A new approach to object recognition
Nikolai Sergeev · OPen Access Repositorium der Universität Ulm (OPARU) (Ulm University) · 2012
In this thesis a new approach to object recognition will be introduced. The basic idea of the approach is to compare two combinations of half ellipses representing two objects in order to find out whether the combinations can be at least partially transformed into each other by an affine mapping. As the transformation can be partial the system is robust to partial occlusion. The transformation does not have to be exact. Two combinations being able to be transformed into each other approximately with an error up to some epsilon-bound are regarded as similar. The epsilon-error tolerance makes the system robust to deformation. Further attractive properties of the approach are: its capability of stable separation of a single object from its background or of several objects partially occluding each other from one another, its capability to learn a new object in a time independent of the number of objects already learned, color information can be ignored or combined with form representation. Additionally a new machine learning algorithm developed to handle the representations based on half ellipses will be introduced. Because of the abstract task that it solves, the algorithm can be characterized as "near neighbor search". It looks for all stored vectors which are similar enough to the query vector with respect to the maximum norm. It is significantly responsible for many properties of the approach.