Lane Information Detection Method Based On Visual And Map

Driverless technology is one of the hot topics in recent years.Lane detection is the key and prerequisite for Driverless technology.Urban roads are important application scenarios for driverless technology.However,the scene information on urban road is complex and the types of roads and lane markings are various and varied.The diversity of road environments poses a challenge for lane information detection.This thesis focuses on the theme of traffic safety and conducts research on two key issues,namely the acquisition of lane environment depth and the detection of the current vehicle lane information.The research content of this paper is mainly the following aspects:(1)A stereo matching algorithm based on double component model(SM-DCM)is proposed to obtain the lane depth by using the binocular vision system.This method decomposes the image into two independent components,Outlier and Inlier,according to the texture features.Outlier image contains only fine texture such as small trees,power poles,signal poles and ground abandoned objects.Inlier image normally contains large-size planar area such as large area road,vehicles,and buildings.Different matching algorithms are designed based on two kinds of texture characteristics to perform stereo matching respectively.Finally,the results are fused to obtain accurate depth information of the lane scenes.The test results show that this method can obtain a detailed depth map in the driving environment and has a good detection effect on the fine and large surface textures.(2)The lane region is solved by matching the pre-designed road map with the video image.Firstly,the forward view image taken by the camera is transform by IPM according to the estimate value of the attitude parameter,and a top view image with small difference from the map projection is obtained.In order to improve the robustness of map and image matching,an enhancement algorithm is designed for lane lines in the top view image.A method of point-to-point association between map lane line and camera image lane line by group mapping is proposed,and the mapping equation is established by corresponding point pairs.The target equation is solved by RANSAC principle,and the projection transformation parameters between the map and the top view image are obtained.The attitude parameter values are updated and the process is repeated until the algorithm converges.Finally,the map is transformed according to the parameter and superimposed on the front-view image to identify the lane area.(3)A lane detection method under the condition of known driving path is proposed.The driving route map is established by public electronic map or the mapping map,and the images taken by the camera and the route map information are correlated by position information;A simple and effective method is designed to fuse the vector route information with the corresponding camera image;The segmentation model based on depth convolution neural network is studied,and the fused data samples are trained.the trained model can effectively detect lanes.The comparative test shows that the proposed method can make full use of the prior knowledge of the driving path.Compared with the simple visual detection method,the proposed method has better detection effect on the lane of turning,pedestrian crossing and other scenes.