An Application of Panoramic Mosaic in UAV Aerial Image
Jinwen Hu
1
, Yihui Zhou
1
, Chunhui Zhao
1
, Quan Pan
1
, Kun Zhang
2
, Zhao Xu
2
Abstract— Panoramic mosaic plays an important role in the
field of computer vision, robot navigation and virtual reality.
This paper summarizes the specific process of panoramic
stitching and proposes a coordinate transformation stitching
method based on down-sampling. In order to reduce the
processing time, images are compressed by down-sampling
before processing, and spliced in the corresponding points of
original images. Especially, the middle image is chosen as the
reference image and the others are directly spliced onto it by
transformation matrix. Considering the illumination of UAV
aerial images, just one overlapping region of the adjacent
images is remained when doing the fusion. In the experiment,
images with different resolutions are tested. The results show
the performance of good efficiency and little time consuming.
I. INTRODUCTION
In recent years, the demand for panoramic image is
becoming more and more urgent, along with the develop-
ment of computer technology. As an emerging technology,
panoramic mosaic has attracted many researchers attention,
and it has been applied in many areas such as geological ex-
amination, military surveillance, Minimally Invasive Surgery,
video conference and so on.
Panoramic stitching [1] is the process of seamlessly align-
ing and combining multiple images with overlapping fields
of view to form a high-resolution output image. In many
literatures [2]–[4], image mosaic is usually considered as
a multi-image matching problem, and uses invariant local
features to find matches between all of the images, which
accompany high computational complexity. But in most real
applications with a single camera, images are mostly taken
in order and each two adjacent have overlapping areas.
The rest of this paper is organized as follows. Section II
summarizes the process and general problems of panoramic
mosaic. In Section III, panoramic mosaic algorithm is pro-
posed for UAV aerial images. The experiment results are
shown and analyzed in Section IV. Finally, we conclude this
paper and conceive future work in section V.
II. OVERVIEW OF PANORAMIC MOSAIC
A. Process of panoramic mosaic
Nowadays, many image stitching methods have been
proposed. Although they are different in the details of the
algorithm, the framework of the steps is the same. As shown
in Fig. 1, panoramic stitching is basically made up of five
1
Jinwen Hu, Yihui Zhou, Chunhui Zhao and Quan Pan are with
the School of Automation, Northwestern Polytechnical University, Xi’an,
Shaanxi, China.
2
Kun Zhang and Zhao Xu are with the School of Electronics and
Information, Northwestern Polytechnical University, Xi’an, Shaanxi, China.
Corresponding author: Chunhui Zhao (Email: zhaochun-
hui@nwpu.edu.cn)
steps: image preprocessing, image registration, panoramic
image projection, image splicing and image fusion. Image
registration, splicing and fusion are the key steps which
determine the performance of the stitching method.
Fig. 1. Flowchart of panoramic stitching.
Image preprocessing is to eliminate the possible adverse
effects before processing. The main purpose of image prepro-
cessing ensure the accuracy of the next image registration,
which is similar to filtering and enhancing. Filtering algo-
rithms we used commonly include median filter, Gaussian
filter, etc. Image registration includes two parts: feature
extraction and matching.
The first step for stitching is to extract the characteristic
information of two adjacent images to provide the basis for
image matching. There are two main types of detecting and
matching features in images [5], area-based methods and
feature-based methods. Area-based methods use correlation
of intensity patterns of a pixel with the intensity pattern
around the corresponding pixel in another image, which
make them sensitive to changes in viewing position, absolute
intensity, contrast and illumination. Feature-based methods
are based on intensities in the images rather than image inten-
sities themselves. There are two features that are commonly
used, edge detectors and interest point detectors. Interest
points, as SIFT feature [6], Surf feature [7], Orb feature
[8], are robust to changes in lighting, rotation, viewpoint,
translation and scale.
Since each image is captured by camera at different angles,
they are not in the same projection plane. If the overlapping
regions are spliced directly, the visual consistency of the ac-
tual scene will be destroyed. So we need to obtain projection
image, and then splice the image. There are several projec-
tion models such as planar projection, cylindrical projection,
cube projection and spherical projection.
If we use the pattern matching, we can do the splicing
work directly by the translation parameters we calculated
before. If we choose the point features, we need to calculate
the homograph between two images through matching points,
and then convert image coordinates into new.
After all the images are processed, in most cases, however,
it is not a perfect panorama, and some seams at the boarder of
2017 13th IEEE International Conference on Control & Automation (ICCA)
July 3-6, 2017. Ohrid, Macedonia
978-1-5386-2679-5/17/$31.00 ©2017 IEEE 1049