IEEE Consumer Electronics Magazine - January/February 2022 - 23

Figure 16. Images acquired from the two cameras
in experiment 2.
in experiment 2. Supplemental video 2 shows the
video created in this experiment. The first 5 s
show the circle detection and stitching stages,
and the remainder shows video capture during
head rotations and walking. Camera vibrations
due to head rotations and walking of course
affect the video, but the two hemispherical cameras
mounted at some separation were nonetheless
able to capture omnidirectional images.
This article describes only the above two
experiments in detail, but we also performed
multiple other experiments that successfully
demonstrated omnidirectional image capture
with both horizontal and vertical separation
between the hemispherical cameras.
Discussion
We succeeded in generating 360 panoramic
images from two hemispherical cameras independent
of their positioning, which was the aim of
this article. This demonstrates that omnidirectional
images from hemispherical cameras could
be captured without being obscured by the
mounting structure. This allows applications such
as 360 image acquisition in the vicinity of two
head- or robot-mounted hemispherical cameras
with some extent of displacement, or mounting on
existing utility poles for 360
surveillance cameras.
Per-frame processing times in the proposed
stitching method were around 240 ms, so there is
a delay of a few seconds after stitching before the
start of omnidirectional image capture, but this
issue can likely be overcome through use of computers
with higher processing speeds.
Images created after stitching by aligning the
center coordinates of detected circles clearly
show central areas, but small deviations arise as
one moves toward the upper and lower edges.
January/February 2022
Figure 17. Results of stitching the images in
Figure 18.
Therefore, further distortion correction may be
required in cases when it is necessary to capture
extremely accurate omnidirectional images. However,
the currentomnidirectional images are likely
sufficient when the goal is simply object detection.
CONCLUSION
We sought to address the issue of omnidirectional
cameras partially capturing their mounting
structure in acquired images. Specifically, we
mounted two hemispherical cameras at sufficient
separation so that the mounting structure
would not be captured and we then stitched the
resulting hemispherical images together into
an omnidirectional image. In this article, we
included a circular object in the images captured
by the two hemispherical cameras, detected the
circle in those images, and stitched the images
together by matching circle parameters. After
performing stitching, the same stitching was
used to perform continuous omnidirectional
imaging. In future work, we will further develop
the omnidirectional image correction performed
in the proposed method and shorten the processing
times required for stitching. In this article,
we have not done the color calibration after the
stitching process. The color calibration is also a
future work of this project.
&REFERENCES
1. T. Kon and Y. Shibata, " Seamless surveillance system
by omni-directional camera network, " in Proc. 14th Int.
Conf. Netw.-Based Inf. Syst., 2011, pp. 156-161.
2. P. H. Yuan, K. F. Yang, and W. H. Tsai, " Security
monitoring around a video surveillance car with a pair
of two-camera omni-directional imaging devices, " in
Proc. Int. Comput. Symp., 2010, pp. 325-330.
23
IEEE Consumer Electronics Magazine - January/February 2022

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