La mise en registre automatique des surfaces acquises à partir d'objets déformables

Authors: Cao, Van Toan
Advisor: Laurendeau, Denis
Abstract: Three-dimensional registration (sometimes referred to as alignment or matching) is the process of transforming many 3D data sets into the same coordinate system so as to align overlapping components of these data sets. Two data sets aligned together can be two partial scans from two different views of the same object. They can also be two complete models of an object generated at different times or even from two distinct objects. Depending on the generated data sets, the registration methods are classified into rigid registration or non-rigid registration. In the case of rigid registration, the data is usually acquired from rigid objects. The registration process can be accomplished by finding a single global rigid transformation (rotation, translation) to align the source data set with the target data set. However, in the non-rigid case, in which data is acquired from deformable objects, the registration process is more challenging since it is important to solve for both the global transformation and local deformations. In this thesis, three methods are proposed to solve the non-rigid registration problem between two data sets (presented in triangle meshes) acquired from deformable objects. The first method registers two partially overlapping surfaces. This method overcomes some limitations of previous methods to solve large global deformations between two surfaces. However, the method is restricted to small local deformations on the surface in order to validate the descriptor used. The second method is developed from the framework of the first method and is applied to data for which the deformation between the two surfaces consists of both large global deformation and small local deformations. The third method, which exploits both the first and second method, is proposed to solve more challenging data sets. Although the quality of alignment that is achieved is not as good as the second method, its computation time is accelerated approximately four times since the number of optimized parameters is reduced by half. The efficiency of the three methods is the result of the strategies in which correspondences are correctly determined and the deformation model is adequately exploited. These proposed methods are implemented and compared with other methods on various types of data to evaluate their robustness in handling the non-rigid registration problem. The proposed methods are also promising solutions that can be applied in applications such as non-rigid registration of multiple views, 3D dynamic reconstruction, 3D animation or 3D model retrieval.
Document Type: Thèse de doctorat
Issue Date: 2016
Open Access Date: 24 April 2018
Permalink: http://hdl.handle.net/20.500.11794/26764
Grantor: Université Laval
Collection:Thèses et mémoires

Files in this item:
SizeFormat 
32450.pdf24.45 MBAdobe PDFView/Open
All documents in CorpusUL are protected by Copyright Act of Canada.