End-to-end Autonomous Driving Decision Model And System Design Based On Imitation Learning

With the accelerated urbanization in China,the number of motor vehicles is increasing day by day,and traffic and safety problems are becoming more and more serious.Autonomous driving,as a new application of artificial intelligence technology,is expected to alleviate these problems.As a new artificial intelligence technology,imitation learning can learn the mapping relationship between image and control information in expert driving data,and then get excellent expert driving strategies.However,in reality vehicle driving is a continuous behavior,imitation learning simply maps image and control information,and the subsequent actions are not completely continuous,which leads to unstable vehicle driving.Especially in complex scenarios such as intersections,imitation learning cannot select the correct road based on its original structure,while conditional imitation learning based on passenger commands to select the road requires human intervention in the commands and each branch network is trained independently,resulting in inefficient decision making.To address these problems,this thesis proposes end-to-end autonomous decision model and system based on imitation learning,the main work is as follows:（1）Aiming at the problem that the continuity of driverless behavior itself is not considered in the original end-to-end imitation learning,which leads to the instability of vehicle driving,an end-to-end imitation learning lane-keeping method integrating temporal semantic information is designed.This method enhances the extraction of image features through the technology of image semantic segmentation and prior knowledge transfer learning,which improves the accuracy of model prediction.Then,the continuity of vehicle driving is effectively guaranteed by associating temporal information with LSTM network and adding vehicle state information.In this thesis,the test experiment is carried out in the simulation environment.The experiment shows that the proposed method reduces the average prediction error of 24.31% and the average continuity error of 17.99% compared with the original imitation learning,so that the autonomous driving control decision-making system can make more accurate prediction and fit the long-term continuous driving behavior in real life.（2）Aiming at the problems of human intervention and low efficiency of data use in the conditional imitation learning of taking passengers’ instructions as road selection,an end-to-end decision-making model of conditional imitation learning guided by waypoint information is designed.By using the waypoint information to select the road instead of the passenger command input on the conditional imitation learning network,the problem of continuous intervention of passengers in vehicle driving is solved,and the utilization rate of training data and system stability are improved.Overall,the model can learn driving strategies with less data and realize end-to-end automatic navigation.Based on the simulation environment,this thesis tests the driving performance in different urban complex scenes,verifies the feasibility and effectiveness of the proposed method,and is superior to the existing conditional imitation learning in the number and quality of unmanned tasks.