Exploitation of the vector field representation for pose estimation, view integration and 3D modeling using natural features

Authors: Nguyen, Van Tung
Advisor: Laurendeau, Denis
Abstract: View integration and registration is an essential and unavoidable phase in a 3D modeling process. The most challenging step of this problem is to estimate relative poses between views without using any initial information of the scanning position or using intervention prior to the acquisition step. We propose an automatic global registration method for point cloud data. The main problem addressed in this thesis is to solve the view integration and registration problem in 3D modeling for point cloud data. The proposed global approach is based on the Vector Field framework and automatically performs coarse to fine registration without requiring any assumption on the initial position between views or manual pre-processing for initial positioning. In particular, we exploit an augmented Vector Field representation to implement segmentation and extraction of features on the surface of an object in order to detect correspondences. In addition, the pose refinement process in the Vector Field reduces the complexity of the search for closest point correspondence since the information is implicitly encoded in the Vector Field representation. Also by exploiting the Vector Field representation, we provide a new method of registration that supports all steps of 3D modeling without requiring transformation of the data representation. An alternative solution using a variation of RANSAC-DARCES based in the Vector Field enables the proposed method to deal with objects of various types of geometry. Finally, the proposed approach is validated on multiple data sets such as standard models as well as real models scanned by hand-held scanners. The performance of the proposed method is evaluated by visual inspection as well as quantitatively by measuring the correspondence error.
Document Type: Thèse de doctorat
Issue Date: 2015
Open Access Date: 23 April 2018
Permalink: http://hdl.handle.net/20.500.11794/26498
Grantor: Université Laval
Collection:Thèses et mémoires

Files in this item:
SizeFormat 
32006.pdf20.51 MBAdobe PDFView/Open
All documents in CorpusUL are protected by Copyright Act of Canada.