3.2. Stereo Matching and Computing 3-D 
Coordinates of Edge Points 
Although the constraints are similar, the matching 
procedure for edge points differs somewhat from that 
used for targets (El-Hakim, 1989). Since no isolated 
points exist, some must be extracted from the edges. 
For the left image in each pair, edge profiles are 
taken at preset intervals to reduce the edge to an 
appropriate set of points and the  sub-pixel 
coordinates of these points are determined by 
convolving a step-edge function with the Gaussian 
image function and equating it to the edge profile 
followed by a least-squares adjustment to find the 
unknown parameters of the step edge. For the right 
image, since the corresponding points are not yet 
determined, profiles are taken every pixel and the 
sub-pixel image coordinates of these edge points are 
computed. We now have a set of points for the left 
image and a denser set for the right image. The 
matching procedure is then applied as follows: 
1. For each edge point from the left image, the 
epipolar line on the right image is computed; 
2. The epipolar line will intersect several edges in the 
right image, out of which only one is correct; 
3. Constraints, such as the disparity and depth 
constraint, plus the added knowledge of the shape 
of the complete edge, are applied to find the correct 
match. 
Some edges are determined from one pair, while the 
others are determined from the other pair to avoid 
those parallel or at small angles to the epipolar line. 
The 3-D coordinates of the edge points are determined 
and separated, or labeled, according to each isolated 
edge. 
3.3. Automated Part Measurement 
The current VCM system is capable of automatically 
measuring a finite set of parts. Some parts can be 
measured automatically with very little operator 
intervention while the automated measurement of 
some other parts require that a setting up process be 
performed by an operator familiar with the 
measurement processes involved. In order to fully 
automate the inspection of a variety of parts, it was 
necessary to develop a strategy for dealing with parts 
in a very general way. Since it would be difficult to 
foresee all the possible requirements associated with 
newly encountered parts, it was recognized that the 
design strategy should allow for straightforward 
system expansion. 
.The following measurement strategy has been 
designed to meet the requirements just outlined (see 
also figure 3). The automatic inspection process has 
been divided into two main steps. In the first step, the 
system acquires all the measurement data, using a 
wire-frame model as a guide (see section 3.1.3). After 
this step, the measurement data will exist as sets of 
labelled points. In the second step, the system uses the 
data to perform the inspection. Dividing the 
inspection process into these two main steps has the 
benefit of allowing for the possible future integration 
of the system with other measurement devices. Since 
the second part of the measurement process receives 
its input as sets of labelled points, these points would 
not have to necessarily be provided by the machine 
vision system. It is possible that the points come from 
several different sources such as a coordinate 
measuring machine or laser scanner. Thus, the 
possibility for a multi-sensor inspection system exists. 
The measurement step consists of two tasks. The first 
task is to locate the part. In order to perform part 
projection, the part must be located by the system. 
     
This can be done using some of the more prominent 
features of the part. Using some of the features to 
locate the part and then using the model to find the 
rest of the part features is preferable to having the 
system find and identify all of the features without 
any guidance. At the current time, target type 
features have been successfully used to locate parts so 
that part models can be applied. Primary edge 
features could also be used to locate parts. Once the 
part has been located, the second task, acquiring data 
points, can be performed. Each distinct feature, as 
defined in the model, is measured separately. Using 
the edge projection technique to mark features 
uniquely or in isolation removes the problem of 
having to deal with multiple edges when performing 
matching of edge points, thus virtually removing the 
possibility for error. Another benefit of measuring 
features separately is that the data points belonging 
to specific features are easily maintained as sets of 
data points labelled according to the feature in the 
model to which they belong, thus alleviating the need 
to perform feature recognition on the data sets. 
The second part of the measurement process - consists 
of dealing with the measured data. This step can also 
be divided into two parts. In the first part, the system 
is provided with the information it needs to derive 
further feature information from the measured data. 
Such features might include corner points, surfaces, 
or even invisible features such as circle centers and 
imaginary lines and boundaries. Once these features 
are obtained they may be used in calculations in the 
same manner as the directly measurable features. 
Inspection calculations can then be driven by a list of 
inspection orders. Each inspection order will give all 
the information necessary to calculate one parameter 
value. The first field in the inspection order will not 
only identify the parameter to be calculated but will 
also uniquely identify the subroutine used to calculate 
it. Subsequent fields will identify the data sets 
involved in calculating the parameter and the 
tolerance specifications for the parameter. For 
example, the entry "distance point to line 63 27 19.99 
20.01" would be a request to calculate the distance 
from “point 63” to “line 27” and make sure the value 
is between 19.99 cms and 20.01 cms. As new types of 
inspection tasks are encountered, new routines can 
easily be added to the system to handle them without 
having to redesign the rest of the measurement 
system. 
4. CONCLUDING REMARKS 
The edge measurement procedure described is useful 
for applications where the object is known 
(dimensional inspection and object tracking) as well 
as those where the object has no available CAD data or 
model and it is required to create such model (reverse 
engineering). In the inspection and tracking 
applications, the procedure is fully automated, except 
for the set-up, while in reverse engineering the 
procedure is interactive. The VCM system which 
applies these procedures is available in two forms: an 
application program, targeted for the above 
mentioned applications, and is hardware specific; and 
library routines which are not hardware specific and 
not targeted for a specific application. The later is 
intended for users expected to develop their own 
application oriented software and user interface.