Large Scene Point Cloud Analysis and Modeling Using Geometric and Photometric Properties
Abstract：Obtaining 3D point cloud data from the real environment, and with filtering, matching, synthesis, rendering, etc., a real scene with a virtual model of the similar objects can be generated and this has a great importance for the production and daily life. The project aims to large scene three-dimensional point cloud and seeking for fast handling and reconstruction algorithms based on geometric and photometric properties. By using mutual information metric, point clouds are registered. Then by sampling and analyzing features, unorganized point clouds are segmented. The local signatures are extracted based on the color, light intensity and other properties, and semi-supervised learning methods are used for scene point cloud classification. For point cloud object formed by the classification, descriptors are extracted for the matching of model and point clouds. Meanwhile, the trees and buildings point cloud objects are reconstructed and visualized using primitive geometries and are eventually converted into large scene models that can be fast rendered. The key scientific problems to be solved is to use the project point cloud geometry and photometric properties with semi-supervised learning for precise point cloud classification, to match point cloud with models by feature descriptor, and to effectively combine primitive geometries which are used for point cloud fitting. The innovation of the project lies in the geometry-based and photometric-based multi-scale classification for point cloud, matching of point cloud and model based on descriptors, as well as the point cloud automatically processing.
Keywords: mutual information; shape descriptor; multi-scale; matching; reconstruction
Contact:
MENG Weiliang
E-mail: weiliang.meng@ia.ac.cn