2mains
'ss will
tion is
ith the
istitute
essing
ducing
cao de
XIX
Jues as
27, pp.
-1/SAR
the 35°
GIS by
a Serra
çäo de
Andrew Bibitchev
STRAIGHT EDGE EXTRACTION FROM MULTIPLE VIEWS FOR RECONSTRUCTION OF
MAN-MADE OBJECTS
Andrew BIBITCHEV
State Research Institute of Aviation Systems, Russia
bibitchev@mail.ru
Working Group III/3
KEY WORDS: buildings, edge extraction, image matching, reconstruction.
ABSTRACT
Man-made object extraction and reconstruction based on edge models are widely used. This paper describes technique
of low level extraction of 3D straight edges from multiple views, each of which is gray scale image of 3D scene. The
main idea of the approach consists in simultaneous maximization of the following functionals: integral intensity step
along 2D line on each image and a special form of correlation between 2D lines on different images, where 2D lines are
projections of the 3D edge to appropriate images. During the functionals calculation Gauss pyramids of the images are
used, resulting in high speed and stability. Result of maximization is set of 3D edges, i.e. edges in scene space,
calculated with sub-pixel accuracy. Moreover, values of the functionals can be treated as weights of appropriate edges
in further scene analysis (selection and grouping). Examples of the proposed technique usage for auto and semi-auto
building extraction and reconstruction from aerial imagery are included.
1 INTRODUCTION
Although feature extraction is the first serious step to object recognition and image understanding it still remains one of
the most complicated problem in computer vision [David M. McKeown, Jr., Chris McGlone, Steven Douglas Cochran,
Yuan C. Hsieh, Michel Roux, Jefferey A. Shufelt]. We suppose such situation can be resolved with the aid of maximum
prior and input information usage even during low-level stages of image processing and analysis. One of the reasons is
following. As a rule there are a number of scene images produced either from different view-points or at different times
or in different spectrum parts and we obtain some kind of prior information about dependence between these images.
(This paper concerns the first case, i.e. case of different view presence.) At the first analysis stage (low-level analysis)
most of the approaches deal with each available view of scene separately, which results in additional data miss. In this
paper attempt to use relation between views even during generation of hypothesis is introduced.
The second idea consists in replacing feature extraction task by maximization of functionals, which have integral
nature. Thus, we avoid vague threshold selections and complex aggregations. Moreover, as a rule such functionals have
sense of feature weight and its values can be used in further analysis. To construct the functionals we can formulate the
following characteristics of the features. First of all, feature on image is characterized by special intensity behaviour.
Thus, the first functional must describe this behaviour. Let me call it as integral intensity step functional. Then, for
feature on one image we can try to find appropriate feature on another image. To do this, special form of correlation
between features on different images is introduced.
For simplicity all aspects, described above, will be discussed for straight edge extraction from gray stereo imagery.
Note that generalization is allowed.
This paper is organized as follows. In Section 2 both functionals are constructed and discussed. Also Section 2 concerns
the problem of “simultaneous” functionals maximization. In the third Section we consider topics related to proposed
feature extraction technique; namely, Gauss pyramid usage during functionals calculation and maximization. In Section
4 application of proposed theory for building extraction and reconstruction is described; moreover, auto and semi-auto
approaches are included.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 71