169 
AN INVESTIGATION INTO THE REGISTRATION 
OF LIDAR INTENSITY DATA AND AERIAL IMAGES USING THE SIFT APPROACH 
Abbas Abedini 3 *, Michael Hahn b , Farhad Samadzadegan 3 
a Dept. of Surveying and Geomatics, Faculty of Engineering, University of Tehran, Tehran, Iran - 
(aabedeni, samadz)@ut.ac.ir 
b Dept. of Geomatics, Computer Science and Mathematics, Stuttgart University of Applied Sciences, Stuttgart, Germany 
michael.hahn@hft-stuttgart.de 
Commission I, WG 1/2 - SAR and LiDAR Systems 
KEY WORDS: SIFT, Registration, Matching, LiDAR, MATLAB program, Transformation, Gaussian Function 
ABSTRACT: 
Various methods exist for automatic registration of different kinds of image datasets. The datasets could be aerial or satellite images, 
maps, LiDAR data, etc. In this paper we propose a method for the registration of aerial images and LiDAR data. Our approach 
utilizes the so-called SIFT algorithm by which distinctive features are extracted from both aerial images and LiDAR intensity data. 
The extracted features are then automatically matched and the aerial image and LiDAR intensity data are registered in the 
subsequent stage. The invariance of the SIFT features to image scale and rotation and the robustness of the features with respect to 
changes in illumination, noise and to some extend changes in viewpoint may qualify those features particularly for our purpose. For 
SIFT feature extraction an open MATLAB implementation is used and integrated in the overall image registration process. This 
paper describes our approach and presents the results obtained from a LiDAR test site in Stuttgart, Germany. 
1. INTRODUCTION 
Automatic registration of airborne or remote sensing images 
with other images or maps has a long tradition in 
Photogrammetry and Remote Sensing. Generally, matching 
works very well if the matched image data are of the same type 
but is often less successful in applications using different kind 
of data like range and intensity data. This is the context of the 
research carried out in this paper. Automated registration of 
aerial images and LiDAR data is investigated based on a 
conceptual approach known as Scale Invariant Feature 
Transform (Lowe, 1999). 
1.1 Related Work 
Schenk and Csatho (2007) discussed fusion of imagery and 3D 
point clouds for reconstruction of visible surfaces of urban 
scenes. They represent the reconstructed surface in a 3D 
Cartesian reference system and introduce a feature-level fusion 
framework with the idea to generate a rich 3D surface 
description, in which surface information explicitly includes 
shape and boundary of surface patches and spectral properties. 
The purpose of the surface reconstruction is to support 
applications such as bold-earth determination and the 
generation of true orthophotos. 
Habib et al. (2004) investigated registration of data captured by 
photogrammetric and LiDAR systems for close range 
applications. Their approach relies on straight lines as the 
features of choice upon which the LiDAR and photogrammetric 
dataset are co-registered. They extract LiDAR linear features 
and image points along the conjugate straights lines in the 
images interactively and estimate the parameters of a 3D 
similarity transform. With experiments, they demonstrated the 
capabilities of the proposed method of co-registration of laser 
scans and photogrammetric data. 
Hu et al. (2006) presented a hybrid modelling system that fuses 
data from three resources: a LiDAR point cloud, an aerial and 
ground view images. The purpose of the system is the rapid 
creation of accurate building models. There are three main steps 
to do the modelling: At first building outlines are interactively 
extracted from a high-resolution aerial image and mapped to the 
LiDAR data. Next surface information is extracted from LiDAR 
data and used for model reconstruction. Finally high-resolution 
ground view images are integrated into the model to generate 
fully textured urban building models. 
1.2 Research Concept 
The basic idea for registration of airborne images and LiDAR 
data is to use the LiDAR intensity data and the aerial images 
and employ the SIFT algorithm for matching these two data sets. 
Substantially, the LiDAR intensities represent radiation that is 
reflected from the scanned surface. LiDAR scanners like the 
one we use in our experiments mostly operate with frequencies 
in the NIR, therefore it is likely that a LiDAR intensity image 
and an aerial image do not match in all image regions similarly. 
On the other hand it is expected that the SIFT algorithm with its 
invariance to image scale and rotation and robustness to 
changes in illumination, noise, occlusion and to some extend 
changes in viewpoint should find an adequate number of 
suitable features for a successful matching of both data sets. By 
using the homologous feature points found by SIFT feature 
matching registration can be carried out. 
* Corresponding author