Three-Dimensional Pose Estimation and Segmentation Methods - 3D Computer Vision: Efficient Methods and Applications - page 127

estimated three-dimensional centre point c of the image-based mean-shift stage. For

all subsequent iterations, the ellipsoid model is moved to the new position

n = 1 s n ·

· g( s n , c j ) q (id) (iBin( s n ))

c j + 1 = , (2.35)

n = 1 g( s n , c j )

· q (id) (iBin( s n ))

where c j is the previous centre position. In the mean-shift procedure a truncated
Gaussian kernel g( s n , c j ) is used according to Cheng ( 1995 ), for which the weight
decreases with increasing distance from the ellipsoid centre c j . The mean-shift-
based three-dimensional tracking approach incorporates an appearance weighting

q (id) (iBin( s n )) obtained by looking up the appearance probability of the moving
three-dimensional point s n in the target model q (id) , such that a three-dimensional

point with an appearance similar to the target appearance is assigned a higher
weight. Figure 2.16 (right) depicts all moving three-dimensional points and the
final result of the two-stage mean-shift procedure for three-dimensional track-
ing.

2.3.3 Segmentation and Spatio-Temporal Pose Estimation

The problem of three-dimensional scene segmentation along with the detection and
pose estimation of articulated objects has been addressed primarily in the context of
human motion capture (cf. Sect. 2.2.1.2 for an overview). A technique for model-
based three-dimensional human body tracking based on the ICP algorithm is pre-
sented by Knoop et al. ( 2005 ). Normal optical flow is used by Duric et al. ( 2002 )
for positional short-term prediction in an image-based system for the detection and
tracking of human body parts. The motion of a camera through a scene is estimated
by Gonçalves and Araújo ( 2002 ) based on a combined analysis of stereo correspon-
dences and optical flow.
This section describes a vision system for model-based three-dimensional detec-
tion and spatio-temporal pose estimation of objects in cluttered scenes proposed by
Barrois and Wöhler ( 2008 ). The presentation is adopted from that work. Further de-
tails are provided by Barrois ( 2010 ). As low-level features, this approach requires a
cloud of three-dimensional points attributed with information about their motion and
the direction of the local intensity gradient. These features are extracted by space-
time stereo based on local image intensity modelling according to Sect. 1.5.2.5 .
After applying a graph-based clustering approach to obtain an initial separation be-
tween the background and the object, a three-dimensional model is adapted to the
point cloud based on an ICP-like optimisation technique, yielding the translational,
rotational, and internal degrees of freedom of the object. An extended constraint
line approach is introduced which allows one to estimate the temporal derivatives of
the translational and rotational pose parameters directly from the low-level motion
information provided by the spacetime stereo data.