Abstract:
This thesis investigates robust and fast methods for single and cooperative 2D/3D
image mosaicing to enhance field of view of images by joining them together.
Image mosaicing is underlined by the process of image registration and a significant
portion of the contributions of this work are dedicated to it.
Image features are identified as a solution to the problem of image registration that
uses feature-to-feature matching between images to solve for inter-image transformations.
We have developed a novel two signature distribution based feature descriptor
that combines grey level gradients and a colour histogram. This descriptor
is robust to illumination changes and shows better matching accuracy compared
to state of the art. Furthermore, we introduce a feature clustering technique that
uses colour codes assigned to each feature to group them together. This allows
fast and accurate feature matching as the search space is reduced.
Taking into account feature location uncertainty we have introduced a novel information
fusion technique to reduce this error by covariance intersection. This
reduced error location is consequently fed to an H∞ filter taking into account
system uncertainty for parameter estimation. We show that this technique outperforms
costly nonlinear optimisation techniques. We have also developed a novel
coupled filtering scheme based on H∞ filtering that estimates inter-image geometric
and photometric transformations simultaneously. This is shown to perform better
than standard least square techniques. Furthermore, we have introduced time
varying parameter estimation using recursive techniques that facilitate in tracking
changing parameters of inter-image transformations, suitable for image mosaicing
between moving platforms.
A method for rapid 3D scene reconstruction is developed that uses homographic
lines between images for semi-dense pixel matching. Triangular meshes are then
used for a complete visualisation of the scene and to fill in the gaps. To tackle
cooperative mosaicing scenarios, additional methods are presented that include
descriptor compression using principal components and 3D scene merging using
the trifocal tensor.
Capabilities of the proposed techniques are illustrated with real world images.
