Research On Vehicle Point Cloud Data Enhancement Method Based On Panoramic Imagery

With the development of smart cities,3D city models have become the basic data for building smart cities,and they have important significance in urban planning and construction of traffic guidance.The existing 3D urban modelling data is based on imagery and laser point cloud data.The image-based 3D reconstruction of urban imagery is influenced by factors such as the complexity of its street view data,and the large area of shadows and inconsistencies in the scale of close-up images.This makes traditional image matching algorithms difficult to meet requirements,and requires a lot of work and time.With the maturity of laser radar hardware technology,a vehicle-mounted LiDAR scanning system has emerged,which can quickly and directly acquire the dense three-dimensional space point coordinates of the ground features,which to a certain extent make up for the lack of traditional image-based three-dimensional information acquisition methods in photogrammetry.It has become an important data source for reconstructing 3D models,but the laser point cloud data still has self-occlusion of the ground objects or the shadow of the background features blocked by the foreground objects,resulting in incomplete 3D information of the ground objects.The laser scanner in the vehicle-mounted mobile measurement system acquires the laser point cloud data,and the panoramic camera sensor equipped with it also acquires the street view image data.For the problem of missing laser point cloud data,this paper uses the vehicle's panoramic image data to fill in the missing information in the laser point cloud.This avoids the need to collect data at the same time,but also obtains real information to achieve the goal of point cloud data enhancement.The main work and contributions of the thesis are as follows:(1)the vehicle point cloud clustering method is studied,and a vehicle point cloud clustering method based on graph cut is proposed.This algorithm improves the Density-Based Spatial Clustering of Applications with Noise based on Density-Based Spatial Clustering Algorithm(DBSCAN),and combines graph-based global optimization clustering to reduce the missed fraction and guarantees a high level of accuracy rate.Firstly,the point clouds are segmented using the improved DBSCAN algorithm to generate point cloud super voxel blocks.Then the super voxel is used as the node to establish the graph model.The global optimal clustering is performed using the graph segmentation method to obtain the final clustering result.Based on the typical features of land features for feature classification and identification of missing regions,it is true that point cloud data needs to be enhanced in the target area.(2)Refinement of registration parameters for vehicle-mounted point clouds and panoramic images.The initial registration parameters for vehicle point clouds and panoramic images may be affected by attitude deviations among the sensors within the system,resulting in certain errors.Semi-automated vehicle point cloud and panoramic image registration parameter refinement methods are used to manually select the same point in the vehicle point cloud and panoramic image data as the control point,and Accurate parameter information is obtained by the least-squares adjustment based on the panoramic spherical collinear equation,making the accuracy of the two sets of data registration higher,and providing a good data support for point cloud data enhancement.(3)To study the missing point cloud data filling problem,this paper proposes a vehicle point cloud filling method based on panoramic images.The existing point cloud filling method is mainly filled in the missing information based on the scan data hollow hole and the point cloud relationship around the void,and lacks authenticity.Taking into account the different perspectives of vehicle image data and point cloud data acquisition,this paper uses the image to fill in missing point cloud data.First,analyzing and detecting the missing regions of the point cloud,and selecting the image pairs that can be used to fill in the sequence images.Then,aiming at the problem of serious distortion of panoramic imagery and similarity of streetscape architectural textures,affine transformation is performed on the subregions of the panoramic image,and the local imagery block after the transformation is subjected to intensive matching of regional growth based on local geometric constraints,and a 3D point is filled and generated,finally the point clouds is integrated to achieve the purpose of enhancing the original point cloud data..