This thesis describes a general framework for parameter estimation, which is suitable for computer vision applications. The approach described combines 3D modelling, animation and estimation tools to determine parameters of objects in a scene from 2D grey-level images. The animation tool predicts images using a 3D model of the scene (virtual reality), describing components like cameras, light sources and objects, and their parameters. The 3D modelling, using primitives like quadrics enables the handling of occlusion. A least squares estimator in combination with the modelling tool is used to estimate the selected parameters from real or animated grey-level images. The non-linear relation between the measurements and the set of parameters is coped with by the iterative application of the linear estimator. The least squares estimation paradigm is applied in a standardised way to the grey-level images of objects by considering the pixels as measurements of the object parameters. Two or more images, even when taken from different points of view, can be included simply by extending the measurement vector. Also the inclusion of a Gaussian filter, to increase the estimator performance by improving the image properties, can be carried out in a natural way.
|Award date||14 Jan 2000|
|Place of Publication||Enschede|
|Publication status||Published - 14 Jan 2000|