Automatic Learning and Detection Of Point-Based Models

The aim of this diploma thesis was the extension of an existing optical infrared tracking (positioning) system. Using such an infrared optical tracking system with two cameras and retro reflective markers (beacons), it is common to define geometric models for the objects that have to be tracked. Normally, for the interaction with Virtual Reality environments, models consisting of only few markers are used. While one marker is enough to determine a position, at least three markers are needed for determining rotations and distinguishing between different models (based on their geometric constellation). For Augmented Reality applications often models with a much larger number of markers (more than five) are needed. To be able to track an arbitrary real object (e.g. in an industrial environment), it has to be equipped with several markers, such that arbitrary movements within the workspace are possible. To overcome the problem of occlusions of markers, models must be identified even if there are only few of its markers (minimum three) recognized by the tracking system. As a result of the increasing number of markers, models became more complex and their description in the system more difficult to perform. Therefore, an automatic learning process to store the model description was designed and implemented as well. This process allows the learning of models, whereas not all markers have to be seen by the cameras at a particular time. This procedure allows fast application of new models with a complex structure in contrast to systems, where the model to learn is not aloud to be moved during the learning phase.