Abstract

In this paper, we present techniques for characterizing complex, multi-tasked activities that require both exemplars and models. Exemplars are used to represent object context, image features, and motion appearances to label domainspecific events. Then, by representing each event with a unique symbol, a sequence of interactions can be described as an ordered symbolic string. A model of stochastic contextfree grammar, which is developed using underlying rules of an activity, provides the structure for recognizing semantically meaningful behavior over extended periods. Symbolic strings are parsed using the Earley-Stolcke algorithm to determine the most likely semantic derivation for recognition. Parsing substrings allows us to recognize patterns that describe high-level, complex events taking place over segments of the video sequence. We introduce new parsing strategies to enable error detection and recovery in stochastic context-free grammar and methods of quantifying group and individual behavior in activities with separable roles. We show through experiments with a popular card game how high-level narratives of multi-player games as well as identification of player strategies and behavior can be extracted in real-time using vision.

Citations