Segmentation For Indoor SLAM Point Cloud Assisted By Geometric Structure Information

3D point cloud has the characteristics of fast acquisition speed, real-time, high precision and all-digital features,which is one of the most important data sources of 3D modeling. Due to some problems of point cloud data, such as the great amount of data, the poor dispersion, the low edge accuracy of point cloud and noises existed in data, data processing requires manual assistance in the existing point cloud segmentation methods for point cloud segmentation. And the result of segmentation is also influenced by human factors.There are three problems in most of traditional point cloud segmentation algorithms. Firstly, the algorithms cannot meet the needs in large-scale point cloud segmentation. Secondly, the degree of automation of point cloud segmentation in the algorithms is low, most are semi-automatic segmentation under artificial auxiliary.Thirdly, the algorithms can only realize the segmentation and extraction of regular objects, and be weak to the segmentation of curved objects. Based on the characteristics of experimental data,this paper utilized the method of roughly segment and finely segment to point cloud data, realized the fine segmentation of indoor SLAM(Simultaneous Localization And Mapping) point cloud.The main contents and results of this paper are listed as follows:1) Preprocessing of indoor SLAM point cloud data. Indoor SLAM point cloud data has the characteristics of large amount of data, poor dispersion, low edge accuracy of point cloud and unavoidable noises. Therefore, data preprocessing is required before segmenting point cloud data. First of all, this paper construct spatial index of point cloud data based on KD-tree to improve the efficiency of neighborhood searching for point cloud. Then, point cloud data has been done to reduce the amount of data. At last, removing the discrete points has been conducted by a method of remove sparse discrete points to reduce point cloud noises.2) Clustering Segmentation of 3D Point Cloud Data. In order to improve the efficiency and the degree of automation of clustering algorithm,this paper propose an efficient representative points selection method for FDBSCAN algorithm. A fast and automatical segmentation methods has been proposed under the combination between the automatic ?-radius estimation algorithm and cluster merging algorithm for the indoor SLAM point cloud.3) Fine segmentation of cloud data. In this paper, first using he color-space region growing segmentation algorithm in the process of coarse segmentation in the same cluster of different object segmentation; Then get the geometric structure information of the indoor SLAM point cloud by the panoramic image semantic segmentation, and mark the corresponding point cloud data. The geometric structure information is matched to the point cloud data to be segmented by point cloud matching; According to the geometric structure of the information of label set RANSAC algorithm model, so as to realize segmented fine segmentation of point cloud data.4) Experimental verification and analysis. The measured point cloud data has been adopted as the evaluation for the proposed methods. The experimental results show that the proposed method can achieve fine segmentation for indoor SLAM point cloud, and improve the efficiency and the automation degree for the point cloud segmentation.