Research On Vision-based Lateral And Longitudinal Positioning Of Autonomous Vehicle

In recent years,the rapid development of autonomous driving technology has greatly promoted the intelligent transformation of automotive technology.However,the high cost of sensor hardware has seriously hindered the application of autonomous vehicles.Therefore,how to use the low-cost sensor to achieve autonomous driving in the specific scenarios has become a matter of great concern to academics and industry.Accurate perception and positioning is the foundation of the application of autonomous driving.Therefore,the research on vision-based perception and positioning is of great significance for accelerating the low-cost application of autonomous vehicles.Focusing on the structured road environment,this paper proposes a vision-based low-cost solution to achieve the lateral and longitudinal positioning of autonomous vehicles.The main research contents are as follows:1)The basic software and hardware platform are established for the research of vision-based lateral and longitudinal positioning.Among them,the basic hardware platform is composed of an autonomous vehicle and the relevant sensors of this study,the basic software platform is composed of several modules such as data reading,data storage,data transmission and parameter calibration of various sensors.Based on the basic software platform,this paper designs the experimental data collection method of related sensors,and completes the parameter calibration of the camera.It provides a platform of data acquisition and experimental verification for the follow-up research.2)This paper focuses on the research of lateral positioning based on the lane feature extraction.First of all,in order to solve the problem that lane feature extraction is susceptible to environmental noise,this paper proposes a method based on the fusion of gradient features and color features to complete lane feature extraction.Secondly,aiming at the problem that the measurement accuracy of angle parameter in the classical vision-based positioning model is difficult to guarantee,this paper proposes an improved lateral positioning model,which only needs to measure two distances and optimize the calibration process of the positioning model parameters.Finally,based on the improved model,a lateral positioning method based on lane features is proposed.The experimental results show that the average absolute error of the static experiment is 0.048 meters,the average relative error is 2.52%,the average absolute error of the dynamic experiment is 0.051 meters,and the average relative error is 2.67%.It basically meets the need of the accuracy of the lateral positioning measurement in the application of autonomous driving.3)This paper also focuses on the research of longitudinal positioning based on the vehicle detection.Firstly,an algorithm of vehicle detection is developed based on the deep neural network.The training of vehicle detection network is completed by using the dataset constructed in this study,and finally a model of vehicle detection with the ability of classification is obtained.Aiming at the problem that the vehicle width in the longitudinal positioning model is difficult to estimate accurately,this paper proposes two methods to estimate the width of the vehicle: one is to estimate the width of the vehicle in the lane based on the information of lane line detection;the other is to take the average vehicle width of different categories of vehicles as a prior width value,in which the vehicle category information is provided by the vehicle detection model.Based on the above-mentioned methods of vehicle width estimation,this paper proposes a fusion algorithm of vehicle width information,and the fused vehicle width is used to complete the vision-based longitudinal positioning.Finally,the experimental results show that the average value of relative error of the longitudinal positioning is less than 3% within 50 meters in front of the vehicle and less than 5% within 60 meters in front of the vehicle.The method proposed in this paper can effectively realize the longitudinal positioning of autonomous vehicles within a certain range.