| In the urban traffic scene,car-following is one of the common traffic phenomena,especially in the case of traffic congestion,car-following behavior is more common.In the traditional research of car-following behavior,behavior modeling is based on vehicle driving data.With the advent of the era of data,more and more high-precision vehicle driving data were saved,which directly promotes the development of car-following control technology research.In the scenario of car-following,too long reaction delay will lead to unpredictable and serious consequences.Therefore,capturing the driving behavior characteristic of reaction delay is crucial to the safe driving of car-following.However,the reaction delay is often ignored in the current research.Therefore,aiming at the specific scene of car-following,the theoretical modeling of car-following behavior is carried out in this work,and a car-following model which could automatically estimate the reaction delay is proposed.At the same time,in order to reduce the influence of too long reaction delay on car-following control,an autonomous car-following control model based on deep reinforcement learning was proposed to realize automatic control of vehicles in following scenarios,shorten the reaction delay time and improve safety.Specific contents include:(1)Modeling and analysis of autonomous car-following control.The driving behavior of car-following model is summarized.Then,the driving behavior of reaction delay is taken as the research object,and its behavior characteristics are analyzed in depth.The environmental perception,behavior decision making and car-following control of autonomous driving vehicle are discussed,which lays a good foundation for the design of car-following control model based on deep reinforcement learning.(2)The vehicle driving data set is cleaned and preprocessed.Firstly,the NGSIM driving data set is introduced,and part of the reasons for the error are analyzed.Then,the original data set is reconstructed to reduce the data error.After that,the car-following process of the reconstructed data is extracted and analyzed,which provides data support for the subsequent model establishment.(3)An attention-based Ensemble Learning Car-following(AEL-CF)model is proposed.Firstly,the data cleaning of vehicle driving data set and the extraction and analysis of carfollowing process are carried out.Then an ensemble learning method is proposed and applied to car-following behavior modeling to establish an integrated learning car-following model.In the first mock exam,the attention and encoder-decoder mechanism and expectation measurement method are integrated.The driving process is driven and verified by the data of the car-following process.The result shows that the model could capture the driving characteristics of the reaction delay automatically and improve the shortcomings of the abnormal acceleration of a single model effectively.(4)A car-following control model of autonomous vehicles based on deep reinforcement learning is constructed.In order to realize the automatic real-time control of the vehicle under the following state,a learning framework of car-following control model is proposed,and a car-following control model is established.Aiming at the scene of car-following,the state,action and reward function of the model are designed,and then the simulation experiment is carried out in CARLA simulation platform.In this work,the model is tested in a single lane car-following scene.The results show that the agent could choose the current optimal behavior in different states and obtain the maximum reward value.Compared with human drivers,it greatly shortens the reaction delay time,realizes the automatic driving of car-following vehicles,and proves the effectiveness of the model. |
