Interpretation of the 3D Visual Environment from Uncalibrated Image Sequences
Ph.D. Thesis Bogdan Georgescu
Abstract
Metric reconstruction of a scene viewed by an uncalibrated camera
undergoing an unknown motion, is one of the most fundamental problems
in computer vision. Recent years have seen significant progress
in reliable analysis of image sequences and the recovered 3D scene
information can be used to generate new viewpoints, acquire 3D models,
track, insert and delete objects, or determine the ego-motion for visual navigation.
Inferring information about a scene starting from an image sequence
is a difficult task and the usual approach is to divide the
problem into several manageable subproblems. The processing
stages are composed from lower-level
tasks such as extracting salient image features to higher-level tasks such as determining
camera positions relative to the viewed scene.
The reliability of the processing chain depends on the robustness
of each module and the ability to cope with incorrect or noisy measurements.
In this thesis we have redefined several of the processing modules and developed
a highly accurate and robust system for the recovery of the 3D visual environment.
The process of image filtering is reformulated in a linear vector
space and the role of different subspaces is analyzed in the context of edge detection.
An edge confidence measure is introduced which allows higher sensitivity to sharp but
weak edges. Based on the distribution of image edge points in the line parametric space,
a method for lens distortion correction is presented.
For the detection of interest point correspondences we have combined the traditional
optical flow with matching color distributions.
Oriented kernels are introduced in the spatial domain to compute the
color distributions, thus obtaining rotation sensitivity. A joint robust minimization
procedure is employed and subpixel accuracy is
achieved under large image transformations.
Estimation of the structural and camera parameters relies on bundle adjustment, a nonliniar
optimization technique which minimizes the
reprojection error. The initial solution is usually obtained by solving a linearized
constraint at each stage of the reconstruction
process. The traditional way to obtain the initial solution is to apply a total least
squares (TLS) procedure which yields a biased estimate because it fails to correctly
account for the noise process
that affects the linearized measurements.
We present a more balanced approach where the initial solution is obtained from a
statistically justified estimator which assures its unbiasedness.
The quality of this initial solution, obtained using the heteroscedastic errors-in-variables
(HEIV) estimator, is already comparable with that of the bundle adjustment output, and thus
the burden on the latter is drastically reduced while its
reliability is significantly increased.
Each module is tested on synthetic data and standard images and the
performance of the 3D reconstruction system is illustrated on several
uncalibrated image sequences.
The thesis contains 166 pages.
Return to Theses