper.Since the imbalance of sample data will reduce the prediction ability of the model for a few classes,this paper applies various sampling algorithms to solve the problem of data imbalance and obtains the optimal sampling algorithm through experimental comparison.The experimental results show that the evaluation indexes of each model are improved to a certain extent after the data processing by the sampling algorithm.The CNN-LSTM model after data imbalance treatment and parameter optimization is compared with the traditional machine learning algorithms of Random Forest,Adaboost and XGBoost,and the results show that the CNN-LSTM has the best fitting effect.Finally,the SHAP framework is applied to interpret the prediction results of the CNNLSTM model from global to individual respectively,focusing on the intrinsic mechanism of the role of each influencing factor on accident severity.The trained CNN-LSTM accident severity prediction model is applied to the Django framework to build an accident severity prediction system,which can be applied to the traffic safety department and can assist the traffic department in dealing with accidents in a more targeted manner and provide a theoretical basis for traffic safety warning.