Road Marking Extraction Based On Registration Of Images And Point Clouds

With the fast development of autonomous driving systems in both the academic and industrial worlds,there have been great needs of high-definition(HD)maps.The mobile mapping system(MMS),equipped with cameras,laser scanners,and navigation systems,can capture high-resolution images?highly-accurate point clouds,and trajec,tory at driving speed.The MMS provides rich textures and accurate 3D geometry in the road scenes,serving as an ideal data source for the generation of HD maps.As a key part of HD maps,the road markings play an important role in various tasks,such as self-localization,behavior prediction,and route planning.Therefore,this thesis aims at road marking extraction that is based on registration of images and point cloudsTo make most of the data captured by the MMS,this thesis first conducts the road marking segmentation in the MMS images and point clouds,respectively,and then carries out the registration between the camera and laser scanner based on that segmentation.This thesis proposes a novel network that has an encoder-decoder ar-chitecture and combines lateral connections,residual learning,and a message passing scheme to learn the details and spatial structures.Also,this thesis uses a loss that combines binary cross entropy and Dice loss to handle the class imbalance.The IoU reaches 73.51%on the ApolloScape dataset,which is close to the state-of-art networks and shows better computational efficiency.This thesis takes two steps to segment the road markings from the point clouds the quick segmentation of roads and the road marking extraction based on intensity This thesis proposes a course-to-fine method of quick road segmentation and conducts intensity correction to eliminate the inconsistency of intensity.Then,an intensity image is obtained using the segmented road points and a self-adapting thresholding method is proposed to extract road markings from this intensity image.The experiments show the effectiveness of the proposed methodsBecause the initial extrinsic parameters provided by off-line calibration cannot match the images and point clouds accurately,this thesis proposes a registration method based on road markings extracted by the above two methods.First,the road marking points are projected to form an image that has the same size of the MMS image based on the initial extrinsic parameters.Then a cost function is established to describe the mismatch between the projection image and the corresponding binary MMS image of road markings.We use particle swarm optimization to obtain the optimal estimation of extrinsic parameters.