Graph Matching Based Scene Understanding And Reconstruction For Point Cloud Scene

The semantic understanding for point cloud scene is one of the key research subjects in computer graphics and stereoscopic vision field.And in which,the semantic information analysis of objects in 3D scene have important application in virtual reality,scene modeling and intelligent housing system.However,the existing method for scene understanding only focus on the location and category,and ignores the shapes of objects in the scene.Therefore,based on the fact that complex objects in real scene usually can be constructed by performing Boolean operations on standard geometry primitives,we propose a graph matching based framework for scene understanding.We take the method of “segmentation first,combination then” to extract and construct the solid objects in the point cloud scene efficiently.The main work of this paper includes adjacent topology graph construction and object extraction and reconstruction based on graph matching.First,the adjacent topology graph is constructed.As the input point cloud scene scanned by Kinect is coarse,we first regularize and thin the input data by voxel filter.Then,we exploit region growing segmentation method to partition the scene into different point cloud patches.The shape and parameters of each patch are estimated by RANSAC(RANdom SAmple Consensus).Finally in this section,we define the minimum distance function to determine the adjacent relation between different patches.Taking each point patch as a node in the graph and the adjacent relation as an edge,we construct the adjacent topology graph for the input point cloud scene.The construction of adjacent topology graph finishes the first round of screening,and excludes the most impossible patches which would not generate a primitive accompanied by current patch.Second,the object extraction and reconstruction based on graph matching is realized.We establish the graph models for common geometry primitives,including sphere,cone,cuboid and cylinder,and the adjacent relation and location of facets on the primitive are represented by an undirected graph.Then,the adjacent topology graph is matched with corresponding graph model to obtain the maximum common subgraph,and the matching rate function is defined to estimate the similarity of the matched subgraph with a standard graph model.Later,the subgraphs with larger matching rate are chosen to generate corresponding geometry primitives,and the shapes and parameters of these geometry primitives are estimated by the parameters of patches on their surface.Finally,the CSG(Constructive Solid Geometry)method reconstructs complex objects in the input scene using geometry primitives with parameters.The graph matching and selection of matched subgraphs finish the last two rounds of screening process,and finally generate the object models in the scene.With four scenes including different objects,it has proved that,on one hand,our method for objects extraction and reconstruction avoids the impact of data defects such as holes and sparse data during point cloud segmentation and RANSAC process,therefore,the final parameters of geometry primitives are of high accuracy.On the other hand,for large areas of data deficiency,our method estimates the optimal shape parameters with limited information,showing the superiority and robustness of the graph matching based object extraction and reconstruction method.