IMPROVEMENT OF SURFACE RECONSTRUCTION PIPELINE USING OPEN SOURCE FOR UNSTRUCTED POINT CLOUD

摘要

As low-cost active sensors have become popular and techniques for extracting point cloud through multi images have been developed, productivity of point clouds has improved. Accordingly, there are growing demands for visualization of point clouds and generation of realistic 3D models. Since point clouds have massive data in general case, surface reconstruction from them requires long time and complex computation. There are two types of point clouds; structured and unstructured type. Point clouds obtained from passive sensors such as multi images are unstructured ones. It is relatively more difficult to generate 3D model from this point cloud type. In general, surface reconstruction process has a five-step pipeline; analysis of point cloud dataset, filtering, point normal calculation, reconstruction, and surface mesh generation. 3D models are generated by applying appropriate algorithm at each step. However implementing complicate algorithms at each step for massive data is not a trivial task. There are some groups who have created these techniques in the form of open source. They provide source codes and API level of methods for surface reconstruction pipeline. In this paper we form a general processing pipeline from open source to generate 3D models by using unstructured point cloud gained from UAV multi images. Then we analyze the problem of the general pipeline and propose an improved pipeline that is suitable for unstructured point cloud.