The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008
173
Aerial
———Last intensity
* - • First intensity
Figure 4: Dependency of matched features on the
number of octaves
3.6 The Impact of the Number of Sub-levels per Octave on
Keypoint Extraction
On each image scale represented by a certain octave a fixed
number of sublevels (differently blurred images) is used in the
SIFT concept to extract the keypoints as explained in Section
2.1. With only two sub-levels a fairly small number of
keypoints is extracted, this increases to a maximum of features
if 4 sub-levels are employed for LiDAR intensity images and 5
sub-levels for the aerial image (Figure 5). The curve for the
aerial images starts from a lower number at sub-level 2 and
intersects the curves for the LiDAR intensities near sub-level 4.
Ie\el=2 level=3 le\el=4 le\el=5 le\«l=6
Number of image per octave
Figure 5: Impact of sub-levels on keypoint extraction
3.7 The Impact of the Matching Threshold on the Number
of Matched Keypoints
To study the dependency of the number of matched features
from matching parameters again image regions of 500 by 500
pixels are used for the experiments. Matching is controlled by
the threshold t (cf. Section 2.2). The lower the threshold t the
more prominent should be the extracted keypoints. With values
of t close to 1 a higher number of error is expected in the
matched keypoints. Figure 6 shows a high sensitivity of the
number of matched keypoints on the threshold t. Around a
threshold of 0.85 a strong growth in the number of matched
keypoints is observed. An increased base level smoothing (the
red curve) produces less matched points.
2000
Threshold of Matching
Figure 6: The number of matched features as a function
of the matching threshold
3.8 Positive and negative LiDAR intensity images
Samples of negative LiDAR intensity, positive LiDAR intensity
and aerial images are depicted in Figures 7A, 7B and 1C. For
the sample depicted in Figure 7A the aerial image is more
similar to the positive LiDAR intensity images, whereas in
Figures 7B and 7C a higher similarity of the aerial image to the
negative LiDAR intensities can be observed. The reason for that
is the Laserscanner operates with a frequency in the Near
Infrared. Whether positive or negative LiDAR intensity images
are used for keypoint extraction has almost no influence on the
number of extracted keypoints. But the extracted keypoints are
different which leads to a higher number of outliers if matched
keypoints related to the positive LiDAR intensity images are
used.
3.9 RANSAC and Data Snooping used for Removing
Outliers
Firstly it should be emphasized that both procedures, RANSAC
and Baarda’s data snooping, are appropriate for removing the
outliers which are among the matched keypoints. If there is a
sufficient number of corresponding keypoints, which match
very well both procedures eliminate the outliers efficiently. If
there is a high percentage of outliers and a weak geometric
distribution of the extracted points RANSAC may stuck to a
local maximum of the search space. But the iterative outlier
elimination also failed by eliminating some good point pairs. A
simulation would probably give a more detailed insight into the
pros and cons of both methods that our experiments with real
image data can do.
A) Aerial image , Negative LiDAR Intensity, Positive LiDAR