Research On And Application Of Deep-Learning-Based End-to-End Autonomous Driving Technology

The development of deep learning has improved the ability of computers to process images and expanded the applications of computer vision.At the same time,end-to-end autonomous driving technology based on vision tasks is constantly developing.However,deep learning is currently relatively shallow in the field of end-to-end autonomous driving.And decision-making models based on deep learning technology are poor in completing autonomous driving tasks.Therefore,the study of decision-making model based on deep learning has important theoretical significance and application value.Based on deep learning technology,this thesis designs a model that can predict driving actions based on road images and conducts an experimental simulation test on a simulator.Finally,based on the model,an end-to-end decision model application system was designed and implemented.The main work of this thesis is as follows:1.A decision model based on deep learning technology is proposed.The model is based on convolutional neural networks and fully connected networks.The model learns and predicts driving behavior through adversarial training.Combining behavior cloning and generative adversarial imitation learning loss items,the loss function of network training is improved;improved the network structure by adding a reward function to enhance the reward signal;use proximal policy optimization algorithm to update the decision model.The model has a simple network and strong driving ability.It can simultaneously predict three driving behaviors of steering angle,throttle and braking.2.The simulation on the simulator realizes the end-to-end autonomous driving task.First,write code to obtain vehicle dynamics information.Then use the decision model to control the car in the simulator to complete the autonomous driving task.The simulation experiment results show that the decision-making model proposed in this thesis can complete the tasks of going straight,turning,changing lanes and overtaking on racing roads.3.Designed and implemented an end-to-end decision model application system.The system includes four modules: user management,driving behavior prediction,model management and data management.Among them,the driving behavior prediction module completes the function of predicting the steering angle,throttle and braking based on the road image;the user management module realizes the user’s identity authentication,authorization and other functions;the model management module provides users with model viewing and downloading services;the data management module provides users with data download and calibration functions.