Research On Building Segmentation Algorithm Based On 3D Point Cloud

In recent years,with the continuous development of 3D point cloud data processing algorithms,point cloud segmentation has gradually become a research focus.Compared with 2d images,3D point cloud data contains more spatial structure information and geometric feature information.It can reflect the real existence form of objects in space,and it is easier to analyze their spatial structure and realize semantic scenes.To correctly realize the scene semantics,a semantic segmentation operation is needed to separate different objects in 3D space,laying a foundation for subsequent 3D point cloud data processing.It is an important research topic in the field of point cloud segmentation to optimize the point cloud segmentation algorithm and improve its efficiency.Based on the traditional point cloud segmentation algorithm,this dissertation has carried out the following research and practice:Firstly,the relevant theories and technologies of 3D point cloud data processing are reviewed.The combined filter in the PCL library is used to filter the data.Then several index structures and search methods of the point cloud are introduced,and the k-D tree structure is selected to realize the search of point cloud data.Finally,the local plane fitting method is introduced to solve the normal vector of the point cloud,adjust the direction of the normal vector uniformly,and obtain the curvature information of the local surface where the point cloud data is located,which lays a foundation for the subsequent point cloud segmentation.Secondly,aiming at the problem that the classical RANSAC algorithm has too much computational cost in solving cylindrical parameters and its idea of random sampling.A point cloud cylinder segmentation method based on prior sampling consistency is proposed.The method firstly calculates the initial interior-point probability of each data point according to the prior information of 3D point cloud data and selects the two sample points with the highest probability as the initial sample subset to fit the initial model.Then,geometric constraints were used to pre-check the model,and the boundary loss function was used to judge the model quality of the pre-checked model to update the optimal model.Finally,the interior-point probability of sample points is updated by Bayes’ theorem,and the next iteration is carried out to continuously optimize the interior point set,and the optimal model is obtained.Experimental results show that the proposed method has certain stability and can segment better cylindrical models with fewer iteration times and time,and the segmentation effect is better when the external dot rate is higher.Finally,the traditional plane segmentation algorithm leads to the phenomenon of plane under-segmentation and over-segmentation.A 3D point cloud plane segmentation method based on improved region growth is proposed.Firstly,all data points are classified based on the dimensional characteristics of points,and all two-dimensional surface points are selected.The one with the lowest curvature value is selected as the seed point each time.The seed point area is grown according to the dimensional characteristics,point surface distance,and growth angle.Until the difference between the curvature of the selected seed point and the minimum curvature is greater than the curvature threshold,the seed point search is completed,the growth is stopped and the initial plane is extracted.Then,the approximate surface combination is carried out according to the point cloud normal vector,the distance between the normal vectors of the surface,and the proximity distance of the surface.Finally,through the distance from the point to the plane and the number of neighborhood points,the non-two-dimensional points are reassigned.Experimental results show that the proposed algorithm performs well in plane segmentation and can segment target point clouds from dense point clouds with high precision,which avoids the phenomenon of under-segmentation and over-segmentation to a large extent.