Research On Autonomous Driving Decision Algorithm Based On End-to-end Learning

As the number of cars continues to increase,autonomous driving technology has played a key role in solving road safety and other issues.Aiming at key technologies such as navigation and positioning,route planning,behavior decision-making and vehicle control in autonomous driving,this article conducts research on the decision-making part of autonomous driving.With the rapid development of artificial intelligence algorithms,the end-to-end decision-making of autonomous driving achieved through deep learning algorithms and deep reinforcement learning algorithms has successfully simplified the structure of traditional autonomous driving systems.This article will practice,improve and verify the automatic driving decision-making algorithm based on end-to-end learning in real driving data sets and virtual driving environments.First of all,the existing end-to-end driving decision-making algorithms mostly use image sequences as input,and directly predict the steering wheel angle of the vehicle through the neural network model,but this method cannot complete the task of automatic vehicle driving.Therefore,this paper proposes an end-to-end driving decision model based on deep learning,which weights the visual features extracted by the deep convolutional neural network,and establishes an end-to-end system to learn anthropomorphic decisions in the process of autonomous driving.This system is composed of VGG16 network,cyclic neural network and fully connected network.The input of the system is a sequence of unmanned vehicle first-view images and vehicle speed sequence that are continuous in time,and the output of the system is two floating-point numbers,which represent the prediction Steering wheel angle and speed of the vehicle.In order to improve the prediction accuracy,an attention mechanism is added to the end-to-end driving decision model based on deep learning,and the effectiveness of the model is verified in the Comma.ai driving data set.In the longitudinal speed prediction link,the end-to-end driving decision model based on the attention mechanism reduces the average absolute error by 71.9% compared with the longitudinal speed prediction model.In addition,the end-to-end driving decision model based on the attention mechanism is compared with no attention mechanism.The average absolute error of the model is reduced by 7.93%.Secondly,a driving decision-making controller based on Deep Reinforcement Learning(DRL)is designed,and the control barrier function(CBF)is added to limit the vehicle motion output by the DRL controller,which improves the driving of the vehicle.safety.Using Deep Deterministic Policy Gradient(DDPG)as the end-to-end driving decision-making controller,the first-view image sequence of the unmanned vehicle and the 2D image sequence of the lidar point cloud projection are used as input.The output is the accelerator and brake pedal opening and closing degree of the vehicle and the steering wheel angle.If the vehicle’s lateral control actions cannot ensure that the vehicle is traveling in a safe area,the vehicle’s lateral control commands that meet the conditions of safe driving are solved through secondary programming.During the experiment,a reward function based on simulation environment information and related traffic rules was designed,and the test was completed in the CARLA driving simulation environment.After the same training period,the DRL-CBF controller based on end-to-end learning proposed in this paper can complete the vehicle driving task.And to ensure basic driving safety,the average step length of the DRL-CBF controller far exceeds the end-to-end driving decision controller based on DDPG,SAC,and DQN algorithms.In the same test scenario,the fluctuation time of the vehicle lateral control based on the DDPG-CBF controller is shorter than that of the SAC-CBF controller,and the vehicle speed controlled by the DDPG-CBF controller is higher and stable.