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 139

discrete event systems. The 3D shape of the body parts is reconstructed by
selectively integrating the apparent contours from three mutually orthogonal
views.
Several methods are proposed for 3D motion recovery from monocular images.
DiFranco et al. (2001) describe a method for computing the 3D motion of
articulated models from 2D correspondences. The authors use kinematic
constraints based on a 3D kinematic model, joint angle limits, dynamic smoothing
and 3D key frames which can be specified by the user. In Sminchisescu & Triggs
(2001), the authors propose a method for recovering 3D human body motion from
monocular video sequences using robust image matching, joint and non-self-
intersection constraints. To reduce correspondence ambiguities, the authors use
a matching cost metric that combines robust optical flow, edge energy, and
motion boundaries. In Howe et al. (2000), the authors present a 3D motion
capture via a single camera. The method depends on prior knowledge about
human motion to resolve the ambiguities of the 2D projection.
A geometric model is an approximation of the shape and of the deformations of
the object. This model can be two-dimensional (modeling the contours of the
projections of the object in the images), or three-dimensional (modeling the
surfaces of the object). 2D shape models are generally made of curves, snakes,
segments, sticks, etc., whereas 3D shape models are either systems of rigid
bodies (spheres, superquadrics, etc.) or deformable surfaces (mesh). The
articulations may be modeled by joints or by the motion of control points of B-
splines. The choice between a 2D or a 3D model depends on the application, e.g.,
needed precision, number of cameras, and type of motion to be recognized.
2D
Several researchers work with 2D features to recognize human movement.
Gavrila (1999), Goddard (1994) and Guo et al. (1994) use model-based recog-
nition techniques, namely stick-figures, for this purpose. Other researchers who
used 2D models are Papageorgiu & Poggio (1999), Comaniciu et al. (2000) and
Isard & MacCormick (2001). Wachter & Nagel (1999) proposed a method to
track the human body in monocular sequences. Their method depends on contour
information and moving regions between frames.
Most of the work in this area is based on the segmentation of different body parts.
Wren et al. (1999) proposed a system, Pfinder, to track people in 2D by using
blobs that represent different body parts. The system uses a Maximum A
Posteriori Probability (MAP) approach to detect and track people. The authors
extend their work to obtain 3D estimates of the hands and head by applying two
Pfinder algorithms (Wren et al., 2000). Pfinder uses blob features to detect a