Three-Dimensional Pose Estimation and Segmentation Methods - 3D Computer Vision: Efficient Methods and Applications - page 109

( a ) Original 8-bit image I t .( b ) Absolute difference image. ( c ) Input image I t

Fig. 2.7 (scaled to

8bit)for λ = 1 according to ( 2.11 )

Hence, in contrast to their lengths, the maximal radii r 1 and r 4 of the hand and the
forearm are part of the parameter vector T and can thus be adapted to the images
during the optimisation process.

2.2.3.2 Principles and Extensions of the CCD Algorithm

The CCD algorithm introduced by Hanek ( 2001 ) and refined by Hanek and Beetz
( 2004 ) adapts a curve model to an image based on the probability distributions of
the pixel grey values on the inner and the outer side of the curve. A computationally
efficient real-time variant of the CCD algorithm is described by Panin et al. ( 2006 ).
A difference to the work of Hanek ( 2001 ) is that Hahn et al. ( 2010a )relyonan
extended input image, since in the regarded application scenario the model-based
image segmentation is challenging due to noise, a cluttered background, and the
coarse object description. In order to obtain an accurate and robust model-based
image segmentation, we take advantage of the constant camera position in our ap-
plication. Our input image I ∗ (t) is computed by
I ∗ (t) = I(t) + λ I(t) − I(t − 1 ) ,

(2.11)

where I(t) is the image at time step t and is the absolute difference

image of the current and the previous image. The factor λ defines the influence
of the absolute difference image. The influence of the absolute difference image
increases the pixel values in areas where motion occurs, which allows a more robust
segmentation, since the CCD algorithm adapts the curve model by separating the
probability distributions of the pixel grey values on the inner and the outer side of
the object curve. Another advantage is that if there is no motion, the input image
I ∗ (t) corresponds to the original image I(t) and it is still possible to fit the model.
We experimented with different values of λ in the range

| I(t) − I(t − 1 ) |

and found
empirically that the dependence of the segmentation result does not critically depend
on the value of λ . For a moving camera, the original image I(t) would have to
be used as the input image I ∗ (t) , corresponding to λ

[ 0 . 5 ,..., 3 ]

0. Figure 2.7 shows the
original camera image, the absolute difference image, and the input image obtained
for λ

=

= 1.