ActivityExplorer: A semi-supervised approach to discover unknown activity classes in HAR systems

Abstract

• A semi-supervised activity recognition approach able to identify unknown activity classes. • Approach designed to deal with small-sample set scenarios with limited amount of training data. • Approach based on affinity propagation clustering able to automatically identify the number of clusters. • Results on public datasets confirm the efficacy of the proposed approach. A semi-supervised activity recognition system is here proposed to deal with partially labeled video-sequences, where the uncertainty in the data comes from two different factors: only a subset of the data has a class label assigned and only part of the activity classes are known. In particular, the paper presents ActivityExplorer, an approach able to identify clusters of similar activity patterns within the dataset and to identify those clusters that might correspond to new activity classes, still unknown to the recognition system. These capabilities are realized thanks to a combination of metric learning, used to determine a suitable subspace for pattern classification, an advanced clustering technique and ad hoc indicators defined to estimate the membership of each pattern to known classes and possibly identify new activities.