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 146

tional simulation. During the experiment, we use the TriMedia Software Develop
Kit version tcs2.20 that includes a compiler tmcc, an assembler tmas, a linker
tmld, a simulator tmsim, an execution tool tmrun, and a simulator tmprof. The
TriMedia system is running on a Dell Precision-210 computer with two TriMedia
reference boards. The TriMedia boards can communicate via shared memory,
which enables fast data communication for stereo vision applications, e.g.,
disparity map generation.
Direct Algorithm for Human Gesture Recognition
In this subsection, we discuss in more detail an exemplar approach for human
detection and activity recognition in the light of previously mentioned algorithms
and real-time aspects. Most of the activity recognition systems are suitable for
a specific application type. The presented example can detect a wide range of
activities for different applications. For this reason, the scheme detects different
body parts and their movement in order to combine them at a later stage that
connects to high-level semantics. Each human body part has its own freedom of
motion and the activity recognition for each part is achieved by using several
Hidden Markov Models in parallel. Real-time performance of two and three-
dimensional activity recognition techniques are compared for this particular
example.
2D
A - Low-level Processing:
This section presents the proposed algorithm for the detection of the human body
parts. The algorithm blocks are displayed in Figure 3. A more detailed explana-
tion of our algorithm can be found in Ozer et al. (2000).
Background elimination and color transformation: The first step is the
transformation of pixels into another color space regarding to the applica-
tion. Background elimination is performed by using these transformed pixel
values for the current and background images.
Skin area detection: Skin areas are detected by comparing color values to a
human skin model. We use a YUV color model where chrominance values
are down-sampled by two.
Segmentation of non-skin areas and connected component algorithm:
The foreground regions that are adjacent to detected skin areas are
extracted and corresponding connected components are found. We com-