Date of Award

2012

Degree Type

Dissertation

Degree Name

Ph.D.

Department

Computer Science

First Advisor

Boufama, Boubakeur (Computer Science)

Keywords

Computer science.

Rights

CC BY-NC-ND 4.0

Abstract

The three-dimensional (3D) metric reconstruction of a scene from two-dimensional images is a fundamental problem in Computer Vision. The major bottleneck in the process of retrieving such structure lies in the task of recovering the camera parameters. These parameters can be calculated either through a pattern-based calibration procedure, which requires an accurate knowledge of the scene, or using a more flexible approach, known as camera autocalibration, which exploits point correspondences across images. While pattern-based calibration requires the presence of a calibration object, autocalibration constraints are often cast into nonlinear optimization problems which are often sensitive to both image noise and initialization. In addition, autocalibration fails for some particular motions of the camera. To overcome these problems, we propose to combine scene and autocalibration constraints and address in this thesis (a) the problem of extracting geometric information of the scene from uncalibrated images, (b) the problem of obtaining a robust estimate of the affine calibration of the camera, and (c) the problem of upgrading and refining the affine calibration into a metric one. In particular, we propose a method for identifying the major planar structures in a scene from images and another method to recognize parallel pairs of planes whenever these are available. The identified parallel planes are then used to obtain a robust estimate of both the affine and metric 3D structure of the scene without resorting to the traditional error prone calculation of vanishing points. We also propose a refinement method which, unlike existing ones, is capable of simultaneously incorporating plane parallelism and perpendicularity constraints in the autocalibration process. Our experiments demonstrate that the proposed methods are robust to image noise and provide satisfactory results.

Share

COinS