Methods for object detection and recognition have been studied for several years. Here the focus of activities in particular lies on perspective based object recognition as well as on image and feature based registration methods.
A statistical approach to view based object modeling and recognition was presented in [Hornegger1994ORU]. In [Reinhold2001IA3 , Reinhold2001ASO] this work has been continued, whereas wavelets instead of point features were used. With these approaches, it is possible to identify and locate objects even before heterogeneous backgrounds. The object detection problem has thus been reduced to a statistical estimation problem. This work has also been continued in [Deinzer2005OAI]. With the help of a statistical eigenspace method it has been shown exemplarily, how a controllable camera can be used for a directed exploration, respectively for the view planning for object recognition. Furthermore basic research on current image-tracking procedures using structural geometry models has been pursuited. In parallel several approaches from the fields of Computergraphics and Image Processing have been investigated in these image-tracking procedures. In the work of [Ewering2006MTM] a prototype for model-based tracking with point and curve based features has been realized. For the point based approach Harris Corners have been detected and then described by an eigendescriptor. Eigendescriptors have the advantage that they can be compared effectively during the matching task. In the line based approach the wire-frame model of the tracking object has been back projected into the picture with the smallest possible error. Therefor the ideas from [Wuest2005ALT], where search-beams with a direction perpendicular to the edge-direction supply appropriate candidates for the line matching, were basically used. The approach had been expanded by additional line distance dimensions. These were able to contribute to a faster convergence of the pose determination. By combining these procedures the tracking became more stable, even when using very simple distance dimensions.
Today, model-based approaches normally use statistics, appearance, shape or
geometry, which mostly work without an explicit geometrical model schema [Cremers2006AMD, Seemann2006MDO].
We only use models where the modeling of the objects is in a explicit
form, like it is used in Computer Aided Design, which rather corresponds to the
human view [Wirtz2010MRO].
Using models for explicit knowledge representation, there are two main strate-
gies how a controlling algorithm can handle the analysis process (cf. [Niemann1990PAA] p. 240ff). On the one hand, there is the data-driven strategy, where segmentation objects, found in the image, serve as an initialization for the analysis. Based on the segmentation objects, the best possible model is sought. The model-driven strategy on the other hand works the other way round. Each model determines whether it is contained in the image or not and tries to locate its elements.
Hybrid forms are feasible and favored in most cases. We choose a hybrid approach where data-driven models become pre-evaluated and then we selectively search model-driven for segmentation objects [Wirtz2010MRO]. These strategies can be called task-independent because they do not refer to knowledge of the domain. Other systems exist which use ontologies such as OWL and combine them with uncertain knowledge for finding concepts in a domain [Hois2007TCO, Reineking2009ECO].
» see Color Image Processing