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 134

1993), and INRIA's system (Faugeras, 1993) are some of the early real-time
stereo systems. Yet, they do not provide a complete video-rate output of range
as dense as the input image with low latency. Another major problem is that the
depth maps obtained by current stereo systems are not very accurate or reliable.
At Carnegie Mellon, a video rate stereo machine was developed (Kanade et al.,
1996) where multiple images are obtained by multiple cameras to produce
different baselines in lengths and in directions. The multi-baseline stereo method
consists of three steps. The first step is the Laplacian of Gaussian (LOG) filtering
of input images. This enhances the image features, as well as removes the effect
of intensity variations among images due to the difference in camera gains,
ambient light, etc. The second step is the computation of sum-of-squares
differences (SSD) values for all stereo image pairs and the summation of the
SSD values to produce the sum-of-sum-of-squares differences (SSSD) func-
tion. Image interpolation for sub-pixel re-sampling is required in this process. The
third and final step is the identification and localization of the minimum of the
SSSD function to determine the inverse depth. Uncertainty is evaluated by
analyzing the curvature of the SSSD function at the minimum. All these
measurements are done in one-tenth sub-pixel precision. One of the advantages
of this multi-baseline stereo technique is that it is completely local in its
computation without requiring any global optimization or comparison.
Schreer et al. (2001) developed a real-time disparity algorithm for immersive
teleconferencing. It is a hybrid and pixel recursive disparity analysis approach,
called hybrid recursive matching (HRM). The computational time is minimized
by the efficient selection of a small number of candidate vectors, guaranteeing
both spatial and temporal consistency of disparities. The authors use cameras,
mounted around a wide screen, yielding a wide-baseline stereo geometry. The
authors compare the real-time performance of their algorithm with a pyramid
approach, based on multi-resolution images, and with a two stage hierarchical
block-matching algorithm. The proposed method can achieve a processing speed
of 40 msecs per frame for HRM algorithm in the case of sparse fields with block
sizes of 8 by 8 pixels.
In Koschan & Rodehorst's (1995) work, parallel algorithms are proposed to
obtain dense depth maps from color stereo images employing a block matching
approach. The authors compare single processor and multiple processor perfor-
mance to evaluate the profit of parallel realizations. The authors present
computing times for block matching and edge-based stereo algorithms for
multiple processing units that run in parallel on different hardware configura-
tions.
A commercial system with small-baseline cameras has been developed by
Videre Design. From two calibrated cameras, the system generates a disparity