Research On Personalized Decision-making Methods For Vehicle Longitudinal Autonomous Driving

At present,the technology of vehicle longitudinal autonomous driving has been widely studied.In order to improve the comfort of autonomous vehicle and people’s acceptance of autonomous driving system,how to design personalized longitudinal autonomous driving system that is close to human driver’s behavior characteristics has become the current research hotspot.However,the current research has the following problems: personalized longitudinal autonomous driving system is generally established through the personalized design of vehicle driving risk characterization methods and longitudinal decision-making algorithms,while the current design of driving risk characterization methods generally relies on TTC,THW,risk field and other physical models,and the mapping relationship between calculated driving risk and driver action is not ideal,which is not conducive to the personalized design of decision-making algorithms.Deep reinforcement learning has become one of the most widely used methods in the design and research of personalized decision-making methods for longitudinal autonomous driving.However,the reward function of deep reinforcement learning methods is generally designed based on the experience of researchers,and the degree of personalization of autonomous driving decision-making policies obtained is difficult to reach a high level.The decision policies obtained through deep reinforcement learning may encounter safety issues in some scenarios due to incomplete training and other factors.How to fully ensure the driving safety of vehicles is a key issue that must be solved in the practical application of autonomous driving decision methods based on deep reinforcement learning.In response to the above issues,this article focuses on the personalized decision-making task of longitudinal autonomous driving,and the specific research content is as follows:Firstly,in response to the problem that the current vehicle driving risk characterization method does not have an ideal mapping relationship with driver actions,which is not conducive to the personalized design of decision algorithms,a personalized vehicle driving risk characterization method is established.Using the High D dataset,design a driver classification method to classify drivers into three categories:conservative,general,and aggressive,and analyze the behavioral characteristics of each type of driver.Using driving data from three different styles of drivers,the longitudinal acceleration of the vehicle is used as a representation of the driver’s perception and understanding of the vehicle’s driving risk.Based on the vehicle’s kinematic equations,the vehicle’s driving risk characterization values under driving conditions that are not included in the driver data are calculated.Using back propagation neural network to establishing a mapping relationship between driving status and driving risk characterization values,and to establish personalized risk characterization models for three types of drivers.The correlation coefficients are above 0.99,which has good fitting goodness.The variation pattern of risk characterization values reflects the characteristics of different driving styles,laying the foundation for the design of personalized decision-making methods for longitudinal autonomous driving.Secondly,a rule-based personalized decision-making method for longitudinal autonomous driving is designed to ensure the driving safety of autonomous vehicles.Divide rule-based longitudinal autonomous driving decision-making methods into two decision-making modes: following driving and emergency braking,and design mode switching logic based on following distance.Establish the kinematic state equation of the workshop,and based on the difference in following distance among drivers of different styles,establish an acceleration decision-making policy for following driving mode using model predictive control method,to achieve personalized following decision-making.Design a trigger type emergency braking mode to apply emergency braking when the distance between the vehicle and the preceding vehicle is too close in case of sudden emergency situations,ensuring the safety of vehicle operation.Thirdly,aiming at the problem that the personalization level of the current longitudinal autonomous driving system is not high enough and the behavioral characteristics of the longitudinal autonomous driving system are greatly different from human drivers,a longitudinal autonomous driving personalized decision-making method based on deep reinforcement learning is established.Using the output of the driving risk characterization method as the state input of the deep reinforcement learning method,design the structure and parameters of the deep reinforcement learning method,and establish personalized reward functions that conforms to the characteristics of different driver styles.In the simulation environment,train and obtain personalized decision-making policies for longitudinal autonomous driving targeting three different styles of drivers.Compare the differences between the driving trajectories of different individual drivers and three personalized policies with different styles,and identify driver categories.Based on individual driver driving data,the weights of the personalized reward function network are updated using deep inverse reinforcement learning to obtain personalized reward functions for individual drivers.Through deep reinforcement learning training,personalized decision-making policies for different driver individuals are obtained,achieving personalized decision-making for longitudinal autonomous driving.Fourthly,to address the issue of insufficient safety in some scenarios of longitudinal autonomous driving decision-making policies based on deep reinforcement learning methods,a hybrid decision-making method is established by combining deep reinforcement learning based decision-making methods with rulebased decision-making methods,achieving complementary advantages between the two methods.Compare the decision-making effects of two decision-making methods and analyze the advantages of each policy.Using the collision risk and the difference between the expected acceleration values output by two decision policies as fuzzy variables,design fuzzy rules to obtain the decision weights of the two policies,and adjust the parameters of the rule-based decision policy based on the weights.Then,a hybrid decision-making method is established by linearly adding the expected acceleration output of two policy decisions through weights,so that the final longitudinal autonomous driving decision policy can make personalized decisions at all times under sufficient safety conditions.Fifthly,this article uses the Carsim&Simulink simulation platform and the Haval H7 intelligent vehicle experimental platform to verify the methods proposed in this article.In the simulation experiment,the differences between the personalized hybrid decision-making policy established in this paper and the driver’s following trajectory are compared,as well as the differences between personalized hybrid decision-making policies targeting different individual drivers.In the actual vehicle experiment,three personalized hybrid decision-making policies targeting different individual drivers are used for driving tests.The experimental results show that the trajectories of the personalized hybrid decision-making policies are relatively close to trajectories of the drivers.The root mean square errors of the relative distance between the trajectories of the personalized hybrid decision-making policies and the trajectories of the drives are3.41,2.09,and 2.96.The root mean square errors of the relative speed are 0.64,0.48,and 0.55.The root mean square errors of the speed are 0.69,0.74,and 0.74.And there are significant differences between personalized hybrid decision-making policies targeting different individual drivers,which are in line with the corresponding behavioral characteristics of the driver.The decision-making effect of personalized hybrid decision-making policies on real vehicles are consistent with simulation experiments.Proving that the method proposed in this article can effectively achieve personalized longitudinal autonomous driving decisions.