Study On Object Labeling Method Of 3D Scene Based On Point Cloud

The labeling of objects is a prerequisite for the understanding of the scene,which is the basic work of environmental cognition,and has extensive application and research in artificial intelligence,computer vision and 3d reconstruction.In this paper,three-dimensional point cloud scene is used as the research object,and the annotation method of objects in the scene is studied.The main research work is as follows:(1)Based on the idea of octree,a method of scene division was established.Firstly,the overall trend of the model is obtained by PCA method,and the minimum bounding box of the scene is obtained for the segmentation of the initial root node.According to the idea of the octree from big to small and layer by layer,the whole process of dividing the termination condition is calculated.The data of point cloud scene is saved in the partition lattice according to the spatial relation,and the structure of the partition tree is constructed by scene segmentation.Finally,the spatial relation of partition lattice is calculated and the neighborhood information of partition lattice is increased.(2)This paper presents a method for rough labeling of point cloud scenarios.Firstly,based on the idea of hough transform,the plane in the scene is detected,and the relative height of each plane is calculated,and the lowest plane is selected as ground and removed.Then,based on the idea of regional growth algorithm,the segmented lattice of scene segmentation is clustered to classify the data points in the point cloud scene and complete the rough labeling of the scene.For the detailed annotations,the correlation classification of the projection density of the three projective planes is used,and the points with similarity of the normal vectors in the neighborhood are identified by the unmarked points,and the detailed annotations of the scenes are completed.(3)Based on the result of scene rough annotation,a method for the detail annotation of point cloud scene is presented.First three projection plane structure,calculation of the partition lattice in the plane of the density value,and according to the division of regional connecting relation between planes,merger related categories,with the result of coarse marked growth problem solved.For the unmarked points in the object part of the scene,the point of similarity of the normal vector in the neighborhood is identified,and the details of the scene object are finished.(4)Based on labeling is presented,a method of finding and expressing topological relations of objects in the scene.Firstly,the calculation matrix of the topological relation of objects is given according to the four-intersection model.Then,the boundary and interior of the object are calculated according to the nominal value,and the topological relation table is established.Finally,based on the idea of KD tree,the tree topology is established,and the topological relations of scene objects are expressed in combination with topological tables.The experimental results show that this method can be used to classify objects in the scene effectively.At the same time,the method of this paper is used to mark the partition lattice structure,and the processing efficiency of the algorithm is obviously improved.