Important areas of research of the chair for pattern recognition in the project VAMPIRE are 3-D reconstruction and image-based object models. There is a close collaboration of the VAMPIRE project with project C2 of SFB 603 in the area of image-based object models. In order to make the generation of object models more robust and efficient, we implemented a structure-from-motion approach that is able to process image sequences with more than 200 frames in a few seconds. The input data for this approach is generated with our feature point tracker, which is based on the Kanade-Lucas-Tomasi tracker. Recent optimizations allow the tracking of 250 feature points at a frame rate of 30 fps. Custom-built optimizations for a high-speed camera developed by TU Graz improved the frame rate from 30 fps to 200 fps, when ten features are tracked.

The reconstruction of 3-D points with structure-from-motion approaches makes it impossible to determine the coordinates of points and cameras uniquely, because a global scaling of the scene does not change the input data. When two reconstructions of one scene have to be compared with each other, their relative scale factor has to be determined. To this end, we enhanced the well-known ICP (iterative closest point) algorithm by adding the possibility to also estimate this scale factor.

Another area of research of the chair for pattern recognition in the project VAMPIRE are object tracking and object recognition. Object tracking algorithms are either data-driven or model-based. Unlike model based approaches, data-driven algorithms do not require a priori knowledge about the object to be tracked. This is advantageous for starting the tracking of an unknown object directly after the detection of a movement in the scene. A disadvantage of data-driven approaches is that they lose the object in case of strong external rotations, because there is no knowledge about the appearance of the object from a different point of view. This disadvantage can be compensated by approaches that are very robust against appearance changes, or by model-based approaches, which gather knowledge about the object during a training stage. After an object was recognized by an object recognition algorithm, a model-based approach can be used to replace the data-driven approach for further tracking.

Existing approaches were enhanced and new approaches were developed for both data-driven and model-based object tracking. The basis for new approaches for data-driven object tracking was formed by an approach that uses color histogram features and the CONDENSATION algorithm. The extension of this approach to multiple cameras allows the robust estimation of the 3-D position of the tracked object with a data-driven approach. Thanks to the probability-based estimation of the 3-D position, the approach can cope with situations in which the object is temporarily hidden from one camera. Another extension combines the approach with a feature point tracker, which makes the resulting approach more robust against tracking an object in front of a similarily colored background. The combination of SIFT features for initialization with the feature point tracker for efficient tracking of the SIFT features is an example for a newly developed model-based object tracking algorithm. This algorithm allows the complete estimation of the 3D pose of the tracked object in real-time.