The Hazardous Driving Behavior Description,Indentifcation,Cause And Prediction Analysis

Several advanced methods are present in increasing road driving safety with the development of technology and information transit,which have the main point in exploring the cause of hazardous driving appear and evolution based on the "drivervehicle-road" variables.However,theories and methods have some disadvantages in describing hazardous driving behavior,including theory and variables system limitations,hazardous driving identification limitations,and hazardous driving prediction limitations.First,existing theories have disadvantages in describing the multi-dimensional time-series features and exploring driver intention of behavior.Meanwhile,the present researchers model driving behavior based on the original dataset,making the overfit of variables,loss of key information,and low level of visualization.Then,driving behavior is identified based on single variables and thresholds,which ignore the feature of multi-dimensional and time-series.Otherwise,the structure of the prediction model is fixed,and most models are like a black box,which can not be expanded for analyzing the cause of hazardous driving behavior.To solve those problems,this paper focuses on exploring the cause and evolution of hazardous driving behavior based on a multi-dimensional time-series dataset,the main steps include dataset building,theory and variables system construction,hazardous driving behavior identification,and hazardous driving behavior prediction.The details are as follows:Constructing the dataset based on the naturalistic driving study(NDS)experiment.8019 trip records are collected through NDS,after the data is clean and filtered,the standard dataset is built,which includes the number of driving behavior variables,such as driver characteristics(gender,age,etc.),road environment(weather,traffic flow,etc.),driving operation data(speed,acceleration,etc.),and the surrounding vehicle data(Time To Collision(TTC),etc.).Those variables are the data basis of the hazardous driving analysis.Constructing the hazardous driving description method.This paper made a"motivation-intention-behavior" hazardous driving theory based on the planned behavior theory.This theory can be used as the theory for constructing the hazardous driving variables structure,hazardous driving recognition,and prediction.Meanwhile,this paper provided a hazardous driving behavior graph construction method through convent the continue data to piece data based on the knowledge graph theory.After verification,graph construction can enhance the model performance.Moreover,this paper extracted driver intention based on the hidden Markov model and then used it for data construction.Proposing the relationship analysis method for multidimensional time-series variables.This paper extracted the common factors based on the factor analysis,then the common factors are used to identify the hazardous driving behavior through the hierarchy clustering and Gaussian mixture model.3 types of driving behavior are clustered to caution driving,normal driving,and hazardous driving.After verification,this model can enhance the recognition accuracy and reduce the model error.Moreover,the relationship between serval driving behavior variables and hazardous driving is analyzed based on a structured equation model,the result shows that driving operation is the main factor of hazardous driving behavior;adverse weather and gender also have an influence on hazardous driving behavior.Proposing the hazardous driving behavior prediction method based on driver intention and attention machine.This paper constructed a prediction method based on the long-short term memory model(LSTM),the driver attention is extracted based on the Markov model,and the variable contribution is calculated based on the attention machine.Moreover,the hazardous driving behavior is predicted using 3 model inputs including hierarchical structure,graph structure,and panel data structure.The result shows that the graph structure and panel data can enhance the model performance.When comparing the performance of the LSTM and CNN model in prediction,the LSTM has more effective and achieves the highest accuracy based on 2 sec time interval.The result also shows that the Markov can enhance the accuracy through extracted driver intention,the attention machine can not only enhance the model performance but also distinguish the factor contribution of each variable,for example,the 1.6 sec and driving operation variables have more contribution.In general,this paper deals with a main scientific problem as to what is hazardous driving and how it evolved.The three steps of hazardous driving analysis as"description-identification-relation-prediction" are proposed face to the three steps(happen-evolution-result)of behavior happen and evolution.Meanwhile,the main scientific problem has three technical problems,what is hazardous driving,how to identify hazardous driving,and how to predict hazardous driving.This paper dealt with these technical problems,come up with a hazardous driving analysis theory and methods and modeled the hazardous driving identification and prediction.It provides a practical solution for hazardous driving warning and prevention and control of daily driving behavior.It also provides theoretical references and technical support for the improvement of traffic safety levels.