A novel sequence representation for unsupervised analysis of human activities

We present a novel activity representation as bags of event n-grams to extract global structural information of activities using their local event statistics. Exploiting this representation, we present a computational framework for unsupervised activityclass discovery, activity classification and anomalous activity detection. To this end, we model activity-classes as maximally similar activity-cliques in an edge-weighted graph of activities, and present a graph-theoretic method for their efficient discovery. Moreover, to detect irregular behaviors in active environments, we formulate a definition of anomalous activities, and propose an information theoretic method to explain the detected anomalies in a maximally informative manner. Finally, for the purposes of online activity classification and anomaly detection, we model the discovered activity-classes as variable memory Markov chains, and propose a method to find their constituent subsequences that are maximally exclusive amongst the discovered activity-classes. Results over data-sets collected from multiple environments are presented to demonstrate the competence of our framework.