Technical Research About Point Cloud Based Indoor Scene Segmentation

The reconstruction and classification of the three-dimension point could scene has been one of important research project in 3D point cloud data processing.Due to the diversity of scene and the complexity of point cloud data,it is necessary to segment point cloud structure reasonably before scene reconstruction.In this paper,a Gaussian mapping based point cloud segmentation strategy is proposed,aim to figure out the problem about segmentation of permanent building structure in indoor scene.According to the model size and complexity,the method takes by coarse-fine,iterative way,to decomposed the building structure into point cloud clusters which are easy to detect layer by layer.Then the model fitting algorithm which with the prior knowledge of models is proposed,to extract the point cloud accurately,then we can get the final products of point could segmentation.The main innovation works of this paper include the following two points:(1)A model fitting algorithm with prior knowledge of model is proposed——PriorMLESAC.In this paper,the model fitting method based on random sampling consistency and its diffraction algorithm are introduced,and the problems about exist algorithms are analyzed.In order to solve the uncertainty about the threshold and the prior probability,based on the original algorithm,a model fitting algorithm with prior knowledge of model is proposed,which combined with the physical characteristics of geometric models.In the algorithm,the dimension feature of point cloud is introduced as a prior probability,aim to solve the fitting problem about the vertical/non-vertical planar and cylindrical structure model in indoor scene.And the experiment is taken to verify the improvement of the proposed algorithm in accuracy,speed,and stability.(2)A iterative Gaussian mapping based segmentation framework of point cloud is proposed,which integrates a collection of existing method.Aim at the segmentation problem of indoor scene which with high complexity point cloud,this paper adopt an iterative segmentation strategy about point cloud to detect richer details in the scene.Firstly,the Gaussian mapping method is taken to cluster the point cloud,which based on the normal vector information of point cloud.Secondly,the proposed model fitting algorithm is taken to extract the points accurately.Finally,the structure of the indoor scene is extracted layer by layer by iterating over the steps above.The whole workflow of point cloud segmentation is described in detail in this paper,and the cluster method DBSCSN,which based on the density continuity of point cloud,is introduced.The fitting method about plane and cylinder model is given.And the model optimization methods are taken to fix out the over-segmentation problem in point cloud extraction,to ensure the quality of the segment product.Multiple experiments are taken to verify the feasibility of the segmentation strategy.And through method comparison and precision analyzing,to show the superiority of the method proposed in this paper.