Research On End-to-End Lane Keeping Method Based On Conditional Imitation Learning

As artificial intelligence continues to evolve,intelligent car companies are mushrooming onto the autonomous driving scene and autonomous driving technology is evolving at an ever-increasing pace.While traditional mathematical lane keeping methods are relatively mature,the application of neural networks in the field of autonomous driving is becoming more and more popular.This dissertation focuses on lane keeping methods in autonomous driving technology and explores the effectiveness of end-to-end methods for lane keeping based on conditional imitation learning.(1)Based on the traditional end-to-end lane keeping approach,which mainly uses RGB images as input data,the neural network is unable to extract deeper feature information,resulting in poor model prediction capability.In this paper,we propose a research method that uses RGB images,semantic segmentation depth images and control volume information together as input and adds advanced control commands before the action value output to improve the robustness of the model.(2)The end-to-end learning approach proposed by NVIDIA is a milestone in end-to-end application in autonomous driving technology.In this dissertation,a 9-layer convolutional neural network is built to extract image feature information and trained to test its lane keeping capability on Udacity simulated roads.As it was with only RGB image input,the Udacity simulation vehicle collided on the test road and failed to complete the driving task in full.Although the end-to-end approach was proved to be able to be applied to lane keeping,the accuracy of the model was low,and the generalisation ability was weak with a single data input.(3)A Conditional Imitation Learning(CIL)network was constructed based on a classical neural network model,and data was collected using a CARLA car simulation simulator built and pre-processed to prevent over-fitting problems.The trained CIL model was later tested to compare its predictive capability with the ResNet network model in straight ahead,turning,and mixed conditions,and to test the lane keeping capability of the car in mixed conditions using the CoRL2017 evaluation criteria,and the ablation experiments to test the RGB,RM,RMD and CIL algorithms in three metrics: average task completion,average distance travelled and average number of violations.The experimental results demonstrate the feasibility and effectiveness of the end-to-end approach based on conditional imitation learning for lane keeping.(4)To further validate the effectiveness and generality of the end-to-end lane keeping algorithm based on conditional imitation learning,its pre-trained model was loaded into the DDPG(Deep Deterministic Policy Gradient)algorithm,and the adaptability of the CIL model was verified through the training and evaluation results;in addition,the experimental data was collected by building a Raspberry Pi intelligent car platform,and the images were semantic segmentation processed,and the CIL network was used for training and testing,and the results demonstrate that the end-to-end approach based on conditional imitation learning has the ability to enable vehicles to complete lane keeping.