POINT BASED REGISTRATION OF TERRESTRIAL LASER DATA USING INTENSITY 
AND GEOMETRY FEATURES 
Zhi Wang a,b ' * and Claus Brenner 3 
a Institute of Cartography and Geoinformatics, Leibniz Universität Hannover, 
Appelstraße 9a, D-30167 Hannover, Germany 
-wangzchina@gmail.com, - Claus.Brenner@ikg.uni-hannover.de 
b State Key Laboratory for Information Engineering in Surveying, Mapping and Remote Sensing, 
Wuhan University, Luoyu Road, Wuhan, China, 430079 
Commission V/3 - Terrestrial Laser Scanning 
KEY WORDS: Laser scanning, Registration, Algorithms, Point cloud, TLS, Geometry 
ABSTRACT: 
Terrestrial laser scanning provides a three-dimensional sampled representation of the surfaces of terrestrial objects. The fully automatic 
registration of terrestrial laser scanning point-clouds is still a question as it involves handling huge datasets, irregular point distribution, 
multiple views, and relatively low textured surfaces. In this paper, we propose a key point based method using intensity and geometry 
features for the automatic marker-free registration of terrestrial laser scans. We apply the SIFT method for extracting feature points 
from the reflectance image and geometric constraint for excluding false matches. To evaluate the performance of proposed method, we 
employ a test scene in downtown Hannover, Germany. Reference orientations were acquired by the standard orientation procedure 
using retro-reflective targets and manually assisted target selection. In the experiments, we present the results of the proposed method 
regarding performance, accuracy and running time for the test scene. 
1. INTRODUCTION 
Terrestrial laser scanning provides a three-dimensional sampled 
representation of the surfaces of terrestrial objects, such as 
buildings, sculptures and so on. In most cases, the acquisition of 
several scans is needed to obtain full scene coverage, and 
therefore the data collected from different locations of a scanner 
must be transformed into one global reference frame. The fully 
automatic registration of terrestrial laser scanning point-clouds 
is still a question as it involves handling huge datasets, irregular 
point distribution, multiple views, and relatively low textured 
surfaces. 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), work well (for a comparison, see 
(Rusinkiewicz and Levoy, 2001)). However, in practical 
applications of laser scanners, partially overlapping and 
unorganized point clouds are usually provided without good 
initial alignment. In these cases, the existing registration 
methods are not appropriate since it becomes very difficult to 
find the correspondence of the point clouds. 
2. CURRENT APPROACHES AND OPEN QUESTIONS 
For practical purposes, the identification of tie points between 
scans is nowadays solved using artificial markers which are set 
up in the scene prior to scanning. Using retro-reflective 
materials, they can be detected automatically in the laser scan 
using threshold methods. However, distribution and collection 
of the targets is quite time-consuming and often exceeds the net 
scanning time by a factor of five (Brenner et al., 2008). 
Furthermore, due to practical requirements, an optimal 
distribution often cannot be obtained, resulting in a non-optimal 
distribution of registration errors. In those cases, it would be 
desirable to select tie points evenly distributed in the 
overlapping scan volume. 
To addresses the above problems, there has been extensive 
research for developing automatic registration of terrestrial laser 
scans. Brenner et al. (2008) investigate two different 
registration methods targeted at the determination of suitable 
initial values. The first one is based on planar patches, using 
corresponding planar features to compute the orientation. In 
their autonomous method, following the extraction of planar 
patches from different scans, plane triples are assigned, 
transformations are computed and scored, and the 
transformation with the highest score wins. However, the 
extraction and assignment of the planar patches may become a 
quite complex task in the case of confusional scenes. The 
second one is a non-symbolic approach based on an iterative 
alignment scheme using the normal distributions transform. 
Bamea and Filin (Bamea and Filin, 2008) present a 
computational approach for the registration problem. They 
exploit 3D rigid-body transformation invariant features to 
reduce significantly the computational load involved in the 
matching between key features. Bae and Lichti (Bae and Lichti, 
2004) use variation in geometric curvature and approximate 
normal vector of the surface to determine the possible 
correspondence of point clouds. This requires the computation 
of the normal vector and the curvature itself. These descriptors 
have high potential to be effected by noise because of the 
dependency on second-order derivatives. 
Recently, there have been approaches to extract tie points in 
laser scans without artificial markers using the well-known 
scale invariant feature transform (SIFT, (Lowe, 2003)). This 
Corresponding author: Zhi Wang: wangzchina@gmail.com 533