In this paper, a relational graph representation, in terms of 
object boundaries, is developed. The graphic representation is 
described by planar surfaces which are bounded by straight 
lines or regular curves. These surfaces are grouped in terms of 
their normal directions and are stored together with their areas 
and perimeters as matching elements. The topological relations 
between surfaces are constructed in terms of the centre of each 
surface and the common edge of two surfaces. The automatic 
procedure for a machine vision system for FMS includes three 
steps: image processing techniques for the extraction of 
features on the industrial components and hence the 3- 
dimensional measurements of these components; representation 
of the 3-dimensional objects in an efficient and convenient data 
structure; artificial intelligence procedures for the recognition 
of the objects by comparing with models stored in the design 
database. Figure 1 illustrates a schema of the procedure by 
means of which objects in the scene can be reconstructed and 
recognised from digital stereo images. 
2. RECONSTRUCTION OF OBJECTS IN THE SCENE 
Images of industrial components are generally characterised by 
sharp discontinuities which represent features on the objects. In 
order to describe features by their boundaries and to use 
feature-based image matching in photogrammetry, it is 
necessary to extract the complete details of linear features of 
the object from its edges, and to represent them in a suitable 
data structure such as straight lines and smooth curves. By 
matching these geometric features in stereo images, the 3 
dimensional geometry of objects can be reconstructed. 
    
Figure 2 : An image of a block 
2.1 Edge Detection 
Edge detection is low level image processing, which serves to 
simplify the analysis of images by drastically reducing the 
amount of data to be processed, while at the same time 
preserving useful structural information in images. The edge 
detection method (Trinder & Huang, 1993) used in the research 
is based on a linear model which locates an edge point in the 
operator window with subpixel accuracy. The linear model 
comprises two aspects: one is to determine the peak of the 
intensity change in gradient direction, which is performed by 
the Förstner Operator (Förstner, 1987); the other is to limit the 
unstable edge location in the edge direction by the introduction 
of a linear constraint which passes through the window centre 
and meets the edge at right angles. 
To improve the accuracy of the edge location, an edge point is 
determined by a weighted average of two points derived from 
both sides of the edge. Another implementation is to use a 
round operator window instead of a square window, so that the 
result will not be influenced by the difference of edge 
orientations. The scattered edge points are then chained by 
following neighbouring pixels, based on the minimum local 
distance. The direction of the edge chaining corresponds to the 
local direction of the edge, which is attached to each edge point 
as a primitive for the succeeding process. Figure 3 displays the 
edges detected from the image of an industrial component 
shown in figure 2. The edges can be located with a precision of 
0.1 pixel, depending on the contrast of edges. Generally, high 
contrast, which reduces the influence of noise, results in high 
precision of edge location. 
  
  
  
  
  
Figure 3 : Edges on the block 
2.2 Line Segmentation 
In industrial environments, regular shapes such as ellipses and 
straight lines, often occur as elements of object boundaries. In 
order to interpret objects in the scene, it is generally more 
relevant to present the boundaries of objects revealed in the 
images in geometric form, because simple geometric functions 
provide more reliable information and are easier to calculate. 
The suitability of the approach developed in the research for 
line segmentation is demonstrated in (Huang & Trinder, 1994), 
based on the analysis of the local directions of edges. The local 
direction of an edge point is a 1-D value which clearly reflects 
the trend of the edge at the edge point with respect to the next 
point. If a group of edge points contain the same trend in edge 
direction, it means that these edge points are of similar 
geometry. The method attempts not to find corners, but directly 
to find the basic components of straight lines or regular curves 
by the assessment of the local direction at the edge points along 
a complete edge. The principles of the line segmentation are: 
all edge points on a straight line should indicate approximately 
the same edge direction; and all edge points on a regular curve 
should indicate the same sign of the difference in direction. 
  
  
  
  
Figure 4 : Results of line segmentation 
254 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B5. Vienna 1996 
The meth 
lines or r« 
at the edg 
curves ar 
parameter 
the same 
be merge 
close bou 
open regu 
surface [ 
geometric 
segmenta 
2.3 Stere 
Using twi 
be constr 
reliability 
based on 
of straigh 
terminals 
straight 1 
terminals 
To find t 
consider 
tangentia 
a matchi 
basic req 
The mat 
straight ] 
figure 5. 
ellipses 1 
correspor 
Fig 
The calci 
Special e 
surface, 
an image 
plane w