XIX-B4, 2012
ingles and positions
cach matching point
3 and 14 show the
Figures 15 and 16
nge. The horizontal
he vertical axes in
inces from the true
he vertical axes in
rom the true value
—J
-90.60
i0 M 12
ber
3-DOF (wide range)
10 H 12
nber
the 3-DOF (wide
10 11 12
Iber
1e 3-DOF (narrow
10 H 12
nber
the 3-DOF (narrow
International Archives of the Photogrammetry, Remote Sensin
g and Spatial Information Sciences, Volume XXXIX-B4, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
4. DISCUSSION
From the results of our experiments, we have confirmed that
our approach can detect locations using a camera and a point
cloud via a fully automated procedure. There are three kinds of
parameter estimation results to be discussed. First, the azimuth
angle estimation for given position parameters was achieved
reliably to within 1.0°, as shown in Figure 10. We have
therefore demonstrated that our approach can be used in an
indoor space environment containing iron frames if accurate
positional data exist.
Second, the X and Y camera position estimations for given
azimuth angles achieved nearly 50 cm accuracy for the wide
spatial range, as shown in Figure 11. The narrow spatial range
result also achieved almost 30 cm accuracy, as shown in Figure
12. From these results, we suggest that our approach will assist
stand-alone positioning using a GPS receiver and existing
indoor positioning techniques to achieve higher positional
accuracy when accurate azimuth data exist.
Finally, both the camera positions and azimuth angles (3-DOF)
were estimated together. These results were less stable than the
independent results because of the increase in estimated
parameters. However, we have also confirmed that our
approach can assist existing indoor positioning techniques to
achieve higher positioning accuracy. For example, if we have
indoor positioning services such as RFID tags and wireless
LAN at 10 m spatial resolution, our proposed approach can
improve the positional data to sub-meter accuracy. In addition,
the positional data are attached to degrec-ordered azimuth
angles.
When we analyze our results, Figure 14 shows that the results
for image numbers 9 and 11 gave large matching errors. Figure
15 also shows that image number 2 gave large matching errors.
We assume that color differences between the camera images
and the rendered panoramic images caused the matching errors,
because the window objects in the gymnasium were regions for
Which the laser scanner failed to measure 3-D points. When
laser-scanning failures exist, the failure points are projected as
missing points from the camera into the panoramic image.
Therefore, a new pixel value (color) is estimated at each
missing point in the panoramic image using neighboring pixel
values in this experiment. The result of the color estimation will
then differ from the pixel value in the camera image. Specular
reflection on the floor also caused matching errors for the same
reason.
Although we detected matching points from 73,800 candidates,
other data could be used in the location detection. A reduced
number of candidates for matching, achieved by using initial
values taken from the various sensors in a mobile device, would
be an effective approach to achieving more stable matching. For
example, gyro sensor data could be used as initial values for
azimuth angle estimation.
Although the spatial resolution of panoramic images was 0.20°,
We could process at approximately 0.01° resolution using
massive point clouds before data reduction in the current state.
In addition, we could apply sub-pixel image processing to
achieve higher spatial resolutions for positions and azimuth
angles,
Currently, there are many challenges to making our approach
useful in practice. Processing-time reduction is one technical
Sue. Our proposed approach has achieved 3-D location
Matching from a 3-D data-processing problem to simple 2-D
Mage processing. This means that graphics-processor-based
Computing might be an effective and low-cost solution for our
Procedure. We can identify three additional challenges as
follows. The first challenge is location detection using a
handheld camera that includes roll, pitch, and yaw angle
estimation. The second challenge is robust estimation in a
changing environment. The third challenge is robust estimation
when occlusion caused by moving objects such as pedestrians
occur.
S. CONCLUSIONS
First, we have focused on the fact that the camera installed in
mobile devices has the potential to act as a location Sensor,
assisting other location sensors to improve positional accuracy.
We have also observed that massive point-cloud data can be
used as a reliable map. Our proposed location-matching
methodology is based on image matching using images from a
digital camera and panoramic images generated from a massive
point cloud in an image-based GIS. When facility information
for construction and maintenance is geocoded onto maps,
higher accuracy and higher spatial resolutions are required.
In this paper, therefore, we have described fine location
matching aiming for 10 cm accuracy to assist indoor positioning
techniques such as RFID and wireless LAN. We have then
developed a matching system to confirm that our location
application can provide location information using a camera
and a point cloud via a fully automated procedure. Although the
current success rate for location detection was below 100%, we
have confirmed that our approach can detect a location using a
digital camera horizontally. We are currently improving the
reliability of our location-matching procedure.
References
[1] Petri, K., Petri, M., Teemu, R., Henry, T, Kimmo,
V.Hannes, W. 2004. Topics in probabilistic location estimation
in wireless networks. In: Personal, Indoor and Mobile Radio
Communications, 2004. PIMRC 2004. 15th IEEE International
Symposium, Vol.2, pp. 1052-1056.
[2] Durrant-Whyte, H., Bailey, T. 2006. Simultaneous
localization and mapping (SLAM): part I, Robotics &
Automation Magazine, IEEE, Vol. 13, Issue: 2, pp. 99-110.
[3] Bailey, T., Durrant-Whyte, H. 2006. Simultaneous
localization and mapping (SLAM): part IL Robotics &
Automation Magazine, IEEE, Vol. 13, Issue: 3, pp. 108-117.
[4] Lars, L., Karsten, M. & Paul R. 2007. Splat-based Ray
Tracing of Point Clouds. Journal of WSCG, Vol.15, Issue: 1-3:
51-58.
[5] Jason, D. 2000. An approach to virtual environments for
visualization using linked geo-referenced panoramic imagery.
Computers, Environment and Urban Systems 24, 127-152.
[6] Edward, V., Sisi, Z. & Sander, D. 2005, Distance-value-
added panoramic images as the base data model for 3D-GIS.
Panoramic Photogrammetry Workshop.
[7] Masafumi, N., 2011. LiDAR VR generation with point-based
rendering. |n: The Urban Data Management Society
Symposium 2011, Delft, Netherlands, Vol.XXXVIII, pp. 223-
230.
[8] Masafumi, N., Satoshi, K., Anna, N., Yuta, S., 2011, 4
point-based rendering application with a high-speed spatial
interpolation. In: The 32nd Asian Conference on Remote
Sensing, Taipei, Taiwan, P 161 8-26-19.
Acknowledgement
This work is supported by Strategic Information and
Communications R&D Promotion Programme (SCOPE) of the
ministry of internal affairs and communications, Japan.