Research On Automatic Driving Image Simulation Technology Based On Data Driven

With the development of autonomous driving,autonomous driving simulation test technology has received more and more attention.The image sensor among them has been favored by autopilot system developers because of its low price and abundant information.The image algorithm has grown rapidly,and the demand for simulation testing of the image algorithm has also skyrocketed.There are currently two image simulation test methods.The first is to use real data collected by the camera for simulation testing,but this method is limited in test scenarios and it is difficult to change driving conditions;the second is to use simulation software for testing.Although this method is free and flexible in scene construction,and can easily change driving conditions,the image obtained through virtual sensors has a serious CG(Computer Graphic)style.This article will be driven by real data,using image separation and restoration,image synthesis,and image style transfer technology to achieve free combination of real data and style changes,not only to ensure the authenticity of image data,but also to freely build scenes and choose freely according to simulation requirements The algorithm is tested under driving conditions,and the superiority and effectiveness of the method in the image simulation test are verified through the data set test and the test of the vehicle in loop system.This article first realizes the free construction of simulation scenarios.Use image separation technology to separate the real image,and use image restoration technology to fill the classified image holes to obtain two databases of pure background image and dynamic traffic participants,and then based on the Poisson formula to the original image synthesis method Improvements have been made to solve the problem of color bleeding and achieve seamless integration between static background and dynamic traffic participants,which will help us build a suitable simulation scene according to simulation requirements.Then realize the free change of driving conditions,build an image style migration network based on UNET,construct a suitable loss function,and use the processed image data to train the network,and realize the image migration from day to night,sunny to rainy,and summer to winter Work,so we can test the algorithm’s robustness under a variety of driving conditions such as day,night,rain,etc.In the experimental part,we use the yolo-v3 target detection algorithm to test the data set obtained by this method.Compared with the real data and the data obtained by the simulation software,the three key indicators are significantly improved compared with the data set of the simulation software,and are closer to the real data indicators.Compared with the data synthesized by the original Poisson synthesis method,the three key indicators are also slightly improved.Information entropy is introduced to evaluate the effect of image style transfer,and compared with the image collected by simulation software.Finally,the real vehicle in the loop test platform is built,and the typical driving environment is built.The ACC algorithm is tested under the two driving conditions of day and night,and the corresponding speed curve and distance curve are obtained and analyzed.The results are in line with the expectations,which proves that the data generated by this method can be used in the real vehicle test.