The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008 
172 
d e f 
Figure 2. (a,d) Aerial image (20 cm), (b,e) reduced resolution 
(40 cm) ,(c, f) further reduced resolution (80 cm) 
a Rollei AlC-modular -L-S digital metric camera with 4080 x 
5440 pixels per image. 
The RGB images are transformed to a gray scale images and 
used as input to the SIFT algorithm. The LiDAR data have been 
interpolated to a regular raster of 0.5 m grid size. Thus four 
images, two for the last and two for the first returns of LiDAR 
scanner are produced which in the following are called “Last 
Range, First Range, Last Intensity and First Intensity". Each 
LiDAR image covers 2000 x 2000 grid points. 
3.2 SIFT Algorithm applied to the LiDAR Data 
In a first experiment, keypoint extraction related to LiDAR 
range and intensity is investigated by using all four LiDAR 
images. Table 3.1 shows a significant difference with respect to 
the number of extracted keypoints. LiDAR intensity images 
have 7 to 8 times more SIFT key point features than LiDAR 
range images which manifests a comparatively low range 
variation in the range images. The reflected signal 
Original image 
500*500 pixels 
original 
aerial image 
reduced 
image 
reduced 
image 
(20 cm) 
(40 cm) 
(80 cm) 
Number of 
3980 
1326 
378 
SIFT features 
Table 3.1. Key points in different LiDAR images 
strength, which leads to the intensities in the range images, 
leads to texturing which is comparable to the greyscale aerial 
image. The spatial distribution of the keypoints is therefore 
sparse in the LiDAR range images and dense in the LiDAR 
intensity images. First return data tend to result in more 
keypoints than last return data, which was expected in advance. 
3.3 SIFT Keypoints extracted from the Aerial Image 
While LiDAR images have a resolution of 50 cm, the aerial 
image resolution is much higher with around 20 cm on the 
ground. To see the impact on the number of extracted keypoints 
the resolution of the aerial image is reduced to 40 cm and 80 cm 
ground pixel size by Gaussian filtering and resampling. Figure 
2 shows the extracted keypoints. The reduction of the 
resolution by a factor of 2 (linearly), i.e. 2 by 2 pixels are 
combined to 1 pixel, leads to a similar reduction of the number 
of extracted keypoints (Table 3.2). 
LiDAR 
Last 
Range 
First 
Range 
Last 
Intensity 
First 
Intensity 
Number of 
Key Points 
1601 
1828 
11993 
13393 
Spatial 
distribution 
poor 
poor 
good 
good 
Table 3.2. Dependency of SIFT keypoints from image 
resolution 
3.4 Impact of Base Level Smoothing of the Datasets on 
SIFT Keypoint Extraction 
While in the previous section smoothing and reduction of the 
resolution of aerial image was carried out, the base level 
smoothing is a filtering of the input image with the Gaussian 
kernel to define the base level of the scale space image 
according to Eqn. (2-1). Figure 3 shows the dependency of the 
number of extracted keypoints on base level smoothing for an 
aerial image, and for LiDAR First Range, Last Intensity and 
First Intensity. All image data sets are regions of 500 by 500 
pixels. Figure 3 also shows that the curves for the aerial image 
and both LiDAR intensity images are close to each other at a 
base level smoothing with a cr of 1 and above. With a lower 
sigma value for base level filtering more keypoints are 
extracted from the aerial image than from the LiDAR intensities. 
A value of 1 for the base level a was used for all further 
experiments. The red line indicates the very small number of 
extracted keypoints in the range data. 
16000 
Figure 3. Impact of base level filtering on keypoint 
extraction 
3.5 Impact of Octave Selection on Keypoint Extraction 
Figure 4 shows that with an increasing the number of octaves an 
increasing number keypoints is extracted. However with more 
than 5 octaves the number of extracted keypoints is almost 
constant. Thus, 5 octaves are used for the further experiments. 
The size of the image was again 500 by 500 pixels. The number 
of extracted keypoints varies with less than 10% between aerial 
image and the LiDAR intensity images.