mbul 2004
n of local
t frequent
ms: If the
ssigned to
reated; b)
inderstood
le version
m number
entries for
instant the
y of data
ction of a
ossible to
notone or
iputational
east more
flexible to
cement of
olding of
daries are
al become
here.
E MAPS
clustering
ion model
itical size.
high-level
'se colour-
acking of
esponding
istribution
mogeneity
id position
‘es us a 5-
separating
imilar way
| recursive
s. So, we
space that
he spatio-
a distance-
Instead of
k between
erties for
divisions).
"w regions
Both local
aw. So, if
i complex
oor scene,
t, the local
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B5. Istanbul 2004
In accompanying video sequence one can see some results of
our image processing method for a static camera and mobile
objects. In this case, we have a fixed geometrical background
with a mobile foreground. Nevertheless, the fixed character of
the background, there appear some meaningful facts relative to
the image segmentation which are linked to variable lightening
conditions Indeed, small lightening differences are increased by
concentrating available information around middle values. This
sequence provides an analysis of mobile objects with easy
kinematic properties relative to depth variations which can be
obtained by means of an evaluation of scale parameter (by
using methods appearing in [Vil02]).
Furthermore, it is possible to observe small variations in
background giving us a better understanding of depth. The
artificial model is linked to a discrete version of traditional
human perception, where graduations are usually continuous
and more slower.
Some advantages for the developed artificial system are the
following ones: it is possible to give a explicit description of
unfolding and regrouping processes involving to the
background (including watershed effects in the ground or sky,
e.g.), and, mainly, it is possible to include an artificial analogue
to the role played by depth planes. In this case, depth planes are
linked not to architectural elements (as it is usual in 3D
Reconstruction), but to a colour segmentation with supports an
optical and metric information.
The analysis for small variations of lightening and small
motions in forest is similar, but including now reinforced
variations of colour. These reinforced variations are supported
in this case on a distance map able of reproducing small effects
such as leaves frightening or more coarse variations, depending
on the parameters selection. In this way, we obtain an adaptive
approach to a more realistic perception, nevertheless the
reduction accompanying any segmentation task.
6. CONCLUSIONS AND FUTURE WORK
In this work we have sketched some general principles for a
hybrid, hierarchised and multilayered approach to video
segmentation. A feedback between low- and high- level
processing is revealed as a guide to combine accurate results
based on local histograms with meaningful regions.
Segmentation is performed in a semiautomatic way on a 5D
space which is managed by using a binary tree search.
Centroids of segmented regions provide sites for a generalized
Voronoi diagram.
Distance maps are introduced for supporting the identification,
evaluation and tracking of meaningful regions. The average
between consecutive frames allows us to extend traditional
accurate spatial methods to a coarse but reliable approach to the
segmentation problem.
Some standing issues where more accurate results would be
convenient concern to exploiting the duality sketched between
the structure of boundaries and the extended Voronoi diagram,
linked to centroids, the analysis of optimal clustering functions
for regions matching, the comparison of coarse live results with
finer results corresponding to 2" order differential spatio-
temporal operators (D’Alembertian of a Gaussian) linked to
mobile data. Further applications to contents retrieval and
tracking in video sequences are currently in development.
