IEEE Consumer Electronics Magazine - January/February 2022 - 21

Figure 9. Circle detection in overlapping regions.
Any certain size circular object can be used
to carry on this process. However, circle detection
is conducted after conducting the edge
detection. Therefore, the edges of the circular
object should clearly appear in the images from
the both cameras. Generally, the objects in
images of hemispherical cameras easily distort.
But, in the case of circular objects, this distortion
does not much interfere on circular shape.
But, the other constant shape objects such as
rectangular objects are difficult to use since the
rectangular shape easily distorts in images from
hemispherical cameras. As a result, detection of
those objects becomes complicated.
Stitching
After acquiring the center points and radii for
the circles detected in Figure 9 in the previous
section, if the radii are approximately equal, the
images are translated so as to overlay the center
points of the two circles while simultaneously
performing stitching. If the difference between
the two radii exceeds a certain threshold, an
upsampling process is applied to the image containing
the smaller circle to match the radii.
Image translations were performed using an
affine transformation, which combines parallel
movement and linear transformation to perform
image scaling, rotation, and translation. The linear
transformation is represented as the product
of original coordinates ðx; yÞ and a 2 2 matrix,
and the parallel movement is represented as the
addition of a 2-D vector ½tx;tyT. The affine transformation
is thus
! ! ! !
þ
x0
y0
¼
ab
cd
x
y
tx
ty
(2)
where ðx0;y0Þ is the pair of coordinates after the
transformation. Figure 10 shows the results of
January/February 2022
Figure 10. Matching center coordinates.
moving and matching the two center points
detected in the image in Figure 9.
The calibration is possible when the circular
object appears in the images from both hemispherical
cameras. Here, it is not necessary to
appear as an entire circular object in images
from both cameras. At least 60% of circular
object's existence in the images from each cameraisenoughtoachieveasuccessful
calibration
process since when at least 60% of the
circular object exists in the image, the center
point also exists within the images. Therefore,
calibration can easily be conducted. If the circular
object's existence in the image is between
50% and 60%, the center point matching is little
difficult since the image area for the matching
process is smaller. If the circular object's existence
in the image is less than 50% the center
point of the circular object does not exist in the
images, therefore the calibration process is not
possible.
EXPERIMENTS
Experimental Environment
We performed several experiments to confirm
the effectiveness and utility of the proposed
method for using two hemispherical cameras to
capture omnidirectional images. The following
describes two typical experiments in detail.
In experiment 1, we placed the cameras 15 cm
apart with a height difference of 20 cm. This separation
distance was selected as a typical width for
city streetlamps and the height difference to demonstrate
the method's robustness to variation.
Figure 11 shows the installation position for the
hemispherical cameras used in this experiment;
note, however, that omnidirectional images can
be captured without problem at any separation
21
IEEE Consumer Electronics Magazine - January/February 2022

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - January/February 2022
