Unsupervised Activity Discovery and Characterization From Event-Streams

Introduction: Recognizing what is happening in an environment has many potential applications, ranging from automatic surveillance systems to supporting users in ubiquitous environments. A key step to this end is to discover the kinds of similar activities that frequently occur in a particular domain. Equally important is the question of finding efficient characterizations for these different kinds of activities. We are interested in the study of activity class discovery and characterization, in the context of analyzing everyday activities. We present a novel representation of activities as bags of discrete n-grams, . We then demonstrate how disjunctive activity groups can be discovered in an unsupervised manner. Finally, we lay out a framework for unsupervised discovery of predictably recurrent event motifs for activity class characterization.