A COMPARISON OF BAYESIAN AND EVIDENCE-BASED FUSION METHODS FOR
AUTOMATED BUILDING DETECTION IN AERIAL DATA
K. Khoshelham 3 ’ *, S. Nedkov 3 , C. Nardinocchi 3, b
d Optical and Laser Remote Sensing, Delft University of Technology, 2629 HS Delft, The Netherlands
- k.khoshelham@tudelft.nl
b DITS, University of Rome “La Sapienza”, 00184 Roma, Italy - carla.nardinocchi@uniromal.it
Commission VI, WG VI/4
KEY WORDS: Fusion, Building detection, Automation, Aerial image, Laser scanning
ABSTRACT:
Automated approaches to building detection are of great importance in a number of different applications including map updating
and monitoring of informal settlements. With the availability of multi-source aerial data in recent years, data fusion approaches to
automated building detection have become more popular. In this paper, two data fusion methods, namely Bayesian and Dempster-
Shafer, are evaluated for the detection of buildings in aerial image and laser range data, and their performances are compared. The
results indicate that the Bayesian maximum likelihood method yields a higher detection rate, while the Dempster-Shafer method
results in a lower false-positive rate. A comparison of the results in pixel level and object level reveals that both methods perform
slightly better in object level.
1. INTRODUCTION
Automated approaches to building detection are of great
importance in a number of applications, including map updating,
city modeling and monitoring of informal settlements. Up-to-
date maps are in high demand with the current widespread use
of navigation systems. Map updating is a tedious task when it is
performed manually by an operator. Each building has to be
inspected in the map and in a recent aerial image (or a stereo
pair). Changes are marked and the map is updated accordingly.
This process is expensive and time consuming. Automation of
this process will save time and cost, making it possible to do
faster and more frequent map updating.
Using data from only one source of data often does not provide
enough information to correctly detect buildings in an
automated fashion. By fusing data from multiple sources the
chances of correctly detecting buildings increase. Several
methods of data fusion have been used for the detection of
buildings from multi-source aerial data (Bartels and Wei,2006;
Lu et al.,2006; Rottensteiner et al.,2004a; Walter,2004). While
relatively successful application of these methods has been
reported, a comparison of the performance of the methods is not
available. The objective of this paper is to provide a comparison
of the two main data fusion methods, namely Bayesian and
Dempster-Shafer, as applied to the detection of buildings from
multi-source aerial data.
The paper has the following outline: in the next section an
overview of the previous research in the field of automated
building detection is given. This includes building detection
using one aerial image, stereo and multiple-overlap images,
using height data and through the fusion of several sources.
Section 3 provides a brief description of the Bayesian decision
theory and the Dempster-Shafer evidence theory. In Section 4
feature extraction and evidence gathering for classification and
morphological post-processing for building detection are
discussed. Experimental results and a comparison of the
performance of the methods are presented in Section 5.
Conclusions are drawn in Section 5.
2. RELATED WORK
Building detection from aerial images has been a hot topic since
the early 1990’s. Early approaches were based on a single
image. Buildings were detected by making use of their shadows
(Lin and Nevada, 1996). Shadows have the disadvantage of not
being visible in all situations. In addition, they can obscure one
another thereby forming complex shapes, which do not
resemble the original buildings. Lin and Nevatia,(1996)
furthermore assume that the roofs of the buildings are
rectilinear, flat, and that the shadows cast by the buildings fall
on flat ground. This is clearly not the case in many urban areas.
Using more than one image supplies more information and
other methods of detection can be applied. Fischer et al.,( 1998)
use multiple images to recognize building features (points, lines
and regions). These features are then used to construct building
comers, wings, and faces, which in turn are combined to
reconstruct a building. Fradkin et al.,(2001) use multiple images
to detect facades of buildings. Several images from different
angles are taken so that the problem of occlusion is reduced to
some extent. After the facade has been found the rest of the
building is detected. Contrary to roof detection (Khoshelham et
al.,2005; Muller and Zaum,2005), facades have less area from
which they can be detected, making their detection troublesome.
Using several images it is possible to construct a height model
of the scene. Weidner and Forstner,(1995) and Brunn and
Weidner,(1997) use the height differences between buildings
and the ground to make a guess about what is a building and
what is not. Using this method, trees can be classified as
Corresponding author.