Study On Dangerous Driving Behaviors Based On Driving Data

In recent years,with the socio-economic development,the number of domestic cars has also increased significantly.And also have been an increasing number of traffic accidents.There is a great significance to study how to prevent traffic accidents occurrence.According to statistics,the number of accidents caused by drivers accounts for more than 90%.More and more scholars are conduct in-depth research on driving behaviors,but domestic research still needs to be improved.Driving data can intuitively reflect driving behavior.Based on this,There analyze the driving data through data processing algorithms to identify the dangerous driving behaviors.The main contents are as follows:Taking cars as the research object,the dangerous driving behavior of "three rapids and one over" are explained,and according to different characteristics of different dangerous driving behaviors,there proposed different judgment conditions for these behaviors.Under different road conditions,the determination of speeding behavior should be set different thresholds.According to the limits of the vehicle speed in different ranges,there are correspondingly different limits for rapid acceleration to determine the driving behavior of rapid shift.Meanwhile,the determination of rapid turn driving behavior should also be based on different lateral acceleration thresholds for determination according to different road conditions.As a large amount of data is needed to support the research,NGSIM data is selected as the basic data for the research.According to the needs of study,there makes data cleaning,unit conversion,standardization and other pre-processing.Conduct initial clustering of driving behaviors.The principle and characteristics of k-means clustering are analyzed,a commonly used clustering algorithm,and finds that the improved k-means++ clustering algorithm can eliminate the shortcomings of k-means algorithm and obtain the ideal clustering result.Therefore,k-means++ algorithm is taken as the clustering method.Three types of driving behaviors are obtained: dangerous,safe and hidden,and "three rapid and one over" is classified as dangerous driving behaviors.Using KNN and SVM,the decision tree and random forest algorithm with k-means++ do joint classification clustering algorithm respectively,and illustrate the principles of four kinds of classification algorithm,through the confusion matrix method to calculate the classification accuracy index said the classification accuracy of each method,the comprehensive comparison of three accuracy indexes,the results show that the SVM and k-means++ joint classification has the highest accuracy,each category classification accuracy can reach approximate to 1.This method was finally determined as the classification method in driving behavior analysis.According to the classification results,the BP neural network recognition model is established to implement that the driving data into the model and output the driving behavior category number.Explain the composition and algorithm principle of BP neural network,set up parameters of neural network according to the requirements,build models and train neural network to identify "three rapid and one over" dangerous driving behavior and three types of driving behavior.After completing the driving behavior analysis model,the model was verified with the actual driving data,and the results of the model were summarized to obtain the frequency of various kinds of dangerous driving behaviors in the actual data.