Search

In this thesis new methods for the automatic recognition of the content of image sequences are presented. Solutions to the following video sequences analysis tasks are developed: temporal decomposition of an image sequence into scenes and classification of the scenes, and the recognition of people and their movements in the image sequence. The temporal segmentation of a image sequence and the classification of the segments can be used for image sequences with a given content structure, like broadcast news. The image sequences have a defined chronology of scenes, which belong to certain content classes. The content classes and their chronology are represented by nested Hidden Markov models during the recognition. Another application of the Hidden Markov Modells are the classification of movements of objects in the image sequence. The recognition of human gestures for the application of human-computer-interaction is investigated. The recognition system is capable of recognizing a set of pre-defined gestures that are performed in the viewing area of a camera. The system is able to identify undefined movements and can distinguish them from the gestures. The final task is recognizing people visible in image sequences. The recognition of the people is done by recognizing their faces. The indexing of the faces is composed of the two sub-tasks: detection of the faces and recognition of the faces. It is shown that the face-based video indexing can be used to find known persons in the image sequence as well as to group the people in the sequence unsupervised.