Abstract
The usual input to a motion analysis system is a temporal image sequence. Even though motion analysis is often called dynamic image analysis, it is frequently based on a small number of consecutive images, sometimes just two or three in a sequence. This case is similar to an analysis of static images, and the motion is actually analyzed at a higher level. There are three main groups of motion-related problems: Motion detection, moving object detection and localization, and derivation of 3D object properties. In this thesis we focused our attention on the second group. More specifically, within this group we will be dealing with four main issues: Moving object boundary detection, boundary delineation, boundary representation and description, and boundary matching for fast future location prediction. To detect moving object boundaries a new theory derived from temporal cooccurrence matrices is proposed, developed and applied. Afterwards, a filter design is developed to get fast and accurate results. As any boundary detection method, the output from this stage is usually a set of unlinked segments of boundaries. To assemble these segments of boundaries into meaningful boundary, a new active contour model has been proposed and developed that is capable of escaping energy minima caused by noise. Since our method for matching we based on the correspondence of interest points (feature points), we needed a proper set of invariant descriptors in order to match contours of two successive frames. For this task, a new theory on boundary representation and description called The theory of variances and varilets has been proposed and developed. We used moments of the variance transform and the normalized variance transform for an initial matching of contours which is in some sense a global matching. Afterwards, an Iterative sub-mappings strategy has been proposed and applied for fine matching. An important issue from the moment function was that extrema of successive derivatives provide as a coarse-to-fine matching, where to each feature point we assigned a proper descriptor induced from the normalized variance transform matrix.