Study On Decision-planning Systems Of The Automated Vehicle Based On Driver’s Personality Under Car Following In Highway Conditions

The intellectualization of the vehicle could effectively reduce traffic accidents,relieve congestions,restrain energy consumption,and improve road utilization,which has become a research hotspot in academic and industrial circles.As a typical representative of intelligent vehicle technology,automated driving integrates information,sensing,communication,control,computer,and artificial intelligence technologies,which symbolizes the commanding height of the future automobile technology and is also the recognized developing direction in worldwide.Due to the advantages of unique road structure,clear traffic flow,relatively closed environment and determinative driving rules,automated highway driving is expected to be the first step for the commercial use of automated vehicle.Based on the projects from National Natural Science Foundation of China(NSFC)"Research on automated driving decision planning and robust control based on driver’s decision behavior representation(51875061)" and the key research and development program of the Ministry of science and technology of the people’s Republic of China "Research on the key basic issues of information acquisition,decision-making and control of intelligent electric vehicles,research on human-machine co-driving theory and driving right allocation(2016YFB10100900)",the study on decision-planning systems of automated vehicle based on driver’s personality is carried out considering the aspects of driver behavior characteristics modeling and the motion planning under highway driving conditions.The main contents of this paper are as follows:(1)The characteristics of typical highway driving conditions were analyzed,then typical driving scenarios were designed,and the vehicle state data of the driving simulator and the field operation test was collected.Next,Gaussian mixture model(GMM)based on Bayesian theory was proposed to model the driver characteristics.Focusing on driver’s statistical characteristics,the importance sampling and multidimensional conditional probability density distribution regression were developed.Applying the proposed methods on car following driving conditions,the key parameters reflecting the driver characteristics were extracted and driver modeling and behavior analysis were implemented.On the other hand,to predict the longitudinal motion of environmental vehicles,the Gaussian process estimation method was proposed,which laid a foundation for accurate assessment of vehicle driving safety risk in the process of automated driving decision.(2)To deal with the personalized driving decision under complicated driving conditions,a vehicle collision risk description method based on mixed logic dynamic system(MLD)was proposed.With the help of the strict definition of vehicle collision risk boundary,a decision planning algorithm based on model predictive control(MPC)theory was established,which realized the safe automated driving in complex driving environment.Considering the statistical characteristics of drivers’ behavior,the weight allocation for driving performance objectives based on entropy weight and variance weight was proposed.The automated driving decision-making based on personalized representation of drivers was implemented.The accuracy of personalized decisionmaking was further verified by using driver classification model.(3)In order to realize real-time and efficient automated driving decision-making,an automated driving decision-making method based on geometric path generation was proposed.The characteristics of several parametric curve paths in path planning were studied.Cubic Bézier curve was selected as the ideal curve path.From the spatial perspective,the candidate paths for vehicle driving were generated by sampling the target positions,and the speed planning based on quadratic programming(QP)was developed from the perspective of time.Vehicle safety evaluation method based on geometric intersection detection was proposed.Then the safe driving trajectory of the vehicle was extracted,and the optimal driving path that satisfied the requirements of planning performance objectives was designed by comprehensively considering driving personality,path rationality,trajectory consistency and speed fluctuation.The personalized automated driving decision-making method based on geometric path generation improved the personalized driving performance and decision-making efficiency in the process of automated driving and ensured the driving safety.(4)The application of reinforcement learning(RL)algorithm in automated driving decision planning was studied.The reinforcement learning interaction environment was designed for the car following conditions.The reward function was designed based on the real vehicle test data.For the continuous action space problem,the Proximal Policy Optimization(PPO),Deep Deterministic Policy Gradient(DDPG)and soft actor critic(SAC)algorithms were used to learn driver behavior in car following.The results showed that reinforcement learning can effectively learn and imitate the driver’s decision-making behavior,and PPO algorithm had better computational performance.Based on the optimal strategy of reinforcement learning and the geometric path generation strategy based on rules,a multi strategy switching mechanism was proposed,which improved the problem that the planning based geometric curve generation methods could only make personalized decisions on the statistical characteristics of drivers.The safety evaluation based on rules ensured the driving safety of deep learning decision-making and realized the personalized application of reinforcement learning in the process of automated driving decision-making.