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 137

dynamic scenes. Image quality is increased with the accuracy of the disparity
maps provided with the recorded video streams. In Li et al. (2003), a simulta-
neous visual hull reconstruction and rendering algorithm is proposed by exploiting
off-the-shelf graphics hardware.
Beside special hardware, the use of parallel algorithms can't be avoided to
achieve high-speed rendering applications. Early systems, such as Pixar's
CHAP (Levinthal & Porter, 1984) and the commercially available Ikonas
platform (England, 1986), had SIMD processors that could process vertex and
pixel data in parallel. Programmable MIMD machines that could process
triangles in parallel, such as the Pixel Planes (Fuchs et al., 1989) and the SGI
InfiniteReality, had complex low-level custom microcodes and were rarely used.
CPU vendors began to introduce graphics-oriented SIMD processor extensions
into general purpose CPU designs. Examples of these extensions include Intel's
MMX/SSE instructions, AMD's 3DNow architecture, and Motorola's AltiVec
technology. Although such extensions accelerate several graphics operations,
more sophisticated graphics coprocessors, e.g., processors that can support
rendering pipelines, are needed. Such a system has been developed by Sony. The
company designed a custom dual-processor SIMD architecture for graphics
called the Emotion Engine (Kunimatsu et al., 2000).
A detailed survey on graphics hardware can be found in Thompson et al. (2002).
The basic steps for image rendering are shown in Figure 1. The input of the
graphics hardware is raw geometry data specified in some local coordinate
system. The hardware transforms this geometry into world space and performs
lighting and color calculations followed by a texture step. The hardware converts
the vector-based geometry to a pixel-based raster representation, and the
resulting pixels are sent into the screen buffer.
Human Detection and Activity Recognition
In this section, we present related work by classifying the research in terms of
visual analysis of multiple cameras, i.e., projection of 3D models onto 2D images
versus 3D visual reconstruction from stereo images. The former involves
Figure 1. Graphics pipeline.