ws
ion, image
multisensor
ion for the
ping of the
5, which we
TISENSOR
liscuss the
Framework
using our
most wide
es [3]. The
work is a
„a. AN. Uf
) process of
0 the set of
input/output
‘es that are
iultaneously
xpress this
t only the
have the
t is obvious
t problem-
> set of data
1es. Due to
the set of
1 the set of
on the basis
ity, we shall
types and
there is the
ng levels in
It usually
cheme (see
Symbolic level
1
Soft symbolic level
1
Feature level
T
Measurment level
Fig. 1. Multi-sensor data processing semantic
levels.
The arrows in this scheme show the
direction of information movement: from the
measured signal through the feature-based
descriptions to the resultant symbolic description.
Last years, the symbolic level is often considered as
a union of two sub levels: the soft (probabilistic or
fuzzy) description level and the proper symbolic
level corresponding to the final decision making
about the elements of the observed scene. One
supposes that the processing starts from the
separate processing of measured data from each
sensor and the fusion takes place at one of the
highest levels of this scheme. Since the fusion has
being done, the further processing of the fused data
is executed if it is required.
Thus, we have the interlevel (down to up)
and on-level transformations of data. The interlevel
transformations extract the usable information from
data of lower level and the on-level procedures
realize the proper data fusion. Today the most
attractive fusion levels are the symbolic and the
feature levels. The measurement (image, pixel)
level is usually rejected because it is not easy to
provide the accurate co-registration of multi-sensor
image data. However, it takes place in the remote
sensing case where the co-registration is accurate
enough. So, the measurement level must be also of
consideration as a fusion level.
We agree all mentioned points but the
introduced terms are too generic and uncertain to
specify the proper data types and processing
procedures. To define the required set of frame
types we need to analyze the problem more detail.
The first new point we propose to consider is that
any data at any processing level represent a
Structure of some elements. According to this point
of view, one can say that fig.1 describes the levels
of data abstraction for elements but not for
structures. The analogous scheme for structures
may have, for example, the following form (fig.2):
385
Temporal-tructured data
T
3D-Spatial-structured data
T
2D-Spatial-structured data
I
Raster data
Fig. 2. Data organization levels.
The basic level of this scheme corresponds
to the sensor-generated level of data organization.
Second level corresponds to the segmented image
data. The third level includes the 3D-scene
descriptions and the top level of this scheme deals
with the time-varying 3D-world. In general, this
scheme is not right because it mixes two different
types of data organization: spatial and temporal.
These organization types are independent from
each other. So, we have to consider three spatial
organization levels and three different levels for
time-varying data: raster, 2D-structured and 3D-
structured. However, it is the /ong-range remote
sensing case where the 2D-temporal data (raster or
structured) are seldom of use. So, we adopt the
scheme above (fig. 2).
All of levels (fig. 2) are the abstraction
levels too, however, there is no any correlation
between these levels and levels mentioned before.
For example, the objects of 3D time-varying world
description may be characterized by the feature
vectors and, conversely, symbolic data may be
stored in the raster form. Thus, we have the 2D-
space of combinations of possible data structures
with different element types. So, any data type can
be described with the use of two "co-ordinates" -
semantic level and organization level. Let's note that
the fusion may also be executed at any level of
scheme 1 and simultaneously at any level of
scheme 2. So, now we can not more characterize
the fusion operators as on-level operators in
scheme 1. The fusion procedure must preserve the
both of data type co-ordinates.
The introduced 2D-space is just enough to
specify the set of data frames. Nevertheless, one
more additional point must be outlined before. It is
the rare case when the object at feature semantic
level is characterized by only one feature. Usually
the feature level presumes the feature vector of
some dimension to be corresponded with each
element of data structure. Analogously, the soft
symbolic description associates any structural
element with the vector of fuzzy measures
(probabilities) of hypothesis. The dimension of this
vector is equal to the number of object classes
known for the proper system. Finally, the
measurement element can be also the vector of
some dimension, e.g. the TV-signal provides the
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B5. Vienna 1996
