Towards adaptive models for classification of technical objects

The recognition of the specific type of a technical object under consideration is often crucial in understanding its capabilities and its potential behaviour. Reliable automatic object type recognition from digital imagery is therefore demanded in many applications like reconnaissance. In this paper the problem of simultaneous estimation of object pose and type-characteristic shape parameters from images is attacked by the use of CAD models. In order to cope with the usually incomplete and even erroneous model data set (base models), we build from these a flexible adaptive model which forms an orthogonal basis. For this purpose a metamodel is constructed by volume-based metamorphosis from the base models, which are then inserted by mesh optimization methods to give the final flexible model. The model and pose parameter estimation is based on the correspondence of image and model edge elements to avoid illumination influence. An objective function for robust estimation is used and optimized with respect to the parameters by a Levenberg-Marquard method. Preliminary results with civil aircraft show the applicability of the methods presented.