Summary
As active computer vision systems are motivated by the visual systems of intelligent beings, they are designed to autonomously explore their environments, to deduce task dependent information and to develop intelligent behaviour in case of unexpected events. Starting from object recognition methods implemented in an existing active vision system, this project will continue the development of a biologically motivated object representation scheme. Its elements view-based recognition, temporal associations, and structural descriptions turned out to be incomplete as long as single views cannot be adequately represented and recognized. It remained an unsolved problem that the matching of feature maps - despite of complex features and flexible matching approaches - does not allow sufficiently reliable comparisons between input images and stored views.
Therefore, we intend to merge recent findings about shape perception and perceptual grouping methods to derive the foundations of a shape description framework. In contrast to known approaches, we will simultaneously perform the decomposition and grouping of shapes in order to bridge gaps between image features and to handle occlusions by a multi-layered representation scheme.
For navigating and manipulating objects, animals, as well as autonomous robots, need
information about the form and the position of objects. The advantage of camera systems is
that there is no need for a complex data fusion between object recognition tasks and depth
perception. Furthermore it is also possible to realize low-cost solutions. Nevertheless, when a
three dimensional scene is projected to the two dimensional receptive field of a camera or a
biological vision system, all depth information is lost. Also in the two dimensional display the
form of an object is reduced to a contour, that may be interrupted due to occluding objects.
In order to reconstruct the lost contour and depth information, animals, especially human
beings, are taken as a model. The most obvious approach to such a reconstruction is the use
of stereo vision. While there has been a great amount of research done on stereoscopic
approaches, which produced many efficient algorithms, little emphasis was placed on
monocular depth criteria so far. In this work, occlusion, being one of the most important
monocular depth criteria, is used for the three dimensional interpretation of a scene and for
the reconstruction of occluded contours. Emphasis is placed on the entire process chain,
starting with the filtering of natural images. With anisotropic diffusion, a method is introduced,
that enables a reduction of input images to salient features, while preserving edge
information.
Contact
Do you have any questions or comments? Please contact:
|
