ISPRS Commission III, Vol.34, Part 3A „Photogrammetric Computer Vision“, Graz, 2002 
A NEW METHOD FOR BUILDING EXTRACTION IN URBAN AREAS 
FROM HIGH-RESOLUTION LIDAR DATA 
F. Rottensteiner*, Ch. Briese 
Institute of Photogrammetry and Remote Sensing, Vienna University of Technology, Gußhausstraße 27-29, 
A-1040 Vienna, Austria — (fr,cb)@ipf.tuwien.ac.at 
Commission III, WG III/3 
KEY WORDS: Building extraction, automation, 3D building models, segmentation, laser scanning 
ABSTRACT: 
In this paper, a new method for the automated generation of 3D building models from directly observed point clouds generated by 
LIDAR sensors is presented. By a hierarchic application of robust interpolation using a skew error distribution function, the LIDAR 
points being on the terrain are separated from points on buildings and other object classes, and a digital terrain model (DTM) can be 
computed. Points on buildings have to be separated from other points classified as off-terrain points, which is accomplished by an 
analysis of the height differences of a digital surface model passing through the original LIDAR points and a digital terrain model. 
Thus, a building mask is derived, and polyhedral building models are created in these candidate regions in a bottom-up procedure by 
applying curvature-based segmentation techniques. Intermediate results will be presented for a test site located in the City of Vienna. 
1. INTRODUCTION 
1.1 Motivation and Goals 
Automation in data acquisition for 3D city models is an 
important topic of research with the goal of reducing the costs 
of providing these data at an appropriate level of detail. In 
addition to photogrammetric techniques relying on aerial 
images, the generation of 3D building models from point clouds 
provided by LIDAR sensors is gaining importance. This 
development has been triggered by the progress in sensor 
technology which has rendered possible the acquisition of very 
dense point clouds using airborne laser scanners. Using LIDAR 
data with point densities of up to one point per square meter, it 
is possible not only to detect buildings and their approximate 
outlines, but also to extract planar roof faces and, thus, to create 
models which correctly resemble the roof structures. 
Building extraction is solved in two steps (Brenner, 2000). 
First, buildings have to be detected in the data, and the 
approximate building outlines have to be determined. Second, 
in the regions of interest thus detected, the buildings have to be 
reconstructed geometrically, which results in 3D polyhedral 
models of the buildings. It is the goal of this paper to present a 
new method for the automatic creation of polyhedral building 
models in densely built-up areas from high-resolution LIDAR 
data without using ground plans. Our method is unique with 
respect to the algorithms used for building detection because it 
is based on robust interpolation. In the detected building 
regions, planar roof patches, their bounding polygons, and their 
neighborhood relations are extracted. Grouping of neighboring 
planes has not yet been implemented. The examples presented 
in this paper were computed using the LIDAR data from a test 
site in the City of Vienna captured by TopoSys. The resolution 
of the original point cloud is 0.1 m (in-flight) by 1 m (cross- 
flight). A grid of 0.5 x 0.5 m? derived from that point cloud was 
used for building extraction. The test data were captured in the 
  
* Corresponding author. 
course of a pilot project for the Municipality of Vienna in order 
to evaluate and compare various techniques for the generation 
of 3D city models. Our intermediate results show the high 
potential of the method presented in this paper. 
1.2 Related Work 
There have been several attempts to detect buildings in LIDAR 
data in the past. The task has been solved by classifying the 
LIDAR points according to whether they belong to the terrain, 
to buildings or to other object classes, e.g., vegetation. 
Morphological opening filters or rank filters are commonly used 
to determine a digital terrain model (DTM) which is subtracted 
from the digital surface model (DSM). By applying height 
thresholds to the normalized DSM thus created, an initial 
building mask is obtained (Weidner, 1997; Ameri, 2000). The 
initial classification has to be improved in order to remove 
vegetation areas. In (Brunn and Weidner, 1997), this is 
accomplished by a framework for combining various shape cues 
in a Bayesian network. Our algorithm for building detection 
from LIDAR points is based on the method for DTM generation 
by robust interpolation presented in (Kraus and Pfeifer, 1998). 
The geometrical reconstruction of the buildings in previously 
detected regions of interest has been tackled in two ways. First, 
parametric primitives can be instantiated and fit to the data if 
sufficient evidence is found. Second, planar patches can be 
detected in a DSM created from the LIDAR points, and 
polyhedral building models can be derived by grouping these 
planar patches. As parametric primitives often have a 
rectangular footprint, they are especially used if 2D ground 
plans giving a precise location of the building outlines are 
available. The polygon delineating a building in a 2D map is 
split into rectangular regions. In each rectangle, the parameters 
of parametric models are determined using the DSM, and the 
model achieving the best fit is accepted (Brenner, 2000; 
Vosselman and Dijkman, 2001). The data driven generation of 
A - 295