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 141

Bregler & Malik (1998) proposed a new method for articulated visual motion
tracking based on exponential maps and twist motions.
Most of the previous work for human detection depends highly on the segmen-
tation results and mostly motion is used as the cue for segmentation. Most of the
activity recognition techniques rely on successful feature extraction and pro-
posed approaches are generally only suitable for a specific application type. The
authors have developed a system that can detect a wide range of activities for
different applications. For this reason, our scheme detects different body parts
and their movement in order to combine them at a later stage that connects to
high-level semantics.
Real-Time 3D Analysis
This section is devoted to our proposed method of real-time 3D human motion
estimation. Multi-camera systems are used to overcome self-occlusion problems
in the estimation of articulated human body motion. Since many movements
become ambiguous when projected into the image plane and 2D information
alone can not represent 3D constraints, we use multiple views to estimate 3D
human motion. First, we discuss real-time aspects of 3D human detection and
activity recognition. In the following two subsections we show how these aspects
affect the software and hardware architectures. A detailed analysis of our 3D
human detection/activity recognition algorithm and a testbed architecture for this
particular algorithm are given in the last subsection.
Real-Time Aspects
Real-time aspects are critical for the success of the algorithm. The authors
analyze various aspects and challenges of 3D human detection and activity
recognition algorithms. These include: the instruction statistics, branch behavior,
and memory access behavior of different program parts, e.g., stereo matching,
disparity map generation, reconstruction, projection, 2D/3D human-body part
detection, 2D/3D tracking, 2D/3D activity recognition, etc., in the Section
“Algorithmic Issues.” Note that it is essential to understand the application
behavior to develop efficient hardware for a 3D camera system. Hardware
related aspects and challenges for real-time applications are discussed in the
Section “Hardware Issues.” Decisions such as the number of processors in the
system, the topology of the processors, cache parameters of each processor, the
number of arithmetic logic units, ISA (instruction set architecture), etc., all rely