Probabilistic Grammar Induction for Long Term Human Activity Parsing

Abstract

We present a method of representing human activities as Probabilistic Context Free Grammars(PCFGs). Our method will allow these grammars to be learned from any source of data that describe sequences of human actions. We describe how representing human activities as PCFGs will allow them to be used for multiple proposed applications. The method proposed is interpretable such that the representation of an activity can be edited by a human annotator for further increase in performance. We also introduce a method of simulating realistic sequences of human actions, and describe how realistic noise is injected into this data. We propose methods of inducting grammars from this synthetic data and experiments to evaluate both the data and the ability of PCFGs to represent human activities.