Research On Driving Behavior Pattern Recognition Method Based On Multi-source Parameters

The rapid development of new science and technology such as the Internet,Big data,and Cloud Computing has provided technical support and strong guarantee for urban transportation interconnection and convenient travel of residents.Under the institutional guarantee of the strategy of rejuvenating the country through science and technology,and a country with strong transportation network,smart transportation,green travel,"Internet+",convenient payment,etc.have increasingly become people’s technical and platform requirements for high quality,excellent service,convenience,high efficiency,real-time and security for a better life.Based on this background,autonomous driving technology will also be among the soft needs of high-quality life,allowing the Internet of Vehicles technology and vehicle assisted driving technology to master the process of vehicle prediction,operation,and feedback throughout the journey,and people only need to give "instructions".In this way,the efficiency of residents’ travel can be achieved with half the effort,and safe,efficient and comfortable driving space and experience can be provided for residents’ travel.In order to ensure that the automatic driving system realizes safe and efficient driving operations during driving,it is necessary to accurately recognize the current driving behavior pattern of the vehicle.Based on this,this paper establishes a reasonable driving behavior pattern recognition model.Based on the actual vehicle test,the paper selected urban expressways and highways as test sections,selected 20 drivers for the entire vehicle test,and applied Kalman Filtering for data preprocessing.In order to accurately identify driving behaviors,the characteristics reflected by people,cars and roads were coupled,and the current driving behaviors were jointly judged and recognized.According to four types of driving behaviors(left lane change,right lane change,car following and free driving),principal component analysis and single-factor variance test were used to determine characterization parameters.Then,analyze the characteristics of multi-source parameters under high-speed conditions,including vehicle motion,traffic environment and visual parameters.Finally,the paper used Random Forest Decision Trees(CART+RF)and Support Vector Machines(SVM)to establish a driving behavior fusion pattern recognition model,compared the recognition accuracy of the two types of models,and optimized the better method to improve the reliability of driving behavior recognition.The main conclusions of the thesis research are as follows:(1)Vehicle movement and driving environment characteristic parameters that can effectively identify driving behavior patterns include: lane line distance,vehicle speed,relative distance,yaw rate,and lateral acceleration.The difference in the lane line distance parameter samples under different driving behaviors is the most significant,followed by the yaw rate and lateral acceleration that have a greater impact on the recognition effect of different driving behaviors;in the process of different driving behaviors,the speed change trend is more obvious,and when changing lanes,The speed change cycle is short,and there are obvious changes in speed before and after lane changing.Under high-speed driving environment,it is usually reflected in the overtaking process.When the lane is maintained,the speed change cycle is long and the speed fluctuation is relatively stable;the relative distance affects the recognition of car-following and free driving behaviors.(2)During the driving process,the driver has a greater proportion of the visual information obtained by the traffic environment.Analyzing the change rules of various indicators of visual eye movement under the current behavior of the driver can improve the accuracy of driving behavior recognition.Visual parameters include indicators for gaze,blinking,saccade,etc.The analysis results show that parameters such as blinking frequency,gaze angle,saccade range,and gaze point distribution have significant differences under different driving behaviors,and can effectively characterize the four types of driving behavior characteristics.(3)Based on the experimental sample set,the paper establishes a driving behavior pattern recognition model,trains the learning model,and tests its recognition effect.The results show that the Random Forest optimization model based on MLP Neural Network has the highest overall recognition accuracy,reaching 91.92%;followed by the CART+RF model with an overall recognition accuracy of 90.49%,which has the best recognition effect for free driving behavior;SVM model has the highest recognition accuracy for car-following mode.It can be obtained from the comprehensive analysis of model recognition accuracy and ROC curve evaluation that for driving behavior fusion pattern recognition under high-speed working conditions,MLP Neural Network is used as the individual classifier of the random forest model to combine,and the MLP+RF recognition model established based on the integrated learning idea can accurately and reliably determine the current driving mode of the vehicle in real time.