The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008 
584 
method has originally been developed for image matching and 
has been shown to be robust against changes in illumination, 
scale, rotation and affine distortion. Since most terrestrial laser 
scanners have the ability to capture images as well, the obvious 
way to use SIFT for laser scans is to extract feature points from 
the images first, and then compute their 3D coordinates using 
the known relative orientation of camera and scanner (Bendels 
et al., 2004; Bamea and Filin, 2008). However, Bohm and 
Becker (2007) have even shown that good results can be 
obtained by applying the SIFT operator to the reflectance image 
of the scanner directly. 
Inspired by this, it has been our goal to improve these 
results.We address the major drawbacks of the methods which 
use SIFT features directly. First, SIFT has been deliberately 
built to work across huge scale and viewpoint differences. 
However, using laser scans, the scale and viewpoint are known. 
Second, applying SIFT in the image or reflectance data only 
ignores the geometric information available in the scan data. 
Therefore, geometric important features may be ignored by 
SIFT method if they are not so distinctive in the reflectance 
image. On the other hand, SIFT will be indifferent to high 
responses that are due to the actual features, or to fake ones 
resulting from objects partially covering an object in the 
background and thus creating a feature-like effect. The latter 
may lead to false matches. Thus, it was our goal to improve the 
extraction of interest points by incorporating geometric 
information. 
In this paper we present an automatic registration method, using 
geometric feature-point matching. Geometric curvature (e.g., 
the Gaussian and mean curvature) is invariant to 3D rigid 
motion. Therefore, the change of geometric curvature of the 
surface formed by a point and its neighborhood are used for 
selecting the possible correspondences of point clouds. We add 
the Gaussian, mean curvature values to the SIFT feature 
descriptor vector so that not only the gray values but the surface 
geometric properties take part in the detecting and matching of 
feature-points to optimize the matching process and reduce the 
computational cost involved in the matching between geometric 
features. We also show how the information embedded within 
the range data is utilized to improve the quality of the selected 
geometric feature points, such as discarding the fake features 
(resulting from partially occluded objects) by distinguishing 
layered surfaces with respect to their distances. 
2.1 Registration methods 
Registration of terrestrial laser scanning data is to find the 
rotation and translation parameters which makes corresponding 
locations in the two point clouds SP! and SP 2 coincide. Due to 
the six degrees of freedom to place and orient the acquired 
point cloud in 3D space, any two corresponding points p u 
p 2 £ R 3 with p\ £ SP b p2 £ SP 2 , are related by a rigid 
transformation. 
p ] =Rp 2 +T (1) 
where R is a 3x3 rotation matrix, and T £ R 3 is the translation 
vector. The transformed point of p' 2 (i.e., p \ - +T), and 
its correspondence p x in SP 1; do not exactly coincide because of 
measurement errors. Then, the transformation parameters for R 
and T can be found by minimization of the sum of 
distance ^ |-p'|| between p\ and p' 2 . Therefore, the 
major task is to calculate rotation and translation parameters 
between the two point clouds SPi and SP 2 . 
If a good priori alignment is provided and the point clouds share 
a large overlapping region, existing registration methods, such 
as the Iterative Closest Point (ICP, (Besl and McKay, 1992)) or 
Chen and Medioni’s method (Chen and Medioni, 1991), 
achieve a good performance. However, those methods fail if the 
initial alignment given is too far away from the true relative 
position and orientation. Therefore, methods to obtain a good 
initial alignment are of importance. 
3. THE TEST TERRESTRIAL LASER SCANS 
In this paper, the test scans have been acquired using a Riegl 
LMSZ360I scanner, which has a single shot measurement 
accuracy of 12mm, field of view of 360°x90°and a range of 
about 200 m. At 0.12° step width, a full scan takes 
approximately four minutes and results in a maximum of 
3000x750 = 2.25 million scanned points. We selected an area 
called “Holzmarkt” in the historic district of Hannover, 
Germany, as an example for a densely built-up area (Brenner et 
al., 2008). In order to obtain reference values, manual 
alignment using artificial targets has been carried out, leading to 
errors generally in the range of a few millimeters. Table 1 
shows the relative positions and orientations of the scans for 
those combinations that have been used for the alignment tests. 
One can verify that the scanner has been placed at approximate 
distances of 5m and with arbitrary orientation. Using the 
reference values, we also calculated the overlap between scans 
(as shown in Table 2). 
Pair 
co(°) 
<zH°) 
AT(°) 
X(m) 
Y(m) 
Z(m) 
01-02 
-1.088 
-0.112 
51.731 
-5.50 
0.96 
0.02 
01-03 
0.551 
0.419 
57.447 
-10.69 
1.87 
0.08 
01-03a 
-25.707 
15.540 
62.495 
-10.64 
1.96 
0.05 
01-04 
1.984 
0.481 
119.261 
-16.77 
2.53 
0.14 
01-05 
-0.692 
0.678 
-118.535 
-21.05 
4.24 
0.16 
01-05a 
40.577 
-19.397 
-111.274 
-21.12 
4.11 
0.09 
01-06 
-0.154 
0.276 
29.409 
-24.71 
2.74 
0.29 
02-03 
1.432 
-0.958 
5.733 
-2.50 
4.64 
0.08 
03-04 
0.824 
-1.174 
61.834 
-2.72 
5.47 
0.01 
04-05 
1.482 
2.238 
122.148 
3.58 
2.90 
-0.08 
05-06 
0.096 
0.665 
147.948 
3.07 
-2.51 
0.07 
Table 1: Reference values for the relative orientation of scan 
pairs. (First part: relative orientation of SP t and all other scans. 
Second part: relative orientation of successive scans. The tilted 
scans, which were marked with an “a” suffix, were acquired at 
the same positions as the upright scans.) 
Pair 
Overlap (%) 
Pair 
Overlap (%) 
01-02 
83.1 
02-03 
82.6 
01-03 
77.7 
03-04 
81.3 
01-03a 
73.3 
04-05 
83.6 
01-04 
68.8 
05-06 
80.3 
01-05 
63.0 
01-05a 
59.7 
01-06 
50.5 
Table 2: Overlap percentage for the scan pairs used for the 
alignment tests. (First column: overlap of SP] with all other 
scans. Second column: overlap of successive scans.)