Hu Xiangyun 
  
The local gray feature that is defined by road ‘profile’ forms the template. Compare to normal gray level and rectangle | 
shape template, this particular template is a single pixel width and binary (0-1-0) template. Within the image segment, | 
the template moves along the vertical direction of the road (vertical to segment axis) to calculate the correlation 
coefficient with the image, see Figure2.The calculation gets several peak values that are stored in a structure list. 
According to the definition of correlation coefficient, suppose the correlation coefficient between the signal p and q is c, | | 
q, =aq+b, then the new value of correlation coefficient between p and | 
  
: ^g 
q,is cízc. 
The correlation of binary template could not only save the time by — 
transferring the correlation to a partial adding calculation, but also Hl. 5s 
remove the influence of local different contrast between the road and Ln | | | 
  
the environment (linear distortion of the binary template function). 
Furthermore, the pixel by pixel moving correlation could take the fast 
algorithm. The algorithm extremely speeds up the time consuming 
correlation computation. 
For the wide road (the width is larger than 10 pixels) on the image, the 
template matching proceeds at a multi-layer image structure, like 
(Heipke C, Steger C. A., 1995). The results of matching are transferred 
from lower resolution layer to higher resolution layer. However, this 
has been proved as a more effective and reliable strategy. The layer 
count is dominated by the road width. This multi-resolution correlation 
lead to the matching for wide road extraction that is as fast as that of 
few pixels width road. 
  
Figure 2. Binary template correlation 
3.4 Segmented Curve Fitting 
After above processing, most of the max-correlation points are the right centerline point, the errors only occur in the 
column direction of the segment. According to the geometric constraint, road is a continuous and smoothly changed 
shape. The local wiggles and burrs are regard as errors. So a segmented curve fitting by least square could be used to 
remove these blunders. Here a quadratic curve defined by the equation: y=a,+a,x+a,x" is to iterate the processing of 
fitting with eliminating the errors that are larger than a restriction until all final errors are less than a given restriction 
value. The length of segmented curve is settled by the road width since most of the wider roads in the image have less 
curvature and variation of curvature. 
  
3.5 Optimization by Hopfield Network 
The segmented curve fitting could remove most of blunders. 
    
But sometimes the segmentation causes the discontinuous 
PJ 
     
among the segments. In the case of heavy shadows and other 
  
noises, some max-correlation in a short segment might be Bi 
wrong. We need a global optimal computation to select an 
| 
I 
| 
I 
I 
| 
| 
| 
| 
optimal solution that accords with the radiometric and | 
| 
  
| 
I 
I 
| 
| 
I 
| 
| 
| 
L 
geometric constraint of the road. TD 
  
  
Figure 3. Hopfield Model 
Artificial neural network (ANN) is a promising and widely 
used method in the field of pattern recognition. If the initial state of the network is the noise-polluted pattern and the 
minimal value of network energy defined a known pattern, the network that is built with the weights, I/O characteristic 
  
996 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.