Research On Automatic Driving Model Based On End-to-End Neural Network

In recent years,with the development of the automobile industry,while the number of global motor vehicles has shown a blowout growth,self-driving cars have gradually become an emerging hot spot.Autonomous driving technology will help us improve driving safety,manage exhaust emissions,and ease traffic pressure.This has brought far greater changes than expected,Domestic and foreign automobile manufacturers,university research institutions and giant technology companies are constantly increasing the layout of research and development of automatic driving technology.For the development of autonomous driving technology,the most important and most difficult part of the industry is the vehicle sensing and positioning part.Li DAR and millimeter-wave radar are used to realize vehicle positioning and environmental perception,but problems such as weather environment sensitivity,non-coded laser mutual interference,recognition speed are subject to radar rotation refresh rate,and the cost of the entire vehicle have also come.As we all know,in the daily driving process,human drivers only need to collect visual information to complete the control of the vehicle.Under the premise of obstacles to the safety of autonomous driving,reducing the use of sensors as much as possible,so that autonomous driving systems can more intelligently learn how humans drive,and simply using image information to control vehicles is a topic of long-term interest for autonomous driving scholars.The problem is to conduct in-depth research from the two aspects of image information perception and human driving behavior learning combined with the deep learning End-to-End network.First of all,the existing End-to-End neural network model and its practical application are researched and analyzed to discuss whether its network characteristics can meet the perception and control requirements of autonomous vehicles.From the birth of the End-to-End neural network to Analysis of the advantages and disadvantages of its internal structure development and optimized replacement,and then Google used the End-to-End neural network model to control the robotic manipulator in actual engineering,and Princeton University introduced the End-to-End neural network idea of Deep Driving autonomous driving.A comparative summary of models.A feasibility study of physical interaction of unmanned vehicles using unsupervised learning through simple video prediction is proposed.Finally,an input is an image,and the output is throttle,brake,speed,and rotation angle.End-to-End neural network an autonomous driving model for human driving behavior learning.Secondly,in order to improve the training effect of the neural network,in the data processing level,this paper performs image data feature enhancement and two types of manipulation data regression models(multivariate linear regression,Sub-linear correlation).At the level of network structure optimization,this article focuses on the learning rate,activation function,and convolution initialization parameters,and conducts comparative experiments to finally select the Adam learning rate algorithm,Elu activation function,and HE-elu parameter initialization method.Finally,the dcgan(depth countermeasure)deep convolution neural network is used to train and regenerate the existing image data in the data set enhancement level,and the training accuracy of the neural network is improved by expanding the data set.By comparing the effect of neural network model before and after optimization,rsme and driving simulation software are used to verify the optimization effect of neural network.The effects of the optimized end-to-end(end-to-end)and dcgan(depth countermeasure neural network)autopilot model are analyzed and discussed in terms of the accuracy of neural network and the statistics of the modification intervention times in simulation experiments.Finally,it concludes that this automatic driving system can learn human driving behavior and predict neural network by collecting only one type of data of the image,and finally achieves the system expectation of realizing the automatic driving function on the simulated driving system.