Research On Analysis And Prediction Of Driving Behavior Based On Highway Scene

In recent years,with the rapid development of the Internet,cloud computing,big data,5G technology and other industries,the level of vehicle intelligence has been greatly improved.However,there are still many technical challenges in promoting the development of higher-level autonomous driving technology.For example,in the face of complex and changing driving scenarios,the driving behavior exhibits strong spatiotemporal characteristics and randomness,that is,the driving behavior will change with different traffic environments.Even in a single driving scenario,driving behaviors will show diversity as the driver’s state changes.The prediction of the vehicle’s motion behavior affects the optimal control of the driving assistance system and the quality of the vehicle’s decision-making and planning,and it plays a vital role in improving the driving safety and comfortability.Therefore,how to effectively describe and predict the driving behavior has become a hot issue in the industry.To address the above problems,based on the highway scenes,this paper quantitatively analyzes the driving behavior characteristics in the path-following and lane-changing mode.Moreover,the driving behavior characteristics have been integrated in two different modes to dynamically evaluate the driving behavior and establish a probability distribution.The main research contents are as follows:(1)Research on longitudinal driving behavior characteristics based on simulated driving data.In order to analyze the driving behavior characteristics of different drivers and consider the impact of driver state factors on driving performance,baseline car following driving test and random N-Back test are performed based on driving simulation platform,and the driver’s physiological information and vehicle operation status information are collected.Entropy weight method is used to quantitatively evaluate the driving behavior characteristics of the car following situations,and the N-Back test is used to induce the driver to produce different driving states.Therefore,the relationship between the driving behavior characteristics change and different driving states can be analyzed.Simulation results show that the more intensive the task,the greater the driver’s psychological load and psychological stress,and the more unstable the driving performance,the aggressive driver intend to take more strongly response to the secondary task.(2)Research on lateral driving behavior characteristics based on natural driving data.Firstly,the vehicle-related lane-changing information is extracted from NGSIM’s US-101 road segment data to summarize the change rule of driver’s driving performance.Then,the vehicle’s lateral position deviation and vehicle’s lateral speed are determined as the effective factors to represent lane-changing behavior.Through the correlation analysis for the driving behavior factors and the lane changing time,one can find that the lateral speed is highly related with the span of the lane changing.To solve the dynamic changing problem of driving characteristics and the problem caused by single driving mode,we have quantitatively evaluated the driving behavior characteristics in the car-following mode and lane-changing mode.The results derived from the comprehensive evaluation of driving behavior characteristics in different modes are testified more objective and effective than that gained from the holistic assessment of no distinction between driving modes.(3)Recognition of driving behavior based on GMM-HMM model.In order to recognize the driver’s uncertain driving behavior during driving,a driving behavior recognition model based on GMM-HMM is established.The left lane change,lane keeping,and right lane change behaviors are unobservable hidden states,and the vehicle lateral position deviation and lateral speed are observable variables.Use the US-101 road data of NGSIM to train the model,use EM algorithm to optimize and set the parameters of HMM and GMM models,and use forward and backward algorithms to predict the probability of vehicles in various behaviors,which effectively guarantees the correct rate of behavior prediction.The model was verified using the I-80 road segment data set of NGSIM,and the results show that the recognition accuracy of the lane-changing behavior1 s before the lane-changing point reaches 95%,and the recognition rate around the moment when the intention occurred is as high as 80%.