Indoor Point Cloud Segmentation Based On Multi-model Fitting

Building three-dimensional model plays an important role in many fields.In recent years,3D modeling of building technology develop rapidly,and bring out the bloom in the digital city and intelligent city construction.With the continuous construction of digital city construction,only the surface model of the buildings has been unable to meet the demand.The public area of the three-dimensional model of building ask for the model within the building.Nowadays,the indoor cloud scanning technology is becoming more and more mature,and the cost of acquisition of the indoor point cloud is decreasing.The indoor point cloud has became an important data source for the three-dimensional reconstruction of the building.Indoor point cloud data acquisition methods are diverse,in addition to high-quality laser scanning point cloud data,ultra-high kinetics and other low-cost low-quality relatively low point cloud data are also included.The segmentation of point cloud data is a key step in the 3D reconstruction of indoor scene,which is of great significance to the study of indoor point cloud.Considering the features of indoor point cloud data,the advantages and disadvantages of multi-model method,this paper mainly completed the following research contents:(1)Proposing a partial threshold adaptive multi-model fitting methodThe current multi-model fitting method usually requires generating a large number of Hypotheses,and then the output models are selected from those hypotheses,thus the computational efficiency can not be guaranteed.At the same time,the result of the algorithm depends on the accuracy of the parameter setting,so it often takes a long time to adjust the parameters.Taking those in to consideration,the method proposed in this paper divides the data from top to bottom,and improves the calculation efficiency while ensuring the accuracy of calculation.For some of the thresholds in the algorithm,the proposed algorithm can adaptively select the appropriate threshold,thus saving some of time used on parameters tuning.(2)Laser scanning indoor point cloud segmentation processIndoor point cloud includes a large number of geometric elements,by fitting the point cloud into cylindrical surface and plane,indoor point cloud can be segmented.Through laser scanning,we can obtain the indoor point cloud whose density and accuracy are high,while containing a lot of models.Multi-model fitting method based on RANSAC can only segment the point cloud when the numbers of models are told,and it is not applicable in complex scenes because of the need to set the number of models in advance.In this paper,a set of multi-model fitting method based on split and merge is proposed.Firstly,the data is pre-segmented,and then the segmentation result is obtained by multi-model fitting on the basis of pre-segmentation.Based on the data distribution characteristics and the interaction between the cylindrical model and the plane model,the region of the cylinder is determined by the pre-segmentation result and the model fitting result,and then the cylindrical model is fitted by the slice technique.Through a set of processes,the indoor point cloud can be fit into plane area and the cylindrical area,and segment the indoor point cloud.(3)RGB-D image cloud segmentation processCompared with the laser scanning point cloud data,RGB-D image is more cheap easy to get,but the depth information obtained by Kinect is often lower in quality,and has more noise and void,the depth of the effective measurement area is narrow,and laser scanning to obtain point cloud data can not be used with the same set of processes.In this paper,we propose a point cloud data segmentation method for the transformation of the depth information obtained by Kinect,and realize the segmentation of the low quality point cloud data by using pretreatment like filtering and down sampling.Multi-model fitting method is applied to different geometric elements(such as sphere,cone,plane,cylinder,etc.)through multi-model fitting,point cloud segmentation can be achieved.The multi-model fitting method can achieve better results for the indoor point cloud,which includes a large number of basic geometric primitives.Based on the experimental process proposed in this paper,the efficiency of point cloud segmentation can be improved obviously compared with other model fitting methods.