A Statistical Examination of Image Stitching
Software Packages For Use With
Unmanned Aerial Systems
John W. Gross and Benjamin W. Heumann
Abstract
There is growing demand for the collection of ultra-high spatial
resolution imagery, such as is collected using unmanned aerial
systems (
UAS
). Traditional methods of aerial photogrammetry
are often difficult or time consuming to utilize due to the lack
of sufficiently accurate ancillary information. The goal of this
study was to compare geometric accuracy, visual quality, and
price of three commonly available mosaicking software pack-
ages which offer a highly automated alternative to traditional
methods: Photoscan Pro, Pix4D Pro Mapper, and Microsoft Im-
age Composite Editor (
ICE
). A total of 223 images with a spatial
resolution of 1.26 cm were collected by a
UAS
along with 70
ground control points. Microsoft Image Composite Editor had
significantly fewer visual errors (Chi Square, p < .001), but it
had the poorest geometric accuracy with a
RMSE
of 34.7 cm
(Tukey-Kramer, p < 0.05). Photoscan had the most visual errors
(Chi Square, p < 0.001), and a
RMSE
of 10.9 cm. Pix4D had
the best geometric accuracy with a
RMSE
of 7.7 cm, however
this was not found to be statistically different from Photoscan
(Tukey-Kramer, p > 0.05). In terms of price, Microsoft Image
Composite Editor was the least expensive while Pix4D was the
most expensive, although specific pricing varies depending on
the type of licensing needed. These results suggest that unless
high geometric accuracy or 3D images are required,
ICE
is the
best option for most
UAS
photogrammetric applications.
Introduction
Small unmanned aerial systems (
sUAS
) are a novel comple-
ment to existing image acquisition platforms. Although they
cannot replace satellites and manned aircraft for all applica-
tions,
UAS
offer three key benefits. They: (a) represent a cost
effective option for the acquisition of imagery over study
areas with limited spatial extents, (b) are capable of collecting
imagery in regions which have traditionally been too danger-
ous or delicate for manned aircraft, and (c) fly at drastically
lower altitudes (less than100 m above ground level), which
can translate into novel fine spatial resolutions on the order of
1 cm, which is difficult if not impossible to replicate on any
other platform (Anderson and Gaston, 2013).
sUAS
use has been well documented in numerous applica-
tions including biological research, precision agriculture, and
archeology (Knoth
et al
., 2013; Verhoeven 2011; Verhoeven
2012; Zaman
et al
., 2011; Zhenkun
et al
., 2013). This growth
in
sUAS
use will likely continue as government regulations
and safety practices adapt to meet application demands
(Zweig
et al
., 2015). The use of such miniaturized platforms
and sensors, however, leads to a number of challenges that
must be addressed, especially with regards to combining
large numbers of images (100+) into a meaningful representa-
tion of the Earth’s surface, a process known as creating an
orthomosaic. This issue is typically addressed using aerial
photogrammetric techniques.
One of the more conventional ways to handle the creation
of orthomosaics in aerial photogrammetry is through the
use of automatic aerial triangulation (
AAT
) and bundle block
adjustment (
BBA
). In this method, software is able to utilize
interior orientation (
IO
) information provided by the camera, a
global positioning system (
GPS
), and an inertial measurement
unit (
IMU
) to match individual images together then adjust
those blocks of images to match the real world (for a more
thorough review of
AAT
and
BBA
readers should refer to Wolf
and Dewitt , 2000). The accuracy and quality of these proce-
dures are highly dependent on the ability to provide the soft-
ware with accurate information (Barazzetti
et al
., 2010; Turner
et al
., 2012). This is not a problem when using sophisticated
survey grade metric cameras for which the
IO
s have been
accurately calculated, and highly accurate
GPS
and
IMU
data
which is typically available on manned aircraft platforms.
Due to the small size and limited power capacity of many
commercially available
sUAS
, the weight and size of mounted
sensors, or any payload, is severely restricted. This often trans-
lates into the use of consumer grade digital cameras,
GPS
, and
IMU
, which typically have insufficient accuracy for conven-
tional orthomosaic techniques (Laliberte
et al
., 2007; Laliberte
et al
., 2008; Turner
et al
., 2012). There can also be significant
variability in rotational and angular camera position, degree
of overlap, and illumination between images (Barazzetti
et al
.,
2010). Such errors make the use of conventional photogram-
metric techniques difficult, and time consuming, to implement.
Fortunately, recent advancements in the fields of photo-
grammetry and computer vision have produced novel tech-
niques which offer the potential to not only handle these
issues, but to do so in a highly automated fashion. One of the
most notable techniques is structure from motion (
SfM
).
SfM
is
of benefit because it does not require
a priori
knowledge of any
camera parameters or scene information, which complicates
the traditional methods (Choudhary and Narayanan, 2012;
Westoby
et al
., 2012).
SfM
utilizes some form of scale invariant
feature transform (
SIFT
) which uses a difference-of-Gaussian
function to identify “important” features in each image known
as keypoints (Lowe, 2004). These keypoints are then matched
in multiple images based on a minimization of Euclidian
distance function (Lindeberg, 2012). By tracking the keypoints
from image to image,
SfM
is also able to accurately estimate a
number of external camera parameters such as camera orienta-
tion (Westoby
et al
., 2012). From this combination of infor-
mation the software is then able to project each pixel into an
Center for Geographic Information Science and Geography
Department, Central Michigan University, Mt. Pleasant, MI
48859 (
).
Photogrammetric Engineering & Remote Sensing
Vol. 82, No. 6, June 2016, pp. 419–425.
0099-1112/16/419–425
© 2016 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.82.6.419
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
June 2016
419
