A HEIGHT ANDTEXTURE INFORMATION INTERGRATED APPROCH FOR OBJECT
EXTRACTION APPLIED TO AUTOMATIC AERIAL TRIANGULATION
; a* a pn a
Li Pan^', Hong Zheng”, Zhuxun Zhang? Jiang Zhang’,
“School of Remote Sensing and Information & Engineering, Wuhan University
129 Luoyu Road, Wuhan 430079, P.R.China
"School of Electronic Information, Wuhan University
129 Luoyu Road, Wuhan 430079, P.R.China
Commission PS WG III/4
Key words: Photogrammetry, Segmentation, Algorithms, Integration, Triangulation
Abstract:
This paper proposes a method of objects extraction based high and color texture features. Trees or forests are interested objects in
this paper, which are higher than their surroundings. According to the disparity image processed, at first, the original color aerial
images are segmented into the high and low objects, next, on the basis of preliminary results, the trees are refined by Fuzzy C-Mean
with the texture and color features from the high and low areas. With above results, in this paper, an architecture is presented that
observed points on forest and on non-forest are divided into two different groups in automatic aerial triangulation. This is because the
matching accuracy and reliability of observed points on trees is lower than on non-trees. To decrease the negative influence of the
observed points on trees, the observed points in two different groups are given the different a'priori standard deviation of photo
coordination measurement and weights. The experiment results have proved that the method of observed points grouped can improve
the accuracy of automatic aerial triangulation in forest-covered regions..
1. Introduction
The analysis of texture has proven to be an important tool
for image segmentation. A lots of approaches were presented
by the researchers who were in different fields. Those methods
are divided into the three major branches: thresholding or
clustering, edge detection, and region extraction. Traditionally,
the analysis of texture models is on the images of two
dimension, while the height feature belonged to the information
of third dimension has received much less attention. Among the
approaches that have been examined, T. Dang (1994) proposed
the algorithm that was to directly use height information for
detection and reconstruction of buildings in aerial images. The
other method (N.Haala, 1994 ) that was via stereo
reconstruction process to obtain height information for
detection of buildings. The above two algorithms assume only
the case that there are simple houses and little trees or no trees
on flat ground and the quality of the images is very good. In
this paper, trees or forests are interested objects. In contrast to
artificial objects, such as houses and roads, trees or forests are a
kind of natural scenes which are not structured and cannot be
represented easily by regular rules. In addition, trees or forests
do not obey strict position rules. Hence, texture and color
features are important cues for trees or forests extraction in
color aerial images. In practice, if only texture and color
features are used, sometimes the results of trees or forests
extraction are inaccurate. There are two reasons: (1) The non-
tree objects have the similar texture features to trees; (2)
Usually different tree type has different texture features. But, it
is the fact that trees are the objects that are higher than their
surroundings. In this paper, an approach is descried for the
recognition of various forest and trees in low resolution aerial
images. Our method combines texture features, color and
height information to overcome those disadvantages. À
technique similar to ours was discussed in (W.Eckstein, 1996).
*Corresponding author. Email: li.pan@126.com
Their method has the advantage that can dealt with realistic and
complex scenes. But it needed high resolution aerial images
and DTM, and assumed that the range of the highest building
was from 20m to 30m. Hence, it was able to extract the huge
buildings and forests. Our method is to first use the parallaxes
data to provide the preliminary segmentation of the image and,
given this results, perform Fuzzy C-Mean to refine and produce
the final segmentation. At first, the stereo models are created
by the digital photogrammetry system (Virtuozo), the
parallaxes data are obtained. Second, the image that parallaxes
are represented by 256 grey levels is divided into different
regions and the edges are detected by Sobel algorithm. Third,
according to the parallaxes image processed, the original color
aerial images are segmented into the high and low objects.
Generally , in the areas covered by forest, high objects mainly
include trees or forest, less houses and bridges and more non-
trees objects are took in the low region. In order to refine trees
or forest, the high and low objects are classified by Fuzzy C-
Mean respectively. With above results, in this paper, an
architecture is presented that observed points on forest and on
non-forest are divided into two different groups in automatic
aerial triangulation, because the matching accuracy of observed
points on trees is lower than the observed points on non-trees.
To decrease the negative influence of the observed points on
forest, the observed points in two different groups are given the
different pre-variances and weights. The experiment results
have proved that the method of observed points divided into
two different groups can improve the accuracy of automatic
aerial triangulation.
This paper is organized as follows: Section 2 describes the
method of the objects extraction based on the high and color
texture features and is divided into two subsection. In section 3,
experimental results are discussed and the conclusion is given
in Section 4.
Internat
2. TI
In th
includir
the text
mean.
2.1 Hig
We s
digital |
the ima,
original
low obj
Algorit
Given:
Step 1.
Step 2.
Step 3.
Step 4.
Step 5.
