Chester F Carlson Center for Imaging Science MS. Thesis Defense MING ZHANG Towards 3D Matching of Point Clouds Derived from Oblique and Nadir Airborne Imagery Thursday, 1 August 2013, 10:00AM Carlson Bldg. 3215 (DIRS Lab)

Chester F Carlson Center for Imaging Science

MS. Thesis Defense

Ming Zhang

Towards 3D Matching of Point Clouds Derived from Oblique and Nadir Airborne Imagery

Advisors: Dr. John kerekes

Thursday, 1 August 2013, 10:00AM

Carlson Bldg. 3215(DIRS Lab)

Abstract

The use of oblique airborne images to reconstruct 3D scenes can overcome the limitation of traditional nadir-viewing images by providing more details of the side facets of buildings and rich texture information. However, because of the large projective distortion, it is difficult to match images taken from different viewing directions. An alternative method is to build a 3D point cloud for each group of images taken from the same direction and combine them to build a more complete cloud using 3D registration. In this thesis, we introduce an automatic way to generate and integrate point clouds extracted from muti-view imageryand test it.

The generation of an initial point cloud is based on a modified version of the RIT 3D Extraction Workflow, using the Affine Scale-Invariant Feature Transform (A-SIFT) instead of SIFT. To refine the point cloud generated, a Statistical Outliner Removal (SOR) method is used to eliminate the sparse noise spreading throughout the space, a Radius Outliner Removal (ROR) method is used to further remove the remained miscalculated floating clusters, and a Moving Least Square (MLS) method is used to smooth the surface.

Then we perform 3D registration to combine different point clouds which come with independent coordinate systems. Fast Point Feature Histograms (FPFH) are used to describe the features of the 3D SIFT keypoints. A multi-scale feature persistence analysis process is used to find the points that have unique features. The preliminary registration uses the SAmple Consensus (SAC) method to find point pairs with similar features from those points and achieve a rough alignment. The final registration uses the Iterative Closest Points (ICP) algorithm to obtain a more exact result. The output transformation matrix contains the information of scale, rotation and translation changes.