Research On Safety Analysis Method Of Driving Behaviors Based On Driving Maneuver Primitives

As a convenient means of transportation for people,the vehicle population is increasing in recent years,but it also brings serious traffic safety problems.The investigation of accident inducement shows that more than 90% of traffic accidents are caused by improper driver behavior.Therefore,it is very important to study the safety of driving behavior.At present,there are some problems in the analysis and evaluation of driving behavior safety,such as using statistical indicators to evaluate from the static perspective,lacking of consideration of the dynamic characteristics of driving behavior;the purpose of the research mainly focuses on the classification of drivers,less research on individual driving behavior;the input characteristics information of prediction model is not sufficient,and the model prediction accuracy is low.Aiming at the above problems,this paper describes the characteristics of individual driving behaviors from the dynamic process perspective,proposes a driving behavior safety analysis method based on driving maneuver primitives,and establishes a risk prediction and evaluation model of driving behaviors based on the analysis results and high information expression feature set.The main research contents are as follows:1)A driving simulation platform is established by combining Unreal Engine 4(UE4),Car Sim automobile dynamics model and Logitech G29 kit.The driving data collection test is carried out,and effective following driving events is extracted based on the collected data.Based on the static standard in the European project UDRIVE,Safety Critical Events(SCEs)are extracted and the driver safety performance is pre-evaluated,which provides a reference for the verification of the research method.2)Drawing on the ideas and methods of natural language processing,Hidden SemiMarkov Model(HSMM)is used to characterize the process of driving behavior and Hierarchical Dirichlet Process(HDP)is used as the prior distribution over the HSMM parameters,so a Bayesian nonparametric learning model is built and the time-series driving data is segmented adaptively.The driving maneuver primitive,the basic component of driving behavior process,is extracted.Based on the results of extraction,the dynamic random process of driving maneuver primitives is visualized,which provides a theoretical basis for subsequent research.3)Based on the physiological and psychological perception thresholds of drivers,the semantic space of driving behavior is defined,the semantic space markers of driving maneuver primitives are analyzed and determined,and the semantic analysis of driving maneuver primitives is realized from the perspective of the dynamic process of driving behavior.Based on this,the safety risk level is quantitatively marked.A frequency map of driving maneuver primitives that can directly represent individual behavior characteristics is constructed to analyze the behavioral characteristics differences among drivers at different risks.Combining the risk level of driving maneuver primitives with the normalized frequency,a quantitative assessment method of driving safety is proposed,and the validity of the method is verified by the correlation calculation with the preassessment results.4)In order to obtain enough samples,risk markers are carried out for each event based on quantitative assessment results and clustering methods.In order to fully represent driving behavior,a high-dimensional driving behavior candidate feature set is constructed,and the candidate set is sorted by the feature optimization algorithm based on joint mutual information maximization(JMIM)according to their contribution to the research problem.A Hierarchical Gaussian Process Classifier(HGPC)algorithm is established to predict the risk of driving behavior.The test results show that the prediction accuracy of the model can reach 96.13%,which is higher than 93.55% of the support vector machine and 93.23% of the random forest.