Photogrammetric data processing and generation of 3D
point clouds
Photogrammetric data processing is needed to generate a georeferenced
3D point cloud from the unordered, overlapping, and airborne
image collection of the surface. Existing Structure from Motion
(SfM) algorithms automatically extracted features in the images, for
example, contour lines, edges, and feature points. Homologous areas, interior
and exterior orientations were computed in a bundle adjustment.
The Exchangeable Image Format (Exif) metadata from each digital
image further provided approximate values for the focal length and
image size.
A detailed review on how to generate 3D point clouds from photo
imagery can be found in related literature by Lowe [25] and Snavely
et al. [26]. They applied SIFT (Scale Invariant Feature Transform) for
key-point detection. Furukawa et al. [27] proposed a multi-stereoview
approach for large unorganized datasets. Furukawa and Ponce
[28] presented an algorithm for generating referenced 3D point clouds
that are based on the computation of rectangular patches in overlapping
areas of adjacent images.
A commercially-existing software solution by AgiSoft PhotoScan
[29], called PhotoScan, establishes the relation between the unordered
image data collection. This software has been recently optimized for
the use with UAV. The software's professional edition allows to georeference
the results in a specific coordinate system. It exports these
to a digital elevation model or orthophoto. The recommended computer
processing requirements for such large data gathering projects with
more than 100 images are a 64-bit operating system with at least 16
GB RAM.
The data processing is not complicated. First, all aerial images were
imported from the camera to the computer. It was necessary that the
images have adequate overlap between each other. From the Exif metadata,
their approximated interior orientation (focal length and image
sensor size) was determined. Thereafter, a step-wise processing of
the image collection began: (1) align photos, (2) build geometry, and
(3) build texture (if required). Each step gave several possibilities to adjust
parameters, which have influence on the accuracy and structure of
the results and the processing time.
It is generally possible to export the results as a colored point cloud, a
digital elevation model with matched texture, or an orthophoto. An automatically
generated report assesses the quality and accuracy to each
step in the data processing process. It is essential to geo-reference the
data for further use in surveying applications. This task can be completed
in two different ways within the software PhotoScan: direct and indirect
geo-referencing.
Direct geo-referencing can be achieved by using time-stamped GPS
data which is recorded during the flight. Synchronization of the internal
camera time and GPS time is achieved automatically. The exposure
position of the image will be integrated in the Exif data as geographical
coordinates in the WGS84 format. PhotoScan integrates all data and
makes an adjustment to the exterior orientation of the images. As a result,
a point cloud is transferred to the given coordinate system.
Indirect geo-referencing can be applied by measuring reference targets
that were deployed on the ground in the area of interest before the
flight. These targets must be clearly visible in the images. Should targets
not be available, it is also possible to use existing features in the environment
which are fixed, for example, manhole covers or road markings.
These reference points must be surveyed with a suitable survey method,
for example, differential GPS or tachymetry. During the data processing
process it is also necessary to manually identify the reference points in
the model provided by the software. The measured coordinates of the
targets will be referenced to the model. Accordingly, the complete
model will be geo-referenced using a spatial transformation. At least
three reference points are needed for this process, but it is recommended
to use significantly more.
Direct geo-referencing offers the advantages that no ground-based
surveying and neither manipulation of the 3D point cloud are necessary.
It is faster than indirect geo-referencing. However, indirect georeferencing
can take full advantage of using the GPS data that was re-
corded during the flight and relate it easily to the GPS data which was
collected in the fi
eld. As such, the user has to decide if the fina
l
product
of a referenced point cloud should be achieved faster (in such case, preference
should be given to a direct geo-referencing method) or if higher
accuracy is required (then, an indirect geo-referencing method should
be used). Existing commercially-available photogrammetric data processing
software can perform the referencing task of 100 images taken
during a UAV flight within about three hours or less.