Research On Human-like Trajectory Planning And Human-like Steering Model Of Autonomous Vehicles For Curve Driving Based On Data-driven Methods

With the rapid development and wide application of artificial intelligence,highperformance sensor,5G network,high-precision map,and other technologies,the research and development of autonomous vehicles have become one of the hot spots in the global automotive industry.A large number of road tests and multiple traffic accidents also make the development route of autonomous technology more clear and scientific.Building a dedicated lane for autonomous vehicles on expressways is an effective measure to ensure the safety of autonomous vehicles.There are huge difficulties in opening up dedicated lanes for autonomous vehicles on urban roads,especially for the current large and medium-sized cities.In the future,urban traffic will inevitably face the practical problem that autonomous vehicles and human-driven vehicles need to share the road.Therefore,it is necessary to study a series of new problems caused in the process of the transformation of the autonomous vehicle from a simple car into a human-like transport participant.The human-like control of autonomous vehicles is one of the ways to improve the trust relationship between autonomous vehicles and other traffic participants.And when driving on the curve,human drivers show more personalized characteristics.For this reason,this paper has studied the human-like trajectory planning,human-like driver model,and human-like steering tracking control of autonomous vehicles on curved roads.The main work is as follows:Firstly,20 novice drivers and 20 experienced drivers are recruited to drive 5 experimental vehicles of different models on 7 urban double lanes with different radii and length of curvatures under different speeds,and a large number of driving data with the steering control characteristics of human drivers(including speed,trajectory,and steering wheel angle)on curves are collected.Then,it is proposed to use virtual landmarks to transform the raw data into state quantities to achieve a unified data representation.Based on this,the characteristics of vehicle speeds and trajectories of novice drivers and experienced drivers are analyzed to provide a data basis for the following modeling research.Secondly,the similarities between the trajectories of experienced drivers,the trajectories of novice drivers,the trajectories generated by RRT,and lane centerline are calculated by dynamic time warping and adjusting cosine similarity.It is found that the similarities between experienced drivers' trajectories are as high as 0.9135;after repeated driving,the average value of adjusting cosine similarity between novice drivers and experienced drivers is close to 0.8930;the trajectory generated by RRT is always along the edge line of the experimental road,which is very different from the trajectories of experienced drivers;the lateral positions of experienced drivers are fluctuant,which are quite different from the lane centerline.Based on the comparison results,this paper decides to study the human-like steering of autonomous vehicles based on the driving data of experienced drivers.Thirdly,according to the stability of experienced drivers' trajectories,the recursive relation of trajectories at neighboring virtual landmarks is represented by binary linear regression model,and the regression coefficients and random error are also calculated.GRNN,Back propagation neural network(BPNN),and Autoregressive moving average model with exogenous inputs(ARMAX)are used to establish the relationships between the regression coefficients and vehicle speed,road curvature,and sight distance,respectively.By comparison,it is found that GRNN model has the highest prediction accuracy,and thus the human-like curve trajectory planning model established by the model-driven method is obtained.However,the above model-driven trajectory planning algorithm cannot solve the problem of automatically identifying the number of neighboring trajectory points with strong correlation and then establishing the relation between them adaptively.In this paper,a more perfect human-like trajectory planning model is proposed by using the data-driven(LSTM neural network)method.The inputs of the model are vehicle speed and road curvature,and the output is lateral position.The prediction performances of LSTM,ARMAX,BPNN,and Nonlinear autoregressive exogenous model(NARX)are compared on the validation data set and test data set respectively.The results show that LSTM has the highest prediction accuracy and the LSTM-based trajectory planning model also has good generalization performance.Fourthly,3 kinds of preview-based driver models are established in PreScan+ Simulink and multiple steering angle sequence data are obtained by parameter traversal.Comparing the steering angles generated by preview-based driver models with the steering data collected by experienced drivers on the same curved road,it can be found that it is difficult for preview-based driver models to represent the driving habits and driving characteristics of different drivers.Inspired by the data-driven modeling method,a human-like driver model based on recurrent neural network(RNN)is proposed.RNN can not only characterize the preview characteristics of human drivers but also connect the past steering maneuver with the current through memory unit.Based on 3 kinds of recurrent neural networks,three multivariable and multi-step standard RNN,LSTM,and Gated recurrent units(GRU)human-like driver models are established.In the verification period and testing period,the prediction results of LSTM model are closest to the real values.Finally,a steering wheel ‘trapezoidal' angle model is proposed based on the common characteristics of experienced drivers' steering wheel angles in the Frenet coordinate system.After defining 4 feature distances and 1 feature steering wheel angle,the relationship between vehicle speed,road curvature,and feature distance and feature steering wheel angle is established using BPNN,and the trapezoidal angle model is established to represent the input characteristics of experienced drivers' steering wheel angles.Combining it with the electric power steering(EPS)system,the validity of the model is preliminarily determined by simulation in Simulink.The steering tracking experiment is carried out on a real vehicle equipped with EPS.The experimental results show that the EPS system controlled by PID can achieve high-precision tracking of the steering wheel angle.To sum up,based on a large number of real driving data of human drivers on urban curved roads,this paper proposes to use data-driven methods to develop the human-like trajectory planning algorithm,the human-like driver model,and the tracking control method based on steering wheel trapezoidal angle model,which provide the theoretical basis for researches on human-like steering on curves of autonomous vehicles.