€ Detection of linear features in images. 
€ Systematic errors correction. 
€ Optimization of the linear features 
3.1. Image Smoothing Filter 
Smoothing filter is a general notion of transforming a 
digitized image in some way in order to improve picture 
quality. It mainly consists of removing noise, debluring 
object edges, and highlighting some specified features. 
The paper use edge-preserving smoothing, which 
searches the most homogeneous neighborhood of each 
pixel and assigns to it the average gray value of that 
neighborhood. The homogeneity is expressed in terms 
of variance. When the pixel under consideration lies on 
an edge there will be, when moving away, directions 
where the variance is low, i.e., the pixel belongs to that 
region, and directions with high variance. The principal 
notion is to rotate with an interval (e.g. 450), an 
elongated mask around the pixel and to compute the 
variance of the gray values in the bar. The average of 
the gray values of the bar with the smallest variance is 
assigned to the pixel. 
3.2. Extracting Edges 
Edges of objects (e.g. buildings) in an image are 
defined as local discontinuities in the gray value 
appearance. This may result from a depth discontinuity, 
a surface normal  discontinuity a  reflectance 
discontinuity, or an illumination discontinuity in the 
scene. 
Edge detection has been an important part of many 
computer vision systems and is widely described in 
textbooks and presented in scientific works. There are 
two main types of edge detection techniques which have 
been widely described in literature: the differential and 
the template matching techniques. The former 
performs discrete differentiation of digital image array 
to produce a gradient field, in which, at each pixel, 
gradients are combined by a non-linear point operation 
to create an edge enhancement array prior to a 
threshold operation. The template matching technique 
is based on a set of masks representing discrete 
approximation to ideal edges of various orientations, 
which are applied simultaneously to produce the 
gradient field. In that case, the enhancement is formed 
by choosing the maximum of the gradient array 
corresponding to each mask. For each type of edge 
detection technique, a large number of operators have 
been proposed by different authors. 
In our method, we employed the Sobel operator to 
strength the edges and then used the dynamic 
programming line following method to extract these 
lines. 
3.3. Systematic Errors Correction 
The edge pixel coordinates are defined in the frame 
reference system. These coordinates must be 
transformed to the beast positions in the image by 
correction systematic errors such as radial distortion, 
decentring distortion, scaling difference in horizontal 
and vertical directions, translations of the principal 
26 
point. The systematic errors are corrected using the 
llowing equations: 
X, 7 X; — X, * (x, — xyek,er" * (x, — xy) ed, 
(18) 
Y; ^ Yr— yo * (y; — y) *k,or* 
where: 
x, and y, are the image coordinates of a pixel 
related to the principal point; 
x, and y, are the coordinates of the same pixel in 
the frame; 
x, and y, are the image coordinates of the principal 
point; 
r is the distance of one pixel to the principal point; 
k, is the coefficient of radial distortion (higher 
order coefficients and decentring distortion are 
neglected); 
d, is the scale factor in x. 
3.4. Optimization of the linear features 
Once the system errors of image have been corrected, 
the straight lines in the image plane can be expressed 
as the form of equation ( 1 ) with the least square 
adjustment at sub-pixel precision. 
4. EXPERIMENTAL RESULT 
One simulated test and one real have been conducted 
in order to check the potential and the effectiveness of 
the developed calibration scheme of camera. The 
simulated test control lines, which were extracted from 
a cube rendered with 3D Studio, were used to describe 
the whole procedure of this method. The real date was 
used to make a comparing with the point-based method 
and the new one presented here. For the former, the 
control target was one cubic box. Figure 3 illustrated 
the calibrating target. Table 1 listed the simulated 
orientation parameters of the simulated camera. 
  
92) 3 (b) 
   
3(c) 3 (d) 
Figure 3. Simulated Control Target. 
( a ). Original Image of Control Target. 
( b ). Edge 
( c ). Edge: 
(d). Stra1 
Leas 
Table 1. € 
Interior 
Orientatio 
Parameter 
Exteriror 
Orientatio 
Parameter: 
Table 2 liste 
and their ol 
Table 2. Simula: 
  
  
10.10 
LO 
188.19 
LI 
154.08 | 
L2 
39.65 . 
L3 
241.4 ; 
LA 
172.91. : 
L5 
42.58 | 
L6 
238.44 ( 
[T 
235.29, | 
  
  
  
The softw 
and execute 
WINDOWS 
optical dist 
coordinates; 
system erro 
principal pc 
exterior orie 
lines can be 
sub-pixel pr 
Table 3 1 
result. 
Table 3.