| At present,autonomous driving is a hot topic in the field of artificial intelligence.The ultimate goal of autonomous driving is that,without the intervention of human being,the vehicle itself could drive according to the given route by the interaction among onboard sensors,computing units,road-side units,etc.Currently available commercial autonomous driving software is quite complex,which leads to high development and maintenance costs.In contrast,the end-to-end system based on deep neural networks does not deal with separated modules such as perception,planning,decision-making,and control,but consider the system as a whole to optimize,directly mapping the information from the sensors to control command.Currently,the end-to-end autonomous driving system attracts more and more attention from the academic and industrial areas.To improve the generalization ability and driving smoothness of the end-to-end systems,this thesis has done the following work:First,this paper designs an autonomous driving system based on reinforcement learning and visual feature extraction.Lane lines are first extracted from the images captured by the on-board camera,such that the environmental noises such as sky and tress are eliminated.Then feature reduction is performed through the variational autoencoder,and only the key information in the original image is retained,which is sent to the reinforcement learning for training.The test based on the simulation platform shows that the autonomous driving system designed in this paper can not only complete the lane keeping task,but also has good generalization ability.Even the road scene is changed,the vehicle can still drive normally.Second,to make the driving more smoothly,this paper proposes an end-to-end autonomous driving system based on LSTM(Long Short-Term Memory).After some preprocessing,the image and corresponding historical speed sequence as inputs are sent to the pre-designed network structure that combines the convolutional neural network and LSTM,which then predicts the rotation angle and the discrete speed command.The test based on the simulation platform shows that the autonomous driving system designed in this paper could perform the lane keeping task,and the driving operation also tends to be stable. |
