Istanbul 2004 
P-2001) Salt 
, 2001. 
, Detection of 
tation, Poject 
mbre 2002). 
A NEW APPROACH TO AUTOMATIC JUNCTION OF OVERLAPPING AERIAL 
IMAGERY DATA 
Yuri B. Blokhinov *, Dmitry A. Gribov n 
State Research Institute of Aviation Systems (FGUP GosNIIAS), Moscow, Russia 
* blokhinov@gosniias.ru, ° gda@gosniias.ru 
Working Group III/2 
KEY WORDS: Automation, matching, recognition, detection, feature, imagery, mosaic, aerial 
ABSTRACT: 
The original approach to image matching is proposed. The method itself can be classified as relational matching, bases on point 
features. For robust extraction and filtration of features the special procedure, based on dynamic resampling technique, was 
elaborated. Further the rotation invariant relations among the features are used to confirm or reject initial hypothesis. All calculation 
procedures are time effective and invariant to images rotation. Finally, the approach given is applied to two different tasks: 
automatic mosaic creation from video camera sequence frames and automatic relative orientation of photographic camera images. 
1. INTRODUCTION 
Image matching is the task, aroused in many different 
applications. Both input data and practical aims can differ, but 
the underlined principles are the same. So the task under 
consideration should be of interest for wide range of specialists. 
Here we try to develop the sort of feature based relational 
matching as the most suitable for comparison of large images. 
Considerable efforts was done by the investigators in this 
branch, many interesting results was obtained (Heipke, C., 
1996, Woozug, C., ., 1996), each optimal to use in it's specific 
domain. The method, described below, was elaborated for real 
technical applications and two properties was obligatory for it: 
to work in the near real time (minutes, not hours) and to give 
reliable results. 
In short, after some kinds of special pre-processing procedures, 
image can be represented as a set of spatially distributed 
features. Each feature is unique and, in general, can be 
described by some digital parameters and hence can be 
distinguished among another features. Main features types are 
(Henricsson, O., 1996) points, lines and regions. For each type 
the specific methods are elaborated to extract it from image. 
When all substantial features in the image are extracted, their 
relative coordinates with respect to each other can be fixed. 
Now we can say, that image is described by the finite set of 
numbers, features’ parameters and their relative coordinates, 
and to compare different images in the formal mathematical 
way. The main problem is that most of methods used at present 
for image recognition require considerable time to implement. 
This is due to the fact that complex feature extraction by known 
algorithms is very time-consuming procedure. This paper 
introduces one approach to relational image matching, suitable 
for performance in near real time. 
2. VIDEOCAMERA SEQUENCE FRAMES JUNCTION 
2.1 Task and data 
The input data are video shooting obtained by swinging camera 
from airplane. Raw material can be cut into sequence frames, 
which are considered as a set of digital images, the overlapping 
is 40-80%. The total sequence sometimes includes thousands of 
frames and cover large area of the surface. Mosaic of these 
frames, built up on-the-fly, is of considerable interest in some 
practical applications. In the given case “to build on-the-fly” 
means to build automatically due to very large number of input 
images. Substantially that all algorithms should rotational 
invariant and non sensitive to variations in brightness level 
among different frames. 
  
Figure 1. Sample of video frames 
2.2 Features extraction 
Proper choice of features is the key part of relational matching 
(Henricsson, O., 1996). A reasonable compromise should be 
found between the informativity and complexity of the features 
at hand. Lines and regions are informative and stable though, 
they requires much time for extraction and handling. So for 
*build on-the-fly" algorithm only point features were taken into 
consideration. As the index for interest points extraction the 
variance of image brightness V (x, y) was taken. Variance for 
window of size NxN, centred at xo, yo is defined as follows: