| Titre : |
Multiphoto geometrically constrained matching |
| Type de document : |
Thèse/HDR |
| Auteurs : |
Emmanuel P. Baltsavias, Auteur |
| Editeur : |
Zurich : Institut für Geodäsie und Photogrammetrie IGP - ETH |
| Année de publication : |
1991 |
| Collection : |
IGP Mitteilungen, ISSN 0252-9335 num. 049  |
| Importance : |
221 p. |
| Format : |
21 x 30 cm |
| ISBN/ISSN/EAN : |
978-3-906513-01-0 |
| Note générale : |
Bibliographie
Doctoral Thesis, ETH Zurich, 1991 |
| Langues : |
Anglais (eng) |
| Descripteur : |
[Vedettes matières IGN] Photogrammétrie numérique
[Termes IGN] appariement d'images
[Termes IGN] appariement géométrique
[Termes IGN] points homologues
[Termes IGN] prise de vues
|
| Index. décimale : |
33.30 Photogrammétrie numérique |
| Résumé : |
(auteur) One of the central problems in digital photogrammetry, Computer and machine vision is the problem of automatically finding corresponding features in different images. Correspondence is necessary for 2-D and 3-D measurements and very often a prerequisite for object detection, Classification and identification. Although research for more than three decades has been devoted to this problem, a fully automated, precise and reliable image matching method, able to adapt to different image and scene Contents, does not exist yet. The aim of this research was to improve matching in the above aspects by developing and examining the Performance and applicability of a new algorithm called Multiphoto Geometrically Constrained Matching (MPGC). MPGC is an extension of Least Squares Matching (LSM), which is an area-based matching method establishing a fit between small image patches by an affine geometric and a
two parameter linear radiometric transformation. MPGC considerably improves LSM by using two new elements: (i) the exploitation of any a priori known geometric information to constrain the Solution and (ii) the simultaneous use of any number (more than two) of
images. These elements reduce the search space (conditional 1-D search), improve the accuracy and especially the success rate and reliability of matching. The mathematical model of the algorithm is formulated in terms of a combined least Squares adjustment. The
Observation equations consist of the equations formulating the grey level matching and those that express the geometric constraints, the two parts being related to each other through common unknown parameters. The geometric constraints that have been used in this
research are the collinearity conditions extended by additional parameters modelling systematic errors, assuming that the interior and exterior orientation of the sensors are known. Thus, MPGC permits a simultaneous determination of pixel and object coordinates.
Additionally, through the connection of image and object space, any number of images can be simultaneously accommodated. Two measurement modes were analysed: determination of X, Y, Z object coordinates (match points are fixed in one reference (template) image), and determination of Z over fixed X, Y object coordinates (match point position must be estimated in all images). For greater flexibility the geometric constraints are treated as weighted Observation equations and not as strict conditions. The radiometric parameters are not included in the mathematical model; instead to increase flexibility and speed, radiometric corrections are applied during the iterations. As an alternative, it is proposed to use a Wallis filter for a radiometric equalisation of the images before matching. The Wallis filter has also been used to enhance contrast, high contrast being necessary for accurate matching. Emphasis has been placed on three aspects of the algorithm. Since MPGC is based on a nonlinear least Squares estimation, it requires the knowledge of reasonably precise approximations. After a theoretical analysis of the required quality of the approximations, a method to derive the approximations by a image pyramid based (coarse-to-fine) approach is presented. Different kerneis for the pyramid generation are compared, problems occuring at image borders are treated, geometric relations between pyramid levels are derived and criteria
for the choice of appropriate matching parameters at each level are proposed. The image pyramid approach, in addition to being an efficient way to derive approximate values, increases the convergence radius (in practical tests parallaxes up to 70 pixels have been
handled), convergence rate and computational speed, and can be exploited for a better quality control and self-adaptivity of the algorithm. Without the image pyramid, convergence has been achieved, in some cases, for errors in the approximate values of up to 10-20 pixels, but optimally these approximate values should be 1 pixel accurate. A critical question for surface measurement is the choice of points which should be measured. These points must be characteristic points of the surface but also well determinable. The proposed strategy consists of choosing good match points in the reference image and matching them in all pyramid levels. Points which are considered good lie on edges vertical to the geometric constraint line. They are derived by a direction selective interest Operator using the first and/or the second intensity derivatives. The operator's parameters can be adapted so that the selection of "noisy" pixels is reduced and the required density of selected points is fulfüled. Methods to ensure the existence of the selected points in all pyramid levels are proposed. Thus, MPGC is a combination of area-based and feature-based, especially edge-based, matching. The third major aspect of the investigations is the quality evaluation of the results, the detection of gross errors and the automatic adaptation of the algorithm to different image contents and object surfaces. This aspect includes measures to express the achieved precision, methods for the automatic detection of Wunders and Observation errors, different tests to check the determinability and the significance of the shaping parameters, and ways of automatically adapting the image patch size to the signal content and object surface. The blunder detection test involves checking each individual object point ray based on a combination of criteria, whose thresholds are adapted to the processed image. Many practical accuracy studies, referring to a comparison of accuracy and precision and the amount of detected and undetected Wunders, are presented. The theoretical precision of the shifts, in the
case of good targets, typically is 0.01 - 0.05 pixels. The achieved accuracy was for good planar targets 0.2 - 1 um, for signalised or good natural points 2-3 (im, and for natural points on general surfaces 10 - 15 p.m. whereby the pixel spacing was typically 10 um. In the performed tests, the matching accuracy was generally similar to the accuracy of manual measurements; in certain cases the matching accuracy was even higher than the manual one. The percentage of blunders automatically detected by MPGC varied from 5% - 25% of the total number of points, depending on the image content and object surface. The percentage of undetected Wunders was 1% - 3% of the points accepted by MPGC as being correct, thus comparable to the error rate of a human Operator. Further, it is shown how MPGC can reduce problems like multiple Solutions, occlusions, discontinuities and radiometric distortions in comparison to other algorithms. Finally, methods for the reduction of oscillations and divergence problems and computational aspects are treated. Different applications and tests are presented. |
| Numéro de notice : |
60397 |
| Affiliation des auteurs : |
non IGN |
| Thématique : |
IMAGERIE |
| Nature : |
Thèse étrangère |
| Note de thèse : |
Doctoral Thesis : Science : Zurich : 1991 |
| DOI : |
10.3929/ethz-a-000617558 |
| En ligne : |
https://doi.org/10.3929/ethz-a-000617558 |
| Format de la ressource électronique : |
URL |
| Permalink : |
https://documentation.ensg.eu/index.php?lvl=notice_display&id=60744 |
|  |
Multiphoto geometrically constrained matching (1991)
