1365 
OBJECT INFORMATION AUTOMATIC EXTRACTION FROM HIGH RESOLUTION 
STEREO PAIRS BY DENSE IMAGE MATCHING AND INFORMATION FUSION 
Fang Yong, Hu haiyan, Chen Hong, Zhang Wu 
The Xi’an Institute of Surveying and Mapping , No.l YanTa Middle Road, Xi’an, ShaanXi, China, 710054 - 
yong.fang@vip.sina.com.cn 
WgS-PS: WG IV/9 
KEY WORDS: Digital Surface Model, High Resolution, Stereo Image Pairs, Feature Matching, 3D Reconstruction, Image 
Classification, Object Information Extraction 
ABSTRACT: 
With the advent of digital sensor, the requirement for new Photogrammetric software is urgent to quick object information extraction 
from high-resolution stereo pairs. This paper describes an effective combined approach for digital surface reconstruction and 
thematic information extraction automatically. Digital surface reconstruction from stereo pairs is realized by multi-level feature 
matching, which use constraints based on the conditions of the point feature, edge feature, area feature, grey feature and epipolar 
line etc from low to high level. Then the highly reliable and accurate corresponding point pairs is obtained, and edge matching 
controlled with TIN constructed from initial points extends the number of requests to meet. Finally the stereo model is 
reconstruction accurately based on the aerial triangulation by iterative reconstruction technique. The dense image matching is used 
for Digital Surface Model (DSM) extraction, in which the adaptive match window is used to compensate for the blurring effect that 
occurs at object boundaries in high-resolution images. These adaptive techniques separate fore- from background information in a 
correlation window using image classification result for multispectral images. DSM refines the initial classification result, and the 
special layers for buildings and vegetation are generated. 
1. INTRODUCTION 
The rapid development of digital aerial-photo camera provides 
the effective means for the obtainment of remote sensing data, 
which not only improves the data attainment efficiency, but also 
enhances the quality of the attained data. At the same time, 
stereo digital images with large overlap degree can be obtained, 
which makes it possible to provide excellent data for auto 
extracting space information without any additional costs. The 
information extraction from high resolution stereo images is 
one of nowadays hot study fields in application of earth 
observation. 
As high resolution images are concerned, the image characters 
which has obviously distinction to that of middle and low 
resolution ones. As the resolution improved, the rate that areas 
with poor texture or areas with inconspicuous characters 
increases gradually, which will lead the similar measure of 
corresponding window area to fail if the window match strategy 
is adopted and largen the matching results difference to the real 
instance. The traditional theory and method, which mainly 
concern point and linear characters suitable to deal mid and low 
resolution remote sensing images, are coming in for challenge 
with the improvement of resolution. The analyze method based 
on area or surface characters fusing will hold more important 
station. 
The paper brings forward one technology to extract thematic 
objects based on high resolution stereo images depends on 
analyze of the techniques that can be obtained currently. The 
main characters of the technology is the integration of such 
techniques as dense image matching, remote sensing images’ 
auto classifying and space analyzing and it bases on the ground 
surface reconstruct automatically to realize 3D information 
reconstruction and thematic information extraction such as 
buildings, vegetation and water system etc. The algorithm of 
dense image matching synthetically takes into account point, 
linear, spectrum characters and the restrict conditions between 
the corresponding character points, which realizes extracting 
the digital surface model automatic efficiently and provides 
basic 3D space information for further thematic information 
extraction. 
2. BASIC THEORIES AND METHODS 
A confederative method is presented to realize efficiently 
extracting digital surface model and ground thematic 
information automatically according to the characteristic of the 
current available high resolution stereo remote sensing images. 
The basic theories are described as following. Firstly, stereo 
reconstruction is realized through multi-level character 
matching, which using points, linear and area characters and 
epipolar condition restrict to get precise and reliable 
corresponding point pairs. At the same time, the corresponding 
point pairs are expanded to satisfy the requirement of high 
precise stereo directional by means of the edge matching 
controlled by irregular triangle net composed of initialize points. 
The stereo model then is reconstructed precisely through 
iterative stereo. Then, dense image matching is used to extract 
digital surface model. In this step self adaptive matching 
window is used to resolve the punch-drunk in edges during high 
resolution image matching. The selection of self adaptive 
window makes use of the results of high spectrum classify. The 
image in the window is divided to background and topic 
information so as to remove the affection of the discontinuous 
ground surface. The final digital surface model is refined by 
spectrum classify and thematic level as buildings and vegetation 
are generated. The basic theories reference to Figure 1.