Incremental Learning and Novelty Detection of Gestures in a Multi-Class System

The difficulties of data streams, i.e. infinite length, the occurrence of concept-drift and the possible emergence of novel classes, are topics of high relevance in the field of recognition systems. To overcome all of these problems, the system should be updated continuously with new data while the amount of processing time should be kept small. We propose an incremental Parzen window kernel density estimator ( IncPKDE) which addresses the problems of data streaming using a model that is insensitive to the training set size and has the ability to detect novelties within multi-class recognition systems. The results show that the IncPKDE approach has superior properties especially regarding processing time and that it is robust to wrongly labelled samples if used in a semi-supervised learning scenario.