September 1-3, 2010 
In: Paparoditis N., Pierrot Deseilligny M.. Mallet C.. Tournaire O. (Eds), IAPRS. Vol. XXXVIII. Part ЗА — Saint-Mandé, France. September 1-3. 2010 
49 
Number of 
false positive 
5 
3 ~ 
t-crossing detection 
is algorithm, which 
uite fast execution 
?ssor). Moreover, it 
preprocessing step, 
Lion. 
ng into account the 
f. by using a road 
r models to specific 
98. Stereo inverse 
. Image and Vision 
>m sample consen 
sus to image anal- 
'M 24(6), pp. 381- 
7 and Nagai. M.. 
ra for driver assis- 
Dongress. Munich. 
Itering in constant 
16(9). pp. 2389- 
!004. Image seg- 
•king analysis. In: 
Electronics, Voi. 1, 
partially sighted, 
ty Conference on 
VPR’00), Voi. 2, 
and Dissanayake, 
ivironments. In: 
nee on Intelligent 
SA, pp. 123-128. 
idant, J., 2006. 3d 
tages of a ground- 
Archives of Pho- 
rmation Sciences 
arity and projec- 
r the visually im- 
• Society Confer- 
tion (CVPR’05), 
>nnier, R, 2008. 
on. In: Pro- 
itelligent Trans- 
ì. pp. 174-181. 
:08.html. 
BUILDING DETECTION FROM MULTISPECTRAL IMAGERY AND LIDAR DATA 
EMPLOYING A THRESHOLD-FREE EVALUATION SYSTEM 
Mohammad Awrangjeb. Mehdi Ravanbakhsh and Clive S. Fraser 
Cooperative Research Centre for Spatial Information. The University of Melbourne 
723 Swanston St. Carlton Vic 3053. Australia 
E-mails: {mawr. m.ravanbakhsh. c.fraser}@unimelb.edu.au 
Commission III-WG I1I/4 
KEY' WORDS: Building detection, LIDAR. point cloud, multi spectral, photogrammetry imagery, orthoimage, fusion, edge feature 
ABSTRACT: 
This paper presents an automatic system for the detection of buildings from LIDAR data and multispectral imagery, which employs 
a threshold-free evaluation system that does not involve any thresholds based on human choice. Two binary masks are obtained from 
the LIDAR data: a 'primary building mask' and a ‘secondary building mask'. Line segments are extracted from around the primary 
building mask, the segments around trees being removed using the normalized difference vegetation index derived from orthorectified 
multispectral images. Initial building positions are obtained based on the remaining line segments. The complete buildings are detected 
from their initial positions using the two masks and multispectral images in the YIQ colour system. The proposed threshold-free 
evaluation system makes one-to-one correspondences using nearest centre distances between detected and reference buildings. A total 
of 15 indices are used to indicate object-based, pixel-based and geometric accuracy of the detected buildings. It is experimentally 
shown that the proposed technique can successfully detect rectilinear buildings, when assessed in terms of these indices including 
completeness, correctness and quality. 
1 INTRODUCTION 
Building detection from remotely sensed data has a wide range 
of applications including change detection, automatic city mod 
eling. homeland security and disaster (flood or bush fire) manage 
ment. Therefore, a large number of building detection techniques 
have been reported over the last few decades. These can be di 
vided into three groups (Lee et al., 2008). The first group of algo 
rithms uses 2D or 3D information from photogrammetric imagery 
(Mayer, 1999). These algorithms are complex due to involve 
ment of detailed information in high-resolution images (Awrang 
jeb et al.. 2010) and complicated and erroneous estimation of 3D 
(height) information (Sun et al., 2005). Algorithms in the sec 
ond group consider building detection as a classification problem 
and detect building regions from LIDAR (Light Detection And 
Ranging) data (Lee et al., 2008). However, the use of raw or in 
terpolated data can influence the detection performance (Demir 
et al.. 2009) resulting in poor horizontal accuracy for building 
edges (Yong and Huayi, 2008). As a result, it is hard to obtain a 
detailed and geometrically precise boundary using only LIDAR 
point clouds (Awrangjeb et al., 2010) 
In fact, the introduction of LIDAR has offered a favourable op 
tion for improving the level of automation in the building detec 
tion process when compared to image-based detection (Vu et al., 
2009). The third category of methods does use both LIDAR data 
and photogrammetric imagery, since each have unique attributes 
for building detection and the advantages of one can compensate 
for disadvantages of the other. More specifically, intensity and 
height information in LIDAR data can be used with texture and 
region boundary information in aerial imagery to improve accu 
racy (Lee et al., 2008). However, the question of how to com 
bine the two different data sources in an optimal way so that their 
weaknesses can be compensated effectively is an active area of 
current research (Yong and Huayi, 2008); only a few approaches 
with technical details have thus far been published (Rottensteiner 
et al., 2005). 
In addition, there is a current lack of uniform and rigorous eval 
uation systems, and an absence of standards (Rutzinger et al.. 
2009). Indeed, evaluation results are often missing from pub 
lished accounts of building detection (Yong and Huayi, 2008), 
and the use of 1 to 2 evaluation indices only has characterized 
many studies (Demir et al., 2009. Vu et al., 2009). The majority 
of these (Rottensteiner et al.. 2005. Rutzinger et al., 2009. Lee et 
al., 2008) use one or more overlapping thresholds while making 
correspondences between detected and reference building sets. 
The problem with threshold-based systems is that they are too 
subjective and likely to be controversial since there is no unique 
way to select the thresholds (Shufelt, 1999). 
This paper aims at a successful integration of LIDAR data and 
photogrammetric imagery for building detection so that the im 
proved detection performance is obtained. It also develops an 
automatic and threshold-free performance evaluation system us 
ing 15 indices from three categories: object-based, pixel-based 
and geometric. The performance of the proposed building detec 
tion approach has been assessed using the proposed evaluation 
system. Note that this paper is a condensed version of (Awrang 
jeb et al.. 2010) with an extended experimental validation. It has 
similarities to that reported by (Sohn and Dowman. 2007) and 
(Cheng et al., 2008) in the sense that it uses line segments and a 
regularization step (adjustment) employing dominant line angles. 
2 RELATED WORK 
Building detection techniques integrating LIDAR data and im 
agery can be divided into two groups. Firstly, there are techniques 
which use the LIDAR data as the primary cue for building detec 
tion and employ the imagery only to remove vegetation (Rotten 
steiner et al., 2005. Vu et al.. 2009). As a result, they can suffer 
from poor horizontal accuracy for the detected buildings. 
Secondly, there are integration techniques (Lee et al., 2008, Demir 
et al.. 2009, Sohn and Dowman, 2007) which use both LIDAR 
data and imagery as the primary cues to delineate building out 
lines. They also employ imagery to remove vegetation. Con 
sequently, they offer better horizontal accuracy for the detected