The Design Of End-to-end Autonomous Driving Model In Virtual Environments

With the development of the automobile industry and urbanization,the number of cars in cities around the world continues to increase.This will bring serious traffic and safety problems.The autonomous driving is adopted to alleviate the above problems,as well as to improve the travel experience of the passenger so as to bring enormous economic benefits.By learning the mapping relationships between the perception and behavior based on expert demonstrations,the imitation learning-based end-to-end autonomous driving technology enables expert-like control strategies and simplifies the rule-based modular approach compared with traditional methods.However,in end-to-end autonomous driving,how to improve the perception and decision-making abilities of the network to enhance driving performance is always a hot topic.Motivated by this end,this thesis aims to design an end-to-end autonomous driving model based on the perception-decision integrated architecture and to improve the existing network structure to have a better driving performance.The main contribution of this thesis are as follows:1.An end-to-end autonomous driving model based on imitation learning is proposed using a deep learning framework.The model is mainly based on the environment feature extraction and back-end behavior prediction network,where the back-end behavior prediction network is used as a decoder to output the control signal.Here,this work has improved the decoder by combining trajectory planning and control prediction to retain the advantages of both methods.In addition,environmental features are extracted by using Transfuser to fuse Li DAR and RGB inputs,which can reduce 33% of collisions compared with the method that extracts RGB features using Res Net network only.2.To handle the problem that the input state features of existing models only contain a single moment but cannot reflect the dynamic changes of driving behavior,this work additional involves a temporal feature extraction network by taking historical continuous state information as input and fuses temporal features by using a GRU network.In this way,the state information such as driving speed,steering angle,and target point can help self-driving vehicles to make better decisions.3.To address the problem that the back-end behavior prediction network for future time-step control command prediction only uses the corresponding time-step waypoint features and lacks the ability to think about the future,this work uses a gated self-attention mechanism in the control branch to focus on the current time-step waypoint features as well as other momentary features according to the learned weights,where higher weights reflect important regions so as to further improving driving performance.4.Finally,the effectiveness of the end-to-end autonomous driving model proposed in this work was verified in the CARLA virtual environment.The driving dataset is first collected in CARLA using Roach as the driving expert.It is then feed into the proposed model for training.Finally,the trained model is validated in a closed-loop task and parameter fine-tuning on different town routes in CARLA.Extensive experiments show that the addition of temporal feature extraction self-networks and gated self-attention mechanisms improves driving performance by11.5%,which validates the effectiveness of our designed module.