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 131

system to investigate the relationship between the activity recognition
algorithms and the architectures required to perform these tasks in real
time. The chapter describes the proposed activity recognition method that
consists of a distributed algorithm and a data fusion scheme for two and
three-dimensional visual analysis, respectively. The authors analyze the
available data independencies for this algorithm and discuss the potential
architectures to exploit the parallelism resulting from these independencies.
Introduction
Three-dimensional motion estimation has a wide range of applications, from
video surveillance to virtual animation. Therefore, reconstruction of visual
information from multiple cameras and its analysis has been a research area for
many years in computer vision and computer graphics communities. Recent
advances in camera and storage systems are main factors driving the increased
popularity of multi-camera systems. Prices continue to drop on components, e.g.,
CMOS cameras, while manufacturers have added more features. Furthermore,
the evolution of digital video, especially in digital video storage and retrieval
systems, is another leading factor.
In this chapter, we focus on real-time processing of multiple views for practical
applications, such as smart rooms and video surveillance systems. The increased
importance of applications requiring fast, cheap, small and highly accurate smart
cameras necessitates research efforts to provide efficient solutions to the
problem of real-time detection of persons and classification of their activities. A
great effort has been devoted to three-dimensional human modeling and motion
estimation by using multi-camera systems in order to overcome the problems due
to the occlusion and motion ambiguities related to projection into the image plane.
However, introduced computational complexity is the main obstacle for many
practical applications.
This chapter investigates the relationship between the activity recognition
algorithms and the architectures required to perform these tasks in real time. We
focus on the concepts of three-dimensional human detection and activity
recognition for real-time video processing. As an example, we present our real-
time human detection and activity recognition algorithm and our multi-camera,
test bed architecture. We extend our previous 2D method for 3D applications and
propose a new algorithm for generating a global 3D human model and activity
classification.