473
AUTOMATIC BUILDING DETECTION FROM LIDAR POINT CLOUD DATA
Nima Ekhtari, M.R. Sahebi, M.J. Valadan Zoej, A. Mohammadzadeh
Faculty of Geodesy & Geomatics Engineering, K. N. Toosi University of Technology, P O Box 15875-4416, Tehran,
Iran - nima_el983@yahoo.com, sahebi@kntu.ac.ir, valadanzouj@kntu.ac.ir, ali_mohammadzadeh2002@yahoo.com
Commission, WG IV/3
KEY WORDS: Building Detection, LIDAR, DSM, DTM, Normalized DSM
ABSTRACT:
This paper proposes an automatic system which detects buildings in urban and rural areas by the use of first pulse return and last
pulse return LIDAR data. Initially both first and last pulse return points are interpolated to raster images. This results to two Digital
Surface Models (i.e. DSM) and a differential DSM (i.e. DDSM) is computed by them. Then using a height criterion, rough and
smooth regions of the DDSM are found. Then last pulse points lying inside smooth regions are filtered using a simplified Sohn
filtering method to find the so called on-terrain points by which the Digital Terrain Model (i.e. DTM) is generated. The normalized
DSM (i.e. nDSM) is calculated using first pulse-derived DSM and the calculated DTM. Afterwards two separated classifications are
applied on the nDSM. The final results of classifications are a set of nDSM pixels belonging to building roofs. The accuracy of the
proposed algorithm is evaluated using some metrics and has proved an overall accuracy of 95.1% and a correctness equal to 98.3%
and a completeness factor equal to 89.5% which show the level of the efficiency and accuracy of the system.
1. INTRODUCTION
Nowadays there is an increasing demand for 3D urban models
produced from Earth Observation data. Such a model contains
all buildings of a city superposed on an accurate DTM. 3D
urban models are being widely used in the development of 3D
GIS databases which has many applications in utility services,
traffic management, air pollution control, etc.
A 3D City Modeling procedure consists of three phases that is
building detection, building extraction, and building
reconstruction. The purpose of first phase is to detect (the pixels
belonging to) some regions representing buildings. The subject
of second phase is to compute the geometry of polygons which
best fit the detected pixels. The third phase aims to compute and
fit the best planar roof type for buildings.
Among all available methods to extract building models of a
city, those who use integrated data sources appear more
successful since the weakness points of either data sources can
be compensated by the other one. Many researches have been
done on the combination of high-resolution imagery and
LIDAR data to detect and extract buildings (Sohn and Dowman
2007; Schenk and Csatho 2002; Rottensteiner et al. 2005; Guo
and Yasuoka 2002).
Considering the capability of dense LIDAR data, there is no
necessity to involving any aerial images in the building
detection task. Even single-source data systems can work faster
and more automatically. Many researches have shown the
capability of LIDAR data in detection and extraction of the
buildings (Vosselman 1999; Maas and Vosselman 1999; Zhang
et. al 2006).
Two ways are often utilized to identify building measurements
from LIDAR data. One is to separate the ground, buildings,
trees, and other measurements from LIDAR data simultaneously.
The more popular way is to separate the ground from non
ground LIDAR measurements first and then identify the
building points from non-ground measurements [Zhang et. al
2006]. The proposed algorithm here is also among the latter
way.
Although there have been some automatic and semi-automatic
methods of building modeling proposed by many researchers,
all the three phases still can be studied and developed more. The
fact that the accuracy of building detection phase has a
dominant direct effect on the buildings extraction task, suggests
that there is a need for more accurately detected building pixels.
Because the geometry of buildings are extracted wherever
building pixels are detected. This paper develops a building
detection system.
Our building detection algorithm (system) is therefore a single
source data system, since it only utilizes LIDAR data to detect
buildings. Our system generates some Digital Elevation Models
by the interpolation of LIDAR points. The result of our system
is presented in a raster format. In other words our algorithm
uses vector data (3D coordinated points) as inputs and gives the
information in raster format (Building pixels with their
elevation).
2. STUDY AREA
The LIDAR point cloud data used to evaluate our system
comprises of two recorded laser pulse returns; First pulse and
Last pulse return points. FP (i.e. First Pulse return) points are
those recorded from the first reflection of the laser pulse. As a
result they might belong to the edges or surfaces of objects on
the terrain rather than the ground beneath them. While the LP
(i.e. Last Pulse return) points are more likely to belong to the
terrain surface, especially for points of vegetation-covered
regions and those near walls of buildings. That’s why we prefer
LP points to create DTM and FP points to create DSM.
The dataset used to evaluate the accuracy of our algorithm
contains both the FP and LP data. The points of either have a