Graphics Reference
In-Depth Information
respectively. For the three sequences, the true lengths correspond to 190, 190, and
217 mm for the forearm and to 203, 193, and 128 mm for the hand. Especially in
the third sequence the hand posture is poorly represented by the model, as the hand
forms a fist with the index finger pointing. However, this does not affect the robust-
ness of the system. Our evaluation furthermore shows that the motion between two
subsequent images is estimated at a typical translational accuracy of 1-3 mm, which
is comparable to the pixel resolution of the images, and a typical rotational accuracy
of 1-3 degrees. Due to its roundish shape, the rotational motion of the hand is esti-
mated less accurately than that of the more elongated forearm (the very large errors
observed for
ω
p
in the first sequence are due to sporadically occurring outliers).
A robust detection and pose estimation of the hand-forearm limb is achieved for
time intervals between subsequent model adaptations as long as 800 ms (
n
16).
The accuracy of the estimated pose and its temporal derivative is largely indepen-
dent of
n
.
The proposed spatio-temporal pose estimation method relies on a new extended
constraint line approach introduced to directly infer the translational and rotational
motion components of the objects based on the low-level spacetime stereo infor-
mation. Our evaluation of this approach in a 'tracking by detection' framework has
demonstrated that a robust and accurate estimation of the three-dimensional object
pose and its temporal derivative is achieved in the presence of a cluttered back-
ground without requiring an initial pose.
=
7.4 Three-Dimensional Tracking of Human Body Parts
In this section, the three-dimensional pose estimation and tracking system described
in Sect.
2.2.3
are evaluated quantitatively and compared with the individual com-
ponents based on a trinocular small-baseline data set. The mean-shift tracking
method outlined in Sect.
2.3.2
is evaluated on the same image sequences. A de-
scription of the data set, an evaluation of different system configurations, and a
discussion of the obtained experimental results are provided. The presentation is
adopted from Hahn et al. (
2010a
) (cf. Sects.
7.4.1
-
7.4.6
) and from Hahn et al.
(
2010b
) (cf. Sect.
7.4.7
). Further details are provided by Barrois (
2010
) and Hahn
(
2011
).
7.4.1 Image Acquisition
To obtain reliable depth information of the investigated scene, a trinocular camera
system (cf. Fig.
7.6
) similar to the SafetyEYE protection system is used. From the
perspective of the desired application of the camera system as a safety system in
an industrial environment, it is required that no in situ calibration is performed but
that the camera system is calibrated once immediately after it has been built. This
Search WWH ::
Custom Search