THESIS
2020
xiii, 109 pages : illustrations (some color) ; 30 cm
Abstract
Shape, dimension and curvature are some important specifications that greatly affect
appearance and functionality of specular surfaces. Given a high production rate, fast
and accurate quality inspection is a crucial procedure in manufacturing to ensure compliance
with engineering specifications. Traditional inspection techniques for specular surface
include laser profile sensor and chromatic confocal sensor. While high measurement
accuracy can be accomplished with these sensors, the point-based measurement compromises
either speed or resolution. This motivates the development of three-dimensional
(3D) measurement by stereovision because of its promising accuracy, speed and resolution.
However, difficulty in correspondence matching is encountered in reconstruction of
specular su...[
Read more ]
Shape, dimension and curvature are some important specifications that greatly affect
appearance and functionality of specular surfaces. Given a high production rate, fast
and accurate quality inspection is a crucial procedure in manufacturing to ensure compliance
with engineering specifications. Traditional inspection techniques for specular surface
include laser profile sensor and chromatic confocal sensor. While high measurement
accuracy can be accomplished with these sensors, the point-based measurement compromises
either speed or resolution. This motivates the development of three-dimensional
(3D) measurement by stereovision because of its promising accuracy, speed and resolution.
However, difficulty in correspondence matching is encountered in reconstruction of
specular surface. Methods designated for specular surface reconstruction also experience
drawbacks such as dependence on prior knowledge of the surface or low accuracy.
The objective of this thesis is to develop a 3D reconstruction method for specular surface
by stereovision that provides accurate measurement and works without prior knowledge
of the surface. Using a binocular-vision setup, an improved reconstruction method is
proposed to solve the correspondence matching problem. The method uses surface normal
as common feature for correspondence matching and quickly searches for the best match
by gradient descent. Meanwhile, analysis shows that the binocular reconstruction result
is not necessarily unique. An error-based selection algorithm is proposed to select the correct result out of the non-unique result. The algorithm fits a local analytical surface
from combination of different reconstruction results over a small local region. The combination
that shows minimal error between the analytical surface normal and measured
normal is selected as the correct result.
Further investigating the binocular system reveals that the system suffers from a bilinear
interpolation error which significantly compromises reconstruction accuracy. To
improve the reconstruction accuracy, a combined reconstruction method is further proposed.
The method uses reconstruction result from the improved binocular method as
reference surface for accurate slope estimation. Integration-based reconstruction method
is then applied to integrate the measured slope into 3D profile of the surface.
Simulations and experiments are conducted to verify feasibility of the proposed algorithms.
Improvement in reconstruction accuracy over binocular-vision method is experimentally
shown by measurement of a stair-like object. While the size of error scales with
measurement range, relative error of 0.59% or less is recorded in measurement of different
step sizes. The overall reconstruction method can provide accurate measurement and
work without prior knowledge of the specular surface.
Post a Comment