Incremental learning and novelty detection of gestures using extreme value theory

The problems of data streaming, e.g. "infinite length" and "concept-drift", require incremental self-adapting classifiers. The performance of the classifier, however, is affected by false labels. Consequently, the classifier is required to detect outliers or samples belonging to unseen classes, i.e. novelties. We propose an incremental Mahalanobis distance based classifier using extreme value theory to detect novelties. Extreme value theory allows for the determination of a global constant threshold that does not change during the adaption of the classifier and thus does not need additional validation data and/or procedures. The results show high accuracy and high efficiency in linear and non-linear spaces with respect to recognition results and computation time.