EVENTS-BASED IMAGE ANALYSIS 
FOR MACHINE VISION AND DIGITAL PHOTOGRAMMETRY 
Yury V. Visilter, Sergei Yu. Zheltov, Alexander A. Stepanov 
(State Research Institute of Aviation Systems, Moscow, Russia) 
KEY WORDS: Image Analysis, Object Recognition, Bayesian theorem. 
ABSTRACT: 
A new approach for model-based image analysis called the Events-based image Analysis (EA) is proposed. From EA point 
of view, any certain procedure of image understanding can be interpreted as a procedure of evidence fusion. Any fact about 
the whole image, about its part or even about one proper pixel can be the evidence, and the any proposition about the scene 
observed is the hypothesis that to be tested based on these evidences. In this paper the EA formalism was outlined in the 
Bayesian terms. This approach allows to compose the power of sample-based methods and the flexibility of model-based 
methods without the direct comparison of objects or images. The most important properties of EA procedures are the 
following: usage of generic models, usage of hierarchical models and easy fusion of non-homogeneous information. Based 
on EA ideas the complex technique for house detection was proposed. It provides the easy fusion of contour and intensity 
information for 3D-model validation. 
1. INTRODUCTION 
Any certain engineering technique can be considered as a 
combination of "method" and "state of art". Here, 
"method" is a regular part of technique that to be the best 
for a common class of problems involved. In the contrary, 
the term "state of art" means the part of technique that 
reflects the peculiarity of the concrete problem and due to 
this - the skill of the developer. It is very attractive to 
define the common method for the most wide spectrum of 
problems, to prove its' optimality and then to concentrate 
on the state of art only. 
Some object models (of different types) are developed for 
detection and measurement of artificial objects, and the 
proper measurement presumes the estimation of numeric 
parameters of these models. So, the detection methods 
based on such models to be preferred. However, though 
some of model-based detectors are developed, usually they 
contain the heuristic matching procedures without any of 
optimal assumptions. Because of this reason, most 
powerful and robust object detection techniques do match 
the sample images but not the models. 
In the earlier works our group intensively used two well- 
known matching techniques: the Pytiev Morphology and 
the Hough Transform. The Pytiev Morphology provides the 
most invariant detection of objects by their samples, but it 
can not work with models. The Hough transform (HT) and 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
the Generalized Hough Transform (GHT) support the 
efficient model-based contour analysis, but can not use any 
other type of information contained in images. The 
common "method" called the "Events-based image 
Analysis" (EA) was developed to compose the Pytiev's 
optimal state of problem and the methodology of Hough 
transform. EA is a "method" for the most generic model- 
based image analysis while its' "state of art" is connected 
with a choice of the adequate models. 
From EA point of view, any certain procedure of image 
understanding can be interpreted as a procedure of 
evidence fusion. Any fact about the whole image, about its 
part or even about one proper pixel can be the evidence, 
and the any proposition about the scene observed is the 
hypothesis that to be proved or escaped based on these 
evidences. There are many possible ways to provide the 
such fusion. However, the Bayesian approach is the most 
popular and meaningful. So, we shall outline the EA 
formalism in the Bayesian terms. 
In section 2 we outline the basic idea of Bayesian EA. In 
section 3 the EA-approach for house extraction is 
described as an example of EA-application for 
photogrammetry and machine vision. 
898 
  
     
   
  
   
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
   
  
  
  
  
  
  
  
  
  
   
  
  
  
  
    
   
  
   
  
   
  
   
   
  
  
  
  
   
   
From 
object 
princi 
and m 
This « 
based 
based 
chara: 
chara 
distan 
criteri 
recog 
the ol 
statist 
distril 
classi 
traini 
comp 
The 1 
meth 
same 
mode 
the m 
In o 
detect 
(Step 
invar 
demo 
comp 
main 
imag: 
So, € 
dista: 
conve 
imag 
llIm'-. 
mapp 
(algel 
Im'=I 
Using 
meas 
coeffi 
corre 
and | 
Imel