Use this URL to cite or link to this record in EThOS: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.524972
Title: Robust surface modelling of visual hull from multiple silhouettes
Author: Shin, Dongjoe
Awarding Body: University of Warwick
Current Institution: University of Warwick
Date of Award: 2008
Availability of Full Text:
Access from EThOS:
Access from Institution:
Abstract:
Reconstructing depth information from images is one of the actively researched themes in computer vision and its application involves most vision research areas from object recognition to realistic visualisation. Amongst other useful vision-based reconstruction techniques, this thesis extensively investigates the visual hull (VH) concept for volume approximation and its robust surface modelling when various views of an object are available. Assuming that multiple images are captured from a circular motion, projection matrices are generally parameterised in terms of a rotation angle from a reference position in order to facilitate the multi-camera calibration. However, this assumption is often violated in practice, i.e., a pure rotation in a planar motion with accurate rotation angle is hardly realisable. To address this problem, at first, this thesis proposes a calibration method associated with the approximate circular motion. With these modified projection matrices, a resulting VH is represented by a hierarchical tree structure of voxels from which surfaces are extracted by the Marching cubes (MC) algorithm. However, the surfaces may have unexpected artefacts caused by a coarser volume reconstruction, the topological ambiguity of the MC algorithm, and imperfect image processing or calibration result. To avoid this sensitivity, this thesis proposes a robust surface construction algorithm which initially classifies local convex regions from imperfect MC vertices and then aggregates local surfaces constructed by the 3D convex hull algorithm. Furthermore, this thesis also explores the use of wide baseline images to refine a coarse VH using an affine invariant region descriptor. This improves the quality of VH when a small number of initial views is given. In conclusion, the proposed methods achieve a 3D model with enhanced accuracy. Also, robust surface modelling is retained when silhouette images are degraded by practical noise.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.524972  DOI: Not available
Keywords: TA Engineering (General). Civil engineering (General)
Share: