|
Human gesture
recognition has been an interesting topic in the research of human behavior
understanding, human-machine interaction, machine control, surveillance, etc.
For the efficient and reliable study in these fields, gesture database is
significantly required for analyzing the characteristics of human
gesture and verifying or evaluating the developed algorithms. Nevertheless,
there has been no database contains various full-body gesture data
performed by many subjects and obtained by 2D and 3D capture system at the
same time. The 2D and 3D Full-Body
Gesture (FBG) database is collected for supporting 2D and 3D gesture
recognition and its related studies.
The FBG database consists of 14 representative normal gesture data in daily life
20 old persons of 60~80 ages and 10 abnormal gesture data which may happen
in emergency for 20 younger persons.The database contains 3D motion data and
3 pairs of stereo-video data taken at 3 different drections for each gesture
using 3D motio capture devices and stereo cameras.
In addition, the 2D silhouette data is synthesized by separating a subject
and background in 2D stereo-data andvideo saved as binary mask images.
All data are saved according to the file naming rule in order that user can
easily recognize information of subject, capture condition and file. Table 1.
Distribution of subjects for normal gestures by age and sex
|
|
[ Back to FBG Home ] |
- 3D Gesture Capture System
In order to obtain
3D motion data of various gestures, we exploit the Eagle Digital System of Motion Analysis Co. The Eagle Digital System consists of Eagle Digital Cameras, the EagleHub, and EVaRT software,
which can capture subject’s motion with high
accuracy. The gesture data is captured through 2 sessions. In the first session, we captured normal gesture data in the studio which is of 10.1m (width) x 11.1m (length) x 5.1m (height) and has 3 sets of digital motion cameras located in each side of studio. In the second session, we captured abnormal gesture data in the studio which is of 6.0m (width) x 9.0m (length) x 4.0m (height) and has 2 sets of digital motion cameras located in each side of studio. All subjects wear a black and blue
color suit, on which 33 markers reflecting light from LED of 3D cameras are
attached. All 3D cameras are synchronized to the system clock and the 3D
position of makers is obtained at 60 frames per second. Figure 1 and Figure 2
show the studios for capturing 2D and 3D gesture data and body suit with makers,
respectively.
|
|
Figure
1. Studio overview
|
|
Figure
2. Body suit for motion capture
- 2D Gesture Capture System
We captured 2D and 3D gesture data,
simultaneously. Especially, 2D video data are captured with stereo camera system
(STH-MDCS2) made by Videre
Design in order to be used not only for gesture recognition system with
mono camera, but also for that with stereo camera. 2D stereo camera systems are
4m away from a subject and placed at +45, -45, 0 degrees for obtaining gestures
at 3 different directions. This system includes 6.0mm, F 1.4, C mount lenses
and the stereo baseline of camera is 9cm. It captures uncompressed video at 320
x 240 resolution, color and 30 frames per second. The 3 pairs of stereo-video
data are captured by progressive scan mode and saved as uncompressed ‘AVI file format. Before capturing
gestures images, each camera is calibrated with black and white pattern, which
is usually exploited for calibration process. In order to easily separate
subject from background, several pieces of white fabric are used.Those cover right, left and rear sides of
studio, which appear in views of 3 stereo cameras.
|
|
[ Top ] |
The Captured Gestures
Although the
human gestures in everyday life are a wide variety, we define the most common
14 normal gestures: (1) sitting on a chair, (2) standing up from a chair, (3) walking
at a place, (4)?touching a knee and a waist, (5) raising a right hand, (6)
sticking out a hand, (7) bending a waist, (8) sitting on the floor, (9) getting
down on the floor, (10) lying down on the floor, (11) waving a hand, (12)
running at a place, (13) walking forward, and (14) walking circularly. We ask a
subject to behavior with his/her own style and captured 14 gestures with 3D
motion capture cameras and 3 sets of stereo cameras at 3 different directions.
Figure 3 shows the examples of normal gestures.
|
Figure
3. Examples of normal gestures
|
Figure
4. Examples of abnormal gestures
|
|
For user's convenience, we define a
simple file naming rule for the 3D motion data, the 2D stereo-video data and
the 2D silhouette data.
From only file name, user can realize the information of a subject such as
subject index, sex and age; that of capture condition such as captured gesture
index, camera direction and camera view; that of file such as frame number,
type and extension. Table 2 shows the file name and its meaning as the file
naming rule.
Table
2. File naming rule
|
||||||||||||||||||||||||||
|
|
|
[ Back to FBG Home ] |
|
Contact: gesturedb@image.korea.ac.kr |