Real-Time Analysis of Human Body Parts and Gesture-Activity Recognition in 3D - 3D Modeling and Animation: Synthesis and Analysis Techniques for the Human Body - page 148

of areas, b) relative position and orientation, and c) adjacency information
between nodes with overlapping boundaries or areas.
Graph matching: Each extracted region modeled with ellipses corresponds to
a node in the graphical representation of the human body. Face detection
allows the formation of initial branches to start efficiently and reduces their
complexity. Each body part and meaningful combinations represent a class
w where the combination of binary and unary features are represented by
a feature vector X and computed off-line. Note that feature vector elements
of a frame node computed online by using ellipse parameters change
according to body part and the nodes of the branch under consideration. For
example, for the first node of the branch, the feature vector consists of
unary attributes. The feature vector of the following nodes also includes
binary features dependent on the previously matched nodes in the branch.
For the purpose of determining the class of these feature vectors, a
piecewise quadratic Bayesian classifier with discriminate function g(X) is
used. The generality of the reference model attributes allows the detection
of different postures while the conditional rule generation r decreases the
rate of false alarms. The computations needed for each node matching are
then a function of the feature size and the previously matched nodes of the
branch under consideration. The marked regions are tracked by using
ellipse parameters for the consecutive frames and a graph matching
algorithm is applied for new objects appearing in the other regions. Details
of the graph matching algorithm can be found in Ozer & Wolf (2002b).
B - High-level Processing:
This section covers the proposed real-time activity recognition algorithm based
on Hidden Markov Models (HMMs). HMM is a statistical modeling tool that
helps to analyze time-varying signals. Online handwriting recognition (Sim &
Kim, 1997), video classification and speech recognition (Rose, 1992) are some
of the application areas of HMMs. Only a few researchers have used the HMM
to recognize activities of the body parts. It is mainly used for hand gestures
(Starner & Pentland, 1995). Parameterized HMM (Wilson & Bobick, 1999) can
recognize complex events such as an interaction of two mobile objects, gestures
made with two hands (e.g., so big, so small), etc. One of the drawbacks of the
parameterized HMM is that for complex events (e.g., a combination of sub-
events) parameter training space may become very large. In our application, we
assume that each body part has its own freedom of motion and the activity
recognition for each part is achieved by using several HMMs in parallel.
Combining the outputs of the HMMs to generate scenarios is an application
dependent issue. In our application environment, smart room, we use the
Mahalanobis distance classifier for combining the activities of different body