Decision Modeling Of Lane Changing Behavior Of Drivers With Different Styles Based On Imitation Learning

With the rapid development of software and hardware technology in recent years,the trend of the new four modernizations of automobiles-"intelligence,networking,electrification,and sharing" is becoming more and more obvious.The intelligent technology of automobiles led by autonomous driving can improve the The advantages of driving safety and vehicle traffic efficiency have received widespread attention in the industry,and good progress has been made,and commercialization is being accelerated.Among the many functional modules of the automatic driving system,the decision-making planning module plays an important role.It comprehensively processes the information of perception and prediction,and sends instructions to the underlying control unit,which is equivalent to the driver’s brain,in ensuring driving safety.At the same time,it is also necessary to enable drivers and passengers to have a better riding experience and a variety of feelings.In more complex driving scenarios,traditional decision-making and planning algorithms based on rule-based design are difficult to reflect the diversity of drivers’ driving styles,and also make passengers intuitively feel that smart cars are not smart,but based on data-driven and deep learning The method,because it is beneficial to solve complex decision-making problems caused by high-dimensional input and strong uncertainty,has attracted the attention of the academic community in recent years.This paper takes the more complex and dangerous left lane changing(ie low-speed lane changing into high-speed lane)scene in the expressway as the starting point,considers different styles of driver driving strategies,and uses the imitation learning method in deep reinforcement learning based on the natural driving data set.The driver behavior decision model is modeled and relevant analysis is carried out.The main research contents are as follows:(1)Establish a left lane changing dataset of drivers with different styles.Through the comparison of different open source natural driving datasets,the German high D dataset was selected as the original dataset,and the data in the original dataset was converted to increase the amount of data in related scenes.Based on the lateral vehicle speed threshold and the ratio of the vehicle centerline deviation from the road centerline to the vehicle body width,the extraction conditions of the vehicle’s left lane-changing trajectory are designed,and the average vehicle speed,average acceleration,average jerk and other trajectory information are selected as features.The extracted trajectories are classified by principal component analysis and K-means++ clustering,and the style of the dataset is divided into two categories:aggressive and conservative by indicators such as silhouette coefficient.(2)Build an anthropomorphic driver policy network.Considering that in many current reinforcement learning modeling research on strategies,most of them take the current environmental state as input,while most experienced drivers take the current moment and previous historical observation sequences as input,so the structure of the strategy network is studied.Targeted design.Taking the self-vehicle historical observation sequence 2s before the current moment as the input of the strategy network,the long-term sequence input is encoded by the bidirectional long short-term memory(Bi-LSTM)network and the attention mechanism,and the vehicle is generated through the processing of the fully connected layer.The lateral and longitudinal acceleration of the trajectory.The anthropomorphic performance is improved from the structure of the policy network,so that the action output of the policy network is more similar to the action of the real driver in this situation.(3)Based on the methods of behavioral cloning and generative adversarial imitation learning,the training efficiency and generalization ability of the model are improved,and the influence of the driver’s driving style on the overall model is compared.The overall decision-making model training framework is mainly designed to generate confrontation.Considering the training of the original GAIL model,because the generator needs to play against the discriminator and constantly interact with the environment to improve its own strategy,the training efficiency is slow,so that the use of behavior For the offline learning characteristics of clone,the policy network is pre-trained to a certain extent,and then a training policy regulator is designed to make the policy network transition to the GAIL algorithm,and the PPO algorithm is used to further improve the policy performance.Finally,compared with other imitation learning strategy networks from the level of learning efficiency,safety,single trajectory and overall data distribution,and used the data of drivers of different styles for training,and compared whether to consider the driver’s style on the strategy model of this paper.impact and model demonstrations in the highway-env simulation environment.