Liang-Chien Chen
EXTRACTION OF MAN-MADE BUILDINGS IN MULTISPECTRAL
STEREOSCOPIC IMAGES
Liang-Chien Chen, Wei-Chen Hsu
National Central University, TAIWAN
Center for Space and Remote Sensing Research
Icchen @csrsr,ncu,edu,tw
Industrial Technology Research Institute, TAIWAN
Energy and Resources Laboratories
f880224.@erl.itri.org.tw
Working Group IV4
KEY WORDS: Building, Feature Extraction, Image Matching, and Multi-Spectral Data.
ABSTRACT
The objective of this investigation is to develop a scheme to extract man-made buildings from color aerial image
pairs. The scheme consists of five stages of processing. In the first stage, an epipolar transformation is applied
according to the predetermined orientation parameters. Second, segmentation for color image pairs is performed
using 3-band region growing approach. In the third stage, we calculate edge strengths for 3-band image pairs.
Fourth, the detection and location of precision corner points are performed. Two steps are included in this stage.
The initial detection for corners is firstly performed according to the segments derived from stage two. Then,
those initial corner points are refined according to the edge strengths as calculated in stage 3. In the final stage, a
new matching idea is proposed to determine the disparities for each building in the stereopair. The proposed
operator improves the matching results with respect to the traditional ones. Experimental results indicate that
proposed scheme achieve (1) high precision for corner point location, (2) highly successful rate in image matching,
and (3) high accuracy of disparity for buildings.
1 INTRODUCTION
Extraction of man-made buildings from aerial stereopair is an important task in topographic mapping and three
dimensional city modeling. Automated extraction of such objects is also the major work in digital
photogrammetry research. A large number of researches have been devoted to this area (Henricsson & Baltsvias,
1997; Gabet, et. al., 1997; Baillared, et. al., 1998; Kim & Muller, 1998; Sahar & Krupnic, 1999). Considering
different shapes, environments, and image intensity for different buildings together with the occlusion and shadow
effects, the automation is essentially an ill-posed problem (Sahar & Krupnic, 1999). In addition to developing
better scheme, increasing higher information content is an essential way for the solution. Henricsson (1998) used
multi-spectral images to improve the performance in roof determination and edge extraction. Spreeuwers, et.
al.,(1997) and Gabet, et. al.,(1997) used multi-view images to reduce the effect of occlusion. Thus, high accuracy
and reliability were achieved. From color image point of view, we will propose a scheme to extract man-made
buildings from aerial stereopairs. The major works of the scheme include (1 )generation of epipolar image pair,
(2)segmentation by region growing, (3)computation of edge strength, (4)precision corner detection, and
(5)precision image matching. A flowchart of the proposed scheme is show in fig.1.
2 THE PROPOSED SCHEME
The scheme consists of five stages of processing. In the first stage, an epipolar transformation is applied to the
stereoscopic pair. Second, segmentation for color image pairs is performed using 3-band region growing. In the
third stage, we calculate edge strengths for 3-band image pairs. Fourth, the detection and location of precision
corner points is performed. Two steps are included in this stage. The initial detection for corners is firstly
performed according to the segments derived from stage two. An algorithm will be provided for the initialization.
Then, those initial corner points are refined according to the edge strengths as calculated in stage 3. In the final
stage, a new matching idea is proposed to determine the disparities for each building in the stereopair. The
proposed matching technique is call “CLR Matcher”. The matcher uses three-window approach to match image
points along building edges.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 169