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 140

single moving person while our hierarchical and parallel graph matching and
HMM-based activity recognition algorithms enable multi-person detection and
activity recognition.
W4 is another real-time human tracking system (Haritaoglu et al., 1998) where
the background information should be collected before the system can track
foreground objects. The individual body parts are found using a cardboard model
of a walking human as reference. There are a few works that aim to obtain a
more compact representation of the human body without requiring segmentation.
Oren et al. (1997) used wavelet coefficients to find pedestrians in the images,
while Ozer & Wolf (2001) used DCT coefficients that are available in MPEG
movies to detect people and recognize their posture.
Self-occlusion makes the 2D tracking problem hard for arbitrary movements and
some of the systems assume a priori knowledge of the type of movement. The
authors (Wolf et al., 2002) developed a system by using ellipses and a graph-
matching algorithm to detect human body parts and classified the activity of the
body parts via a Hidden Markov Model-based method. The proposed system can
work in real-time and has a high correct classification rate. However, a lot of
information has been lost during the 2D human body detection and activity
classification. Generating a 3D model of the scene and of the object of interest
by using multiple cameras can minimize the effects of occlusion, as well as help
to cover a larger area of interest.
3D
One of the early works on tracking articulated objects is by O'Rourke & Badler
(1980). The authors used a 3D model of a person made of overlapping spheres.
They synthesized the model in images, analyzed the images, estimated the pose
of the model and predicted the next pose. Hogg (1983) tracked human activity
and studied periodic walking activity in monocular images. Rehg & Kanade
(1995) built a 3D articulated model of a hand with truncated cones. The authors
minimized the difference between each image and the appearance of the 3D
model. Kakadiaris & Metaxas (1995; 1996) proposed a method to generate the
3D model of an articulated object from different views. The authors used an
extended Kalman filter for motion prediction. Kuch & Huang (1995) modeled the
hand with cubic B-splines and used a tracking technique based on minimization.
Gavrila & Davis (1996) used superquadrics to model the human body. They used
dynamic time warping to recognize human motion.
Munkelt et al. (1998) used markers and stereo to estimate the joints of a 3D
articulated model. Deutscher et al. (1999) tracked the human arm by using a
Kalman filter and the condensation algorithm and compared their performances.