The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008
174
B) Aerial image , Negative LiDAR Intensity, Positive LiDAR
¥ V«
C) Aerial image , Negative LiDAR Intensity, Positive LiDAR
Figure7: Effect of the texture
3.10 Visualisation of the Rectification Result
After estimating the parameters of the transformation between
the images, one image can be transformed into the coordinate
system of the other one. The standard method of image
rectification is used for this purpose. Figure 8 shows the
rectified LiDAR image, which is superimposed to the aerial
image.
I
■ф
.
'V *
-, >zm
' ?
ч » Rn I
m
SI
Figure 8 :Superimposed LiDAR and aerial images
4. CONCLUSIONS
Registration of aerial images and LiDAR data can be carried
out successfully by using SIFT technique if the LiDAR
intensity images are used for this purpose. In general, a large
number of SIFT keypoints can be extracted and matched
between this two data sets.
Experimental findings with our data are:
1. The number of keypoints extracted from LiDAR intensity
images is 7-8 times higher than the number of extracted
keypoints from LiDAR range data.
2. The grey value distribution (texturing) of a LiDAR intensity
image and an aerial image might be locally very different.
Sometimes the negative (inverse) LiDAR intensities fit much
better to the aerial image. Those differences have to be
expected as the Laser scanner operates in the near infrared
while the aerial is recoding radiation of the visible
electromagnetic spectrum.
3. The following parameters for extracting SIFT features are
found to be most suitable: A sigma level of 1 for the base level
smoothing of LIDAR and aerial images and 5 octaves and 4
sub-levels (number of images per octave) for creating the scale
space representation.
4. The ratio between the distances to the first and second
nearest descriptors has to be lower than a given threshold. A
threshold of 0.8 produced a low outlier quota but at the expense
of a small number of matched keypoints.
5. Regarding the outlier removal with RANSAC and Baarda’s
data snooping no clear preference could be identified.
Limitations that have been observed are:
The invariance of SIFT features to small changes in
illumination does not result in an extraction of suitable features
from shadow areas in an aerial image.
The polynomial transformation of second order which we used
for image registration is only an approximation for a more
general transformation model, e.g. a transformation based on
rational functions.
5. REFERENCES
W. Baarda, "A Testing Procedure for Use in Geodetic Networks,
"Publications on Geodesy, vol. 2, no. 5, pp. 1-97, Netherlands
Geodetic Commission, 1968.
Fischler, M.A. and R.C. Bolles, 1981. Random sample
consensus: A paradigm for model fitting with application to
image analysis and automated cartography. Communications of
the ACM, 24(6):381-395.
Habib, A., Ghanma, M. and M. Tait, 2004. Integration of
LIDAR and Photogrammetry for Close Range Applications,
XXth ISPRS Congress, Istanbul, Turkey.
Hu, J.,You, S. and U. Neumann, 2006. Integrating LiDAR,
Aerial Image and Ground Images for Complete Urban Building
Modeling. 3DPVT, Chapel Hill, USA.
Lowe, D.G., 1999. Object recognition from local scale-invariant
features. International Conference on Computer Vision, Corfu,
Greece, pp. 1150-1157.
Lowe, D.G., 2004. Distinctive image features from scale-
invariant keypoints. International Journal of Computer Vision,
Vol. 60, pp. 91-110, No. 2.
Schenk, T. and Csatho, C., 2007. Fusing Imagery and 3D Point
Clouds for Reconstructing Visible Surfaces of Urban Scenes,