XIX-B4, 2012
rocessing uses only
3. This processing is
without reducing
'othing effect.
averaging filter is
cked to see whether
valid pixels in the
-extraction steps, a
ct valid points. The
distance from the
the extracted points,
s a defined distance
hin the search range
rameter depends on
cloud. Finally, an
to the valid points is
overwritten by the
has a distance value
ed approximately as
ng processing is not
quence is applied to
pint-cloud image
coordinate values in
oint cloud. Azimuth
« can be calculated
mera coordinates to
camera image after
he spatial resolution
r data. Therefore, a
anorama coordinates
cessing time. Image
rted to azimuth and
‘Rectified camera image plane
X
>
Principal point
hae Spherical surface
a spherical surface
tes can be expressed
sing || f
0 of (1)
cos 9 0
es
face coordinates arc
ength of the camera.
to image coordinates
otation angles and à
ration parameters.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B4, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
3. EXPERIMENT
We conducted experiments on location matching using a digital
camera and a point cloud in a gymnasium, which acted as a
large indoor test field. First, we acquired digital camera images
as input data for location detection and a point cloud as a
reference map. Next, we applied image matching to estimate the
three-degrees-of-freedom (3-DOF) parameters. Three kinds of
parameter estimation were conducted in our experiments.
Because horizontal position and azimuth angle are the essential
parameters in general navigation applications, we therefore
focused on estimating the camera position (X, Y) and the
camera azimuth angle as the 3-DOF parameters.
3.1 Data acquisition
We used a digital camera to supply the input data for location
detection and a point cloud taken from a terrestrial laser scanner.
Digital camera mounted on total station
We acquired the camera images (3648 X 2736 pixels) using a
SONY DSC-HXSV. We installed the camera on a total station
(SOKKIA SET550 RXS) to capture images in 12 directions at
intervals 30? in the horizontal plane, as shown in Figure 6.
Digital camera
(SONY DSC-HX5V)
Total station
(SOKKIA SET550 RXS)
Figure 6. Digital camera mounted on total station
The 12 images were projected onto the spherical coordinate
space with a 0.20° angle resolution after camera calibration,
giving 12 images of 360 x 300 pixels each, as shown in Figure
Image 04 $
Image 06
Imagelo Image 11
Figure 7. Projected camera Images
Image 12
Terrestrial laser scanner
We Prepared a point cloud taken from a terrestrial laser scanner
(RIEGL VZ-400). This scanner can acquire panoramic distance
data and corresponding color data over 360°horizontally. The
Scanner was placed at two points. The distance between the two
points wag approximately 15 m. The camera mounted on the
lotal station was placed on the line between the two points, as
shown in Figure 8. The input laser-scanner data set comprised
7,000,000 points in total after 3 cm? spatial filtering.
Figure 8. Sensor arrangement
The color points measured by the laser scanner were then
rendered onto a spherical space with an arbitrary viewpoint.
Figure 9 shows a rendered panoramic image (1800 x 450
pixels) with 0.20? angle resolution. The average processing
time for the panoramic image conversion and iterative filtering
(three iterations) using parallel C programming was 0.33 s for
one-shot generation of multiple viewpoints without file I/O. The
processing involved an Intel Core i7 2.80 GHz processor using
eight-thread execution.
‘ Figure 9. Rendered panoramic image using a point cloud
3.2 Image matching
We estimated the azimuth angle and the horizontal position (X,
Y) of the camera via template matching in this experiment.
We applied a template-matching approach based on the sum of
absolute difference to achieve simple processing of the camera
image and panoramic images in our image matching. Camera
images were used directly as template images and panoramic
images from the point cloud were used as reference images.
The camera was set perpendicular to the line from the laser
reflector on the floor. The relative height of the camera position
from the reflector was measured with a tape measure. The 3-D
position of the reflector was measured with the laser scanner. In
this way, the height value of the camera position was acquired.
The camera was set horizontally using the tripod in the total
station. The horizontal search line in the panoramic image was
therefore given by an elevation angle (0°). The search interval
in the panoramic image was one pixel. The search range in the
panoramic image was therefore from 0.20° to 360.00° at
intervals of 0.20? horizontally and the number of search points
was 1,800 points per panoramic image.
These estimations were conducted over a wide spatial range (10
m at 50 cm spatial resolution) and a narrow spatial range (1 m
at 5 cm spatial resolution). The number of arbitrary viewpoints
was therefore 21 x 21 = 441 points.
As a result, we generated 12 template images from the camera
images and 441 panoramic images from arbitrary viewpoints
for our template matching. The experiment therefore provided
12 estimation results taken from each matching point and
detected from 1800 x 441 = 73,800 candidates.
The average processing time for the matching-point detection in
the template matching via single-thread C programming was
363.90 s using Intel Core i7 2.80 GHz processor.
