Pavement Landmarks Detection And High-precision Localization Of Vehicles Based On Vehicle Image

The most commonly used GPS localization is low accuracy and has signal blind area,advanced visual localization technologies such as visual odometer,simultaneously localization and mapping(SLAM)have high computational complexity and poor algorithm stability.At the same time,the construction of road structures and scenes is becoming more and more standardized,which enriches the visual features and geometric structure information of landmarks,and provides necessary reference conditions for vehicle localization.This thesis studies the vehicle localization by constructing a high definition map model,combining with GPS localization and visual localization,the research work mainly as follows:(1)A high definition map model based on pavement landmarks is proposed,which is constructed by vehicle images,high-precision position information and geometric structure information of landmarks.In the construction process,when landmarks exist,on the one hand,the geometric structure information of landmarks is measured by measuring tools,on the other hand,images are collected by vehicle camera,and the high-precision position information is acquired by GPS receiver and DGPS base station simultaneously.Therefore,each landmark contains a high-precision position information,a frame of image information(and visual features extracted from the image)and geometric structure information.(2)A multi-scale localization algorithm based on pavement landmarks is proposed,firstly,the nearest landmark is calculated,and the positional relationship with landmark is calculated.Combining with the three elements of high definition map model,a coarse localization is first derived based on GPS matching.Afterwards,the nearest landmark is calculated based on pavement marking detection by utilizing Histogram of Gradient(HOG)and local feature matching to achieve landmark-level localization.Finally,the metric localization is accomplished by using the geometrical relationship from the pre-built landmark map.The landmarks such as straight arrow,turn right arrow and manhole cover are selected in the experiment in sunny and rainy day respectively.In the sunny environment,the maximum localization error was 16.1cm and the average localization error was 8.4cm.In the rainy environment,the maximum localization error is 23.3cm and the average localization error is 12.2cm.The results show that the proposed method can locate the vehicle accurately and has high robustness.(3)A vehicle localization method based on visual,GPS and high definition map model is proposed,which corrects the results of GPS localization using visual localization.First,the vehicle GPS coordinate is matched with the high-precision position information of the landmark in the map model,then the nearest reference landmark is calculated.According to the conversion relationship between the coordinate systems,the vehicle GPS coordinates are converted into plane orthogonal coordinates under the reference landmark coordinate system.Then calculating distances between the current vehicle to each side of the landmark using the geometric structure information,which overcomes the nonlinear constraint between points,and keeps the Kalman filter in a linear.Finally,the GPS localization results are taken as the prediction data and the visual localization results are taken as the observation data of the Kalman filter.Different types of pavement landmarks are selected for vehicle localization experiments.When straight arrow is used as a reference,the average localization error before and after fusion is 8.313 m and 0.802 m respectively,when turn right arrow is used as a reference,the average localization error before and after fusion is 8.76 m and 0.79 m respectively.The results demonstrate that the proposed method can effectively improve the localization accuracy.