Redwood Neuroscience
Title "Hierarchical stochastic models
for learning and recognizing human activities.”
Dr.
Hung Bui
Artificial
SRI
International
Abstract:
Understanding the typical patterns in the user’s daily
activities and routines from low-level sensory data is an important problem in
many applications such as smart environment and intelligent assistance technology.
Due to the sequential nature that activities unfold in time, dynamic models
such as the Hidden Markov Model (HMM) has been widely used for activity
learning and classification. However, as the space of activities becomes more
complex, it is necessary to make use of known structural properties of
activities, especially their hierarchical organization and decomposition.
In this talk, I will describe several hierarchical
extensions of the basic HMM including the Abstract HMM (AHMM), and the
Hierarchical HMM (HHMM) for modeling a hierarchy of activities at different
levels of abstraction and resolution. Inference in these models is considerably
more complex due to the dependency between the hidden layers in the model. I
will discuss techniques for offline parameter learning and online real-time
recognition and show how they can be applied to activity modeling from human
movement data. Experiments in this domain demonstrate the advantages of using
these models over the basic HMM.