Image point matching in multiple-view object reconstruction from imagesequences

Abstract

This thesis is concerned with three-dimensional (3D) reconstruction and point

registration, which are fundamental topics of numerous applications in the area of

computer vision.

First, we propose the multiple epipolar lines (MEL) shape recovery method for

3D reconstruction from an image sequence captured under circular motion. This

method involves recovering the 3D shape by reconstructing a set of 3D rim curves.

The position of each point on a 3D rim curve is estimated by using three or more

views. Two or more of these views are chosen close to each other to guarantee

good image point matching, while one or more views are chosen far from these

views to properly compensate for the error introduced in the triangulation scheme

by the short baseline of the close views. Image point matching among all views

is performed using a new method that suitably combines epipolar geometry and

cross-correlation.

Second, we develop the one line search (OLS) method for estimating the 3D

model of an object from a sequence of images. The recovered object comprises a

set of 3D rim curves. The OLS method determines the image point correspondences

of each 3D point through a single line search along the ray defined by the camera

center and each two-dimensional (2D) point where a photo-consistency index is

maximized. In accordance with the approach, the search area is independently reduced

to a line segment on the number of views. The key advantage of the proposed

method is that only one variable is focused on in defining the corresponding 3D

point, whereas the approaches for multiple-view stereo typically exploit multiple

epipolar lines and hence require multiple variables.

Third, we propose the expectation conditional maximization for point registration

(ECMPR) algorithm to solve the rigid point registration problem by fitting

the problem into the framework of maximum likelihood with missing data. The

unknown correspondences are handled via mixture models. We derive a maximization

criterion based on the expected complete-data log-likelihood. Then, the point

registration problem can be solved by an instance of the expectation conditional

maximization algorithm, that is, the ECMPR algorithm.

Experiments with synthetic and real data are presented in each section. The

proposed approaches provide satisfactory and promising results.