Zur Erkennung verformbarer Objekte anhand ihrer Teile

The recognition of deformable objects by the means of digital image processing is a crucial, but widely unsolved problem yet. In many industrial and other areas there is a strong need to automate processes which take place in a changing or not completely controllable environment. However, technical systems are presently characterised by fixed operational procedures and little interaction with their environment. The main problem lies in the interpretation of the camera data. The existing object recognition methods work only in simple special cases.

Therefore, in this thesis a novel approach is studied which allows for a simultaneous classification and localisation of the objects present in an image. To this end, a compositional model is introduced which describes an object as a hierarchy of parts and sub-parts. Between parts, geometrical relationships are modelled. A major emphasis is placed on the analysis of the relationship between the position and the attributes of parts. To account for the strongly varying appearance of deformable objects, the model stores multiple views. This is in contrast to many other recent approaches.

The model is build by analysing sample images of the objects to be recognised. The training comprises both the automatic selection of appropriate parts as well as the identification of characteristic views. Due to differing boundary conditions, the resulting sets of parts are optimised individually for every level in the hierarchy. The comparison between the model and a test image is done by a voting method which has similarities to the Hough transform or to radial basis functions.

The performance of the new methods is demonstrated by the recognition of a character from a cartoon data-base with strongly varying appearance. Two model configurations are presented achieving either a precision of 97 percent or a recall of 93 percent with a general accuracy of at least 78 percent for both cases.