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 133

Basic Algorithm Steps
3D scene synthesis and analysis, by using visible light and multiple cameras, has
been studied by many researchers. Before considering some of these methods,
it is beneficial to review general stereo vision issues with respect to their real-
time applicability. There are three basic problems, namely correspondence
(disparity map), reconstruction, and rendering.
Disparity map generation
One well-known technique for obtaining depth information from digital images
is the stereo technique. In stereo techniques, the objective is to solve the
correspondence problem, i.e., to find the corresponding points in the left and right
image. For each scene element in one image, a matching scene element in the
other image is identified. The difference in the spatial position of the correspond-
ing points, namely disparity, is stored in a disparity map. Whenever the corre-
sponding points are determined, the depth can be computed by triangulation.
Attempts to solve the correspondence problem have produced many variations,
which can be grouped into matching pixels and matching features, e.g., edges.
The former approach produces dense depth maps while the latter produces
sparse depth maps. The specific approach desired depends on the objective of
the application. In some applications, e.g., the reconstruction of complex
surfaces, it is desirable to compute dense disparity maps defined for all pixels in
the image. Unfortunately, most of the existing dense stereo techniques are very
time consuming.
Even though stereo vision techniques are used in many image processing
applications, the computational complexity of matching stereo images is still the
main obstacle for practical applications. Therefore, computational fast stereo
techniques are required for real-time applications. Given the algorithmic com-
plexity of stereo vision techniques, general purpose computers are not fast
enough to meet real-time requirements which necessitate the use of parallel
algorithms and/or special hardware to achieve real-time execution.
Two main performance evaluation metrics are throughput, that is, frame rate
times frame size, and range of disparity search that determines the dynamic
range of distance measurement. There is still a great deal of research devoted
to develop stereo systems to achieve the desired performance. The PRISM3
system (Nishihara, 1990), developed by Teleos, the JPL stereo implemented on
DataCube (Matthies, 1992), CMU's warp-based multi-baseline stereo (Webb,