Graphics Reference
In-Depth Information
Fig. 2.12
Construction of the
reference templates. The
forearm rectangle is cropped
from all camera images and
transformed and interpolated
to a predefined size and
orientation
2.2.3.5 Verification and Recovery of the Pose Estimation Results
Pose Verification
To verify and rate a set of calculated pose parameters, three
quality measures are used. The first quality criterion is the point distance of the seg-
mented point cloud to the model. Accordingly, a thin hull around the object model
is used to determine the points inside the object hull. The weighting value
σ
p
is the
quotient of object points inside the hull and all segmented points.
The second quality criterion is the orientation similarity
σ
o
, which is computed
for a three-dimensional pose by extracting the contour of the three-dimensional
model according to Sect.
2.2.3.1
and projecting it into the images acquired by the
camera system. For calculating the quality of a contour, the algorithm 'walks along'
small perpendiculars in each projected curve point of the contour. For each pixel on
the perpendicular the image gradient is calculated. The image gradient orientation,
which is between 0
◦
and 180
◦
, is then compared with the orientation of the model
contour in the current curve point itself. While the contour gradient describes a ref-
erence orientation, the grey image gradient is the compared actual orientation. By
simply counting the curve points that do not differ too much and normalising by all
projected curve points in all camera images, one obtains an intuitive quality measure
in the range
.
The third quality criterion is the appearance similarity
σ
c
for a three-dimensional
pose at time step
t
, which is based on a comparison of the current object ap-
pearance with the previous object appearances. With the three-dimensional pose
at time step
t
and the known camera parameters we crop the image of the fore-
arm in all camera images and transform the cropped images to a predefined size
and orientation. The final pose estimation result is used to add the current refer-
ence template at time step
t
to a database (cf. Fig.
2.12
). This database describes
the appearance of the tracked object in the last 20 time steps. The appearance
similarity
σ
c
of a new pose estimation result is computed based on a normalised
cross-correlation of the current object appearance at time step
t
with the previ-
ous object appearance at time step
(t
[
0
...
1
]
1
)
. Since we have three cameras, there are
three correlation results for a pose comparison, where we choose the worst of the
three as the final result. The assumptions made for this similarity measure are that
−
Search WWH ::
Custom Search