Online learning of appearance for robust data association in person tracking

Robust tracking of persons in video sequences is an important task in many applications, like for example video surveillance, human-machine-interaction or adaptive public advertisement. Modern tracking schemes usually first apply person detection in subsequent image frames, perform an association of detections to tracked person and then update the state of the individual tracked persons. Therefore, the overall performance of the tracking strongly relies on correct assignments of observations (detections) to tracked persons. In order to improve identification among different images, visual appearance can be used to distinguish individuals. As individuals and their appearance are usually not known in advance, appearance has to be learned online, while the person is being tracked. In this thesis, an approach to online learning of person appearance will be implemented and evaluated. Online learning will be based on Random Forests®, a tree-based classifier that has proven to be suitable for real-time learning. In order to improve identification, several image features will be evaluated (e.g. RGB, HOG), based on published datasets, which are commonly used in the literature. For colour-based features, background-subtraction will be evaluated as a means of avoiding the common problem that backgroundrelated appearance information is often mistaken by the classifier for being related to a person.