' require- 
likely ag- 
mapping, 
Our pro- 
nterest is 
à densely 
m size of 
00 pixels. 
ot occlu- 
re (figure 
1SOT noise 
on within 
y overlap 
eviations 
LL) 
age 1 
est (roi 
  
ig LII] 
'action of 
a natural 
odules of 
res of the 
both half 
intensity 
ulated pi- 
' differen- 
. À value 
e, à value 
s smaller 
to be fo- 
und experimentally. Choosing the threshold is a 
weak point of the procedure. 
Morphological operations 
The binary difference image will represent at least 
one large region with the unknown object. In ad- 
dition, noise and other effects will lead to a lot of 
small regions. These small regions are eliminated 
by ’opening’, i.e. the binary image is processed 
by morphological operations. By ’closing’ small 
gaps are filled and the contour is smoothed. 
Regions 
The explicit determination of regions in the 
binary image is done by a simple region growing, 
in which the 8-neighbourhood of the points is 
checked sequentially.. If the opening has not re- 
moved all small regions this can be done in this 
step by thresholding the extracted regions with a 
required minimum area of a region. 
Border line 
A second closing with a larger operator size aims 
at filling gaps in the region of interest. Further 
filling is obtained by the following procedure. If 
background pixels appear within a horizontal line 
between object pixels they are turned to object 
pixels. The same algorithm is applied the verti- 
cal lines. The border line then will appear as a 
smooth closed contour of the region of interest. 
The border line is simply extracted by binary 
edge detection and edge linking. 
First Segmentation 
The border line circumscribes a region in the dif- 
ference image which includes the object in both 
images simultaneously. Because of the displace- 
ment of the object during the acquisition of the 
sequence in each region of both images some back- 
ground is included. In the first segmentation step 
the regions within the border line are extracted 
in each of both images separately. 
Displacement estimation 
The estimation of the two-dimensional displace- 
ment or, more generally, the displacement vector 
field between the segmented regions of both ima- 
ges can be done by cross-correlation or other well- 
known matching techniques. The search for the 
corresponding parts of both regions has to take 
into account that in both regions different parts 
of background are included. 
Second Segmentation 
The displacement vector is used to eliminate the 
background in both regions individually. This 
is achieved simply by extracting those parts in 
both regions which are identified as correspon- 
ding areas within the matching step. The result 
of this second segmentation step is the region of 
interest which is identified and located in both 
images individually and which only captures the 
object. 
Within the whole task of object recognition the 
detection and location of the region of interest is 
an important process. An example which shows 
the single steps of the procedure is shown in figure 
2. Just to give an idea on some experimental data 
the thresholds and other quantities are listed. For 
the first opening and closing a 3 x 3 operator is 
used. The threshold for the determination of the 
binary difference image was 16 grey values. A 
minimum size of an object of 50 pixels was re- 
quired and in the second closing a 9 x 9 window 
was used. The estimate for the displacement was 
1 pixel vertically and 8 pixels horizontally. The 
result of the processing is the region of interest 
(the two pictures in the lower right of figure 2). 
For the identification of the object this region has 
to be analysed further. 
3. DETERMINATION OF INVARIANT 
FEATURES AND CLASSIFICATION 
The extracted region of interest is the input for 
the identification the unknown object. As sta- 
ted earlier an intermediate process is proposed to 
eliminate motion blur. Because motion blur can 
be avoided by using shuttered video cameras this 
step of processing is not obligatory. Thus a dis- 
cussion of this topic is omitted in this paper. A 
description of the techniques for the deconvolu- 
tion of motion blur can be found in Geiselmann 
(1992). 
463