Research On Vehicle Collision Prediction Based On Car Usage Habits And Driving Spatiotemporal Characteristics

A report from WHO shows that traffic accident has become the 8th leading cause of human deaths,and the number one killer of teenagers ranging from 5 to 29 years old,accounting for 2.5% of global deaths.Therefore,to study the influencing factors of vehicle collisions and start targeted driving publicity and education,have important social significance for reducing the risk of vehicle collisions.In addition,with the rise of UBI auto insurance based on driving behavior in recent years,the prediction of vehicle collisions severity also has broad commercial prospects.Based on study of historical documents,this article finds two directions worthy of further research: First,the existing research on vehicle collision factors does not fully consider subjective and objective factors.Meanwhile,with the introduction of new data sources,there exists the space for further digging on conducted factors.Second,in terms of vehicle collision prediction,existing research mainly focuses on whether a vehicle will collide or not,but it is not enough to predict on collision degrees such as non-collision,ordinary collision and severe collision.The prediction of collision degree,however,has a huge impact on the life and property of drivers and passengers.Aiming at the directions mentioned above,this paper uses the data from Shanghai area as a sample.By using big data technology,it comprehensively analyses vehicle driving behavior,maintenance behavior,driving environment and other multi-dimensional data,and discovers new key factors that affect vehicle collisions.Furthermore,on this basis it constructs a prediction model of the vehicle collisions severity based on machine learning technology.The main research methods and contents are as follows:First of all,use the new data source to construct a data set and preprocess the features.In addition to introducing basic vehicle information,connected vehicle driving control data,and weather data that are currently used in collision research and auto insurance pricing,this article also includes relatively few researched vehicle connected GPS data,connected vehicle collision data,and after-sales maintenance data.In the feature preprocessing stage,operations such as missing value filling,normalization,dummy variable setting,and data binning are performed respectively to generate the most suitable data set for model training.Secondly,from the perspectives of vehicle habit factors and driving time and space factors,a model for predicting the severity of vehicle collisions is established,and the high-impact feature groups and characteristic variables for vehicle collisions are discovered.This paper constructs 77 characteristic variables,and divides them into seven characteristic groups based on business background and characteristic attributes,namely,vehicle speed,acceleration,turning,maintenance,severe weather,unfamiliar environment and driving time.Based on these feature groups and feature variables,this paper proposes two classification models for predicting the severity of a vehicle collision: one is a hierarchical two-class model,which first classifies whether a collision occurs,and then distinguishes the severity of the collision;the other is a three-class model,which directly classifies non-collision,normal collision and severe collision.Then,the high-impact feature groups and feature variables are excavated through experiments.The steps are as follows:1.Establish a basic model by using the feature groups studied by the predecessors as the basic feature groups;2.Add successively the feature groups and feature variables to be studied,according to the two dimensions of vehicle habit factors and driving time and space factors,to obtain high-impact feature groups and high-impact feature variables;3.Add all feature groups to the basic model in permutation and combination,to analyze the combination effect between each feature group,aiming at obtaining the most efficient feature group combination.In the overall model experiment,in order to avoid the errors caused by some contingency,we adopted multiple models,multiple evaluation indicators,multiple experiments and other methods to avoid related error risks.Finally,based on the econometric model,the interaction effects between the influencing factors of vehicle collisions are studied.Through experimental research,this paper finds that the pairwise combination of maintenance and severe weather has an extremely obvious improvement effect on predicting vehicle collisions,which exceeds any other pairwise combination of characteristics.Then,why the combination between them produces such an effect,and what is the internal mechanism,is a question worthy of discussion.Based on the econometric model,the article studies this issue.Based on the above methods,this paper obtains the following four results:First,the prediction model established in the paper can better predict the severity of vehicle collisions.As for the predictions of "no collision,normal collision,and severe collision" by the three models,the average AUC is between 0.72-0.83,and the prediction effect is good.Second,among the subjective factors that affect a collision,vehicle maintenance habits are the most important feature that affects a collision,and the "average maintenance interval" is the single variable with the highest impact.Third,among the objective factors that affect collisions,severe weather travel is the most important feature that affects collisions,and "low temperature travel" is the single variable with the highest impact.Fourth,through the study of interaction effect between the feature groups of maintenance and severe weather,it is found that:(1)The adjustment effect of the maintenance quantity: Higher maintenance frequency will reduce the impact of severe weather on vehicle collisions;(2)The adjustment effect of the maintenance quality: Higher maintenance quality will weaken the impact of severe weather on vehicle collisions.This result provides a direction for further clarifying the interrelationship between the features that affect vehicle collisions and explaining its mechanism of action.In conclusion,this article has carried out the discussion on the aspects that have not been intensively studied in vehicle collision prediction,and achieved the following three innovations:1.By introducing new data sources,new features that affect vehicle collisions are discovered.2.Propose a prediction model of vehicle collision severity based on machine learning.3.Propose the combined adjustment effect of vehicle collision impact features.All these provide valuable enlightenment for insurance companies in the strategies of pricing and cost control for UBI auto insurance.