Remarks on a Real-Time, Noncontact, Nonwear, 3D Human
Body Posture Estimation Method
Kazuhiko Takahashi, Tatsumi Sakaguchi, and Jun Ohya
ATR Media Integration and Communications Research Laboratories, Kyoto, Japan 619-0288
SUMMARY
This paper proposes a new real-time method of esti-
mating human postures in three dimensions from trinocular
images. The proposed method extracts feature points of the
human body by analyzing contours of human silhouettes.
The feature points are extracted by using the subtraction
images when self-occlusions occur in the silhouette images.
Dynamic compensation is carried out with a Kalman filter
so that all feature points are tracked. The 3D coordinates of
the feature points are reconstructed by considering the
geometrical relationship between the three cameras. Ex-
perimental results confirm both the feasibility and the ef-
fectiveness of the proposed method. © 2000 Scripta
Technica, Syst Comp Jpn, 31(14): 110, 2000
Key words: 3D human body posture estimation,
trinocular image, human silhouettes, dynamic compensa-
tion.
1. Introduction
While people usually communicate with each other
by speech, nonverbal information, such as gestures and
body posture, is also useful for human-to-human commu-
nication. From the viewpoint of humanmachine interface
engineering, measurement, recognition, and understanding
of nonverbal information have become an important re-
search subject [1]. Therefore, estimating human motions
and/or postures in real time with a computer is necessary
for making interactions between humans and machines
more natural, and applications of such technology are ex-
pected in various fields such as entertainment, education,
and welfare.
Approaches to measuring the posture or the motion
parameters of a human body can be roughly classified into
contact and noncontact types. Contact-type devices, such
as magnetic sensors, are easy to use and the acquired data
are quite accurate, but they cause stress and are hard to
handle since the person using them must wear them at all
times. With non-contact-type systems, in contrast, the im-
age analysis methods employed can lighten the burden on
users. Accordingly, computer-vision-based technologies for
real-time estimation of human postures in 3D are essential.
Several methods of estimation using image analysis
methods have been studied. In estimating 3D human pos-
tures from a still scene [24] or tracking 3D human motions
from image sequences [510], methods of template match-
ing between human body part models and human images
are often utilized with distance information and/or con-
straint conditions. However, these methods are not easily
applied to arbitrary human postures since they require
various models, knowledge, and constraints. The methods
proposed in Refs. 11 to 15 can deal with 3D posture but
require body models and high computing costs, and the
processing rate is only suitable for off-line jobs. In fact,
there are few methods that work in real time. Pfinder [16]
uses contour and color information of the body to track
body parts based on kinematics and/or dynamics models of
the human body, but it requires a high computational cost.
W4 [17], another real-time method, can only estimate hu-
man postures in 2D.
The authors have also proposed a real-time method
[1820] for estimating 3D human body postures. In this
© 2000 Scripta Technica
Systems and Computers in Japan, Vol. 31, No. 14, 2000
Translated from Denshi Joho Tsushin Gakkai Ronbunshi, Vol. J83-D-II, No. 5, May 2000, pp. 13051314
1