Abstract (eng)
Motion estimation is an omnipresent goal in image analysis and computer vision.
An important task within is optical flow computation in a sequence of images.
It addresses the issue of inferring a vector field from intensity variations, thereby describing the displacements of moving objects.
Typically, optical flow is computed in the plane but is readily generalised to non-Euclidean settings allowing, for instance, for cell motion analysis in time-lapse microscopy data.
Today, fluorescence microscopy enables high-resolution observations of biological model organisms, such as the zebrafish, on the scale of single cells.
Despite its importance for tissue and organ formation, only little is known about cell migration and proliferation patterns during the zebrafish's early embryonic development.
Many of the questions raised involve estimating cell motion.
In view of increasing spatial as well as temporal resolutions resulting in tremendous amounts of data, manual analysis through visual inspection by humans is impracticable.
Therefore, automated cell motion estimation is key to the large-scale analysis of above-mentioned data.
Optical flow delivers quantitative methods and leads to insights into underlying cellular mechanisms and the dynamic behaviour of cells.
The primary biological motivation for this thesis is the desire to analyse cell motion in a living zebrafish embryo during early embryogenesis.
The data at hand depict endodermal cells expressing a green fluorescence protein.
Laser-scanning microscopy allows recording (volumetric time-lapse) 4D images of these labelled cells without capturing the background.
During early development these cells float on a so-called monolayer, meaning that they form a round surface in a single layer.
We exploit this situation and model this layer as a two-dimensional surface deforming over time.
The main idea of this thesis is to conceive cell motion only on this evolving surface.
As a direct consequence, one is able to reduce the spatial dimension of the data, resulting in more efficient motion estimation of afore-mentioned microscopy data.
We formulate the problem of cell motion analysis as a variational optical flow problem on evolving two-dimensional manifolds.
Naturally, this surface is subject to geometrical approximations.
In the first part, we focus on the embryo's changing geometry and assume that the cells' layer deforms over time.
To this end, we translate the original (Tikhonov-regularised) Horn-Schunck functional and the spatio-temporal extension by Weickert and Schnörr to this non-Euclidean and dynamic setting.
In the second part of this thesis, we pay close attention the topology of the embryo's surface.
First, we assume that it is a static round sphere and investigate several vector field decomposition functionals.
In particular, we follow recent trends in image decomposition and study u+v and hierarchical decomposition models for optical flow.
The chosen numerical method solves the problem in a finite-dimensional space spanned by tangential vector spherical harmonics and is advantageous in two ways.
It provides great flexibility with respect to the regularisation functionals and, as a by-product, yields a Helmholtz decomposition of the flow field.
Second, we consider a more appropriate geometrical model for the zebrafish's embryo, namely evolving sphere-like surfaces.
These surfaces can be parametrised from the 2-sphere and constitute a more natural approximation of the embryo's shape.
We extend the optical flow functional of Lefèvre and Baillet for surfaces embedded in R^3 to this new setting.
The variational problem is solved by means of a Galerkin method based on tangential vector spherical harmonics.
In order to find the sphere-like surfaces from cell microscopy data we devise a method for surface interpolation by means of scalar spherical harmonics expansion.
Finally, we present numerical results on the basis of the afore-mentioned cell microscopy data of a live zebrafish and picture the results in a visually adequate manner.