Structure from Motion in nD Image Analysis
Year: 2004
Type: dissertation
Abstract: In this thesis we investigate the measurement of local properties in multi-dimensional grey-value images. Special attention is given to orientation representation and curvature estimation. Furthermore, the possibility to compute global shape properties from these local properties. We aim to derive sampling-error free measurements. A sampling-error free measurement of a properly sampled image yields exactly the same result as the sampled analog measurement (chapter 3). We propose a new robust curvature estimation algorithm in nD grey-level images for lines (chapter 4) and surfaces (chapter 5). A better understanding and a generalization of the continuous orientation representation introduced by Knutsson (chapter 2) enables us to compute derivatives of the orientation field leading to the new curvature estimator. In our approach images are represented as multi-dimensional functions and we assume that locally only a single orientation is present. No parametric description of the objects is needed. The curvature can reliably be computed on ridges and in valleys, where classical isophote curvature fails. Isophote curvature cannot deal with areas of vanishing gradient as the curvature is normalized by the gradient magnitude. The local property curvature is used to estimate global properties as bending energy and Euler characteristic (chapter 4 and 5). The integration over a digital object surface is accomplished by a summation of the gradient magnitude over the whole image after soft-clipping. These local and global measurements are applied to time series:(x,t)-images. Here orientation corresponds to velocity and curvature to acceleration. We will discuss several motion estimation algorithms (chapter 6) and show the connection between spatial and temporal sampling(chapter 3). The developed algorithms are employed in three applications: 1) the characterization of the rising behavior of bread dough, 2) the flow field analysis of the motion of microspheres in nanoliter vials during evaporation of the liquid, where the liquid flow and the Brownian motion are discussed, 3) a biological application where fluorescent labeled chromatin is observed in a living cell with a confocal microscope. We propose a robust method to correct for the superimposed motion/drift of the cell on the chromatin during the acquisition, which enables a better study of the dynamics of the chromatin.
Author: Bernd Rieger
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W1587509359
