Research On End-to-End Automatic Driving Method Under Typical Urban Operating Conditions

With the development of deep learning,it has made remarkable achievements in computer vision,and end-to-end automatic driving technology based on computer vision has also made great progress.However,the general end-to-end automatic driving method only predicts the steering angle at the current moment,it is difficult to complete the horizontal and vertical control of the vehicle,or predict the steering angle and the vehicle speed at the same time,but it is not well mapped,resulting in a prediction accuracy that is not as good as the single target And the working conditions are often in simple scenes with few people and cars,and there is a lack of research on other working conditions in the city.In response to such problems,this dissertation selects several typical urban working conditions for research,and A dual-target end-to-end autonomous driving model fused with image denoising model is proposed,which can accurately predict the steering wheel angle and vehicle speed at the same time.main tasks as follows :(1)The traditional end-to-end automatic driving method hardly enhances the data set,especially the image data,and as the cost of the image equipment decreases,the imaging quality will definitely decline.,Generative Adversarial Networks)added to several common end-to-end autonomous driving models,using image denoising as a sub-task of the model,and the effectiveness of the image denoising model added to the end-to-end automatic driving model was verified in actual measurement;(2)Traditional end-to-end automatic driving does not make good use of and map vehicle behavior information,resulting in low accuracy of prediction.For this problem,this dissertation adds vehicle behavior information(vehicle speed and steering wheel angle)to the convolution In the output of the neural network,input the long short-term memory(LSTM,Long Short-Term Memory)at the same time,the time series mapping relationship between the vehicle behavior and the image is added,and an end-to-end automatic driving model is designed based on this;(3)The accuracy of dual-target prediction(vehicle speed and steering wheel angle)is often not as good as that of single-target.This dissertation proposes a dual-target loss function that adds a single-target loss influence factor to model the single-target loss of vehicle speed and steering wheel angle It is added to the dual-target loss function according to a certain weight and added to the designed end-to-end automatic driving model to form a dual-target end-to-end automatic driving model.The test and comparison are conducted through the actual vehicle test set to verify the prediction performance of the model;(4)After completing the design of the image denoising model and the improvement of the traditional end-to-end model,this dissertation fuses the image denoising model with the dual-target end-to-end automatic driving model,and draw on the mobileNet lightweight network method to The model is lightened to improve the running speed of the model.Finally,the original test set and the test set of another city are used for model testing to verify the performance and versatility of the fused model.The end-to-end automatic driving researched in this dissertation is based on the typical urban working conditions.Through the combination of convolutional neural network and long short Term Memory Network,the time-space sequence relationship in the image is extracted,and a fusion loss function is proposed to improve the The end-to-end automatic driving algorithm’s dual-target prediction accuracy,and considering the reduction in image quality caused by the cost reduction of camera equipment,an image denoising model is fused in the end-to-end automatic driving model,which is Under the circumstances,the application of more complex scenarios provides a few reference.