Beijing 2008 
A '-.nus, 1992. 
, IEEE 
30(5):pp. 
BUILDING DETECTION AND RECOGNITION 
FROM HIGH RESOLUTION REMOTELY SENSED IMAGERY 
ds Automatic 
tal Elevation 
and Remote 
approximate 
IE TPAMI, 
S. Y. Cui a *,Q. Yan a , Z. J. Liu a , M. Li b 
a Key Laboratory of Mapping from Space of State Bureau of Surveying and Mapping, Chinese Academy of 
Surveying and Mapping, Beijing 100039, China - gisyong@126.com; 
b Geoinformation Science and Engineering College, Shandong University of Science and Technology, 
Qingdao 266510, China - limin82128@163.com 
Commission III, WG III/4 
KEY WORDS: Building Extraction, Active Contour Model, Hough Transform, Convex Hull, Matrix Search 
ABSTRACT: 
This paper supposes a schema to deal with the tough task of building detection and recognition from high resolution remotely sensed 
imagery. It is a region-based and semi-automatic schema combining with Hough transform and computation of convex hull of the 
pixels contained in the building areas, which can produce a precise result when the contrast between flat building rooftop and the 
background is high enough. The first step of this strategy is applying seed region grow algorithm to collect pixels contained in the 
building region to form the approximation shape of building. In order to retrieve the precise shape of building, we devise two 
approaches, which are based on Hough transform and convex hull computation, to deal with different scenes. Based on the fact that 
most buildings in real world can be represented by a convex polygon, the first schema uses this idea to compute the shape of the 
building. The second schema search the desired shape represented by a related orthogonal comer from the node matrix constructed 
by the dominate line sets of the building. Extraction result shows this schema supposed is robust and applicable to most high 
resolution remotely sensed imagery. 
1. INTRODUCTION AND BACKGROUND 
1.1 Introduction 
With the successful launch of some high resolution satellites 
including IKONOS and Quick Bird in recent years, large mount 
of high resolution remotely sensed imagery can be utilized to 
extract man-made objects to update for geographic information 
system database. And man-made object detection and 
recognition from remotely sensed imagery is also of significant 
practical importance for mapping, cartography, photo 
interpretation, military activities and so on. Traditionally, 
manual plotting is deployed in man-made object extraction, but 
it is time consuming and expensive, so automatic or semi 
automatic acquisition and update of building data is greatly 
needed, especially after the availability of high resolution 
satellite imagery such as IKONOS and QuickBird. In the last 
three decades, a significant amount of work that has been done 
in the field of aerial image understanding has concentrated on 
development of efficient algorithm to automatic or semi 
automatic detection(at present, semi-automatic methods are 
applicable in production) and extraction of typical man-made 
objects, such as building. Consequently, various strategies and 
methodologies have been brought forward to deal with the 
tough task of building extraction. In the following section, we 
briefly review the previous research in this field. 
1.2 Previous works 
A collection of state of the art articles can be found in the 
periodical proceedings edited by Grim et al. (1995), Grim et al. 
(1997) and Baltsavias et al. (2001b). Mayer (1999) presented a 
comprehensive survey on the techniques used for image based 
building extraction. Previous research on the building detection 
and extraction is briefly reviewed as follows. Morhan and 
Nevatia (1989) used perceptual organization to detect and 
describe building in aerial images. They recognize the 
usefulness of the structural relationships made explicit by 
perceptual organization in complex image understanding. They 
first detect linear features, which are then grouped into parallels. 
Parallel collation with aligned endpoints triggers the formation 
of a U structure. Two U structures trigger the formation of a 
rectangle hypothesis. A constraint satisfaction network is used 
to select the best consistent rectangles by minimizing the cost of 
the network. This kind of approach is usually comprised of a 
complicated process of bottom-up grouping. Detecting 
buildings in aerial images is also the goal of Heurtas and 
Nevatia (1988). The search for rectangle hypotheses is made by 
local contour tracing techniques. Shadows are used to confirm 
hypotheses and to estimate the height of buildings. Contour 
tracing with some structural guidance as oriented comers and 
depth from shadows has been used in (A. Huertas, R. Mohan 
and R. Nevatia, 1986). These kinds of methods are often 
confronted with the issue of fragmentation of edges. Scott Lee 
and Jie Shan (2003), etc. use the classification result of 
IKONOS multi-spectral images to provide approximate location 
and shape for candidate building hypothesis. Then the fine 
extraction is carried out in the corresponding panchromatic 
image through segmentation and building squaring based on the 
Hough transform. Sohn and Downman (2001) used a local 
Fourier transformation to analyze the dominate orientation in a 
building cluster and extract rectilinear building outline from 
IKONOS imagery based on a binary space partitioning tree. Fua 
and Hanson (1987) segment the scene into regions, find edges 
Corresponding author. This is useful to know for communication with the appropriate person in cases with more than one author. 
411