Research On Complex Scene Segmentation And Modeling Based On Supervoxel

Three-dimensional laser scanner is very popular in the field of surveying and mapping because of its high accuracy and abundant information,which can collect surface informat ion of ground objects in a large area.Three-dimensional point cloud data can be used for object classification,feature extraction,three-dimensional modeling,and measurement to obtain relevant information.Therefore,it is very important to optimize the original data and become one of the research hotspots today.Therefore this paper mainly focuses on point cloud denoising,subsampling,deletion of irrelevant data,point cloud segmentation and 3D modeling based on point cloud,in order to improve the data quality and the effect of segmentation and modeling.The specific research contents and achievements of this paper are as follows:(1)The principle of point cloud generated by 3D laser scanner and the work flow of outdoor data collection are described in detail.Then the basic knowledge of point cloud organization and management and geometric information estimation is introduced.Finally,the traditional point cloud segmentation algorithm is introduced,and the experimental verification of data collection is carried out,and its advantages and disadvantages and its application scope are analyzed.Aiming at the problem that the European clustering algorithm is not ideal for data segmentation with ground points,this paper proposes to remove ground points by elevation and optimize the effect of European clustering segmentation.(2)Rough points of three-dimensional point cloud and deletion of redundant information.Firstly,the reflection intensity information of point cloud is used to remove most of the noise,then statistical filtering is used to remove the residual noise,then the sampling algorithm under octree is used to remove redundant points,and finally the improved region growing algorithm is used to remove pedestrian data that accidentally intrudes into the scanning area and noise groups that are difficult to remove.(3)Segmentation of outdoor scene by Supervoxel combined with improved region growing algorithm.Firstly,the Supervoxel over-segmentation algorithm is used to segment the data into Supervoxel with regular shapes and clear boundaries.Then,for each region to be merged,the elevation difference between the seed Supervoxel and its first adjacent Supervoxel is calculated first,and the region is predicted as an uneven surface or plane.Then,the similarity measure fusion growth between the current cluster region and the adjacent Supervoxel is carried out by using different normal vector angle thresholds and automatically calculated orthogonal distance thresholds.If a new Supervoxel meets the similarity condition,its geometric information will be updated after it is added to the current cluster area.The experimental results show that the algorithm in this paper has a good clustering effect for rough surfaces,such as walls,stairs and rockeries.Finally,in order to meet the needs of subsequent 3D modeling and mapping,the RGB color information of the point cloud is also saved during segmentation.(4)Research on 3D modeling based on point cloud data.Firstly,the features segmented by point cloud are extracted.Secondly,Geomagic Studion software is used to model them separately and RGB information is used to generate texture images for mapping.Then City Engine is used to model the objects with poor segmentation effect separately and select suitable plants from its own plant model library to add to the scene.Innovation of this paper:(1)In view of the fact that statistical filtering can’t completely remove the dense noise cluster,this paper proposes to use reflection intensity to reduce noise density first,and then use statistical filtering to remove the remaining noise.For the pedestrian point cloud generated by passing through the scanning area by accident,this paper also improves the region growing algorithm,so that it can manually select seed points to cluster and grow to delete the pedestrian point cloud.(2)To solve the problem of poor segmentation effect of complex scenes,this paper first uses hyper-voxel over-segmentation algorithm to over-segment data into hyper-voxels,and then uses region growing algorithm to cluster and grow them.In the process of growing and clustering,the effect of hyper-voxel fusion is poor,and there is a phenomenon that objects are divided into multiple categories in the segmentation results.Therefore,the region growing algorithm is improved so that the algorithm can correctly segment and extract objects.