ED 
late, 
ation 
have 
racy, 
plate 
plate 
m at 
road 
ad of 
oach. 
years 
and 
mage 
ds to 
e are 
So, 
other 
e the 
road 
ming 
ation 
m of 
from 
ental 
Hu Xiangyun 
2 OVERVIEW OF THE ALGORITHM 
To realize a practical system of semiautomatic road extraction, we should take four factors into account. They are good 
accuracy, steady output, high speed and good interactivity. Steady output is to say the algorithm is not sensitive to noise 
and the slight variation of position of seed point. High speed means that operator needs not to wait for the extraction 
result after giving the seed points and good interactivity make it easy to use. 
Most of semiautomatic extraction is based on optimization by radiometric and geometric constraint of the object (Gruen, 
Li H H, 1995). An object function that describes the constraints is often derived to model the object. The procedure of 
extraction is just to match the model with the image, once the match is successful, some parameters of the function is 
the result (eg. coefficients of spline). This is called top-down or model driven strategy. For example, the algorithm of 
graph searching realized by dynamic programming (Gruen A, Li H H, 1995) and active contour (snakes) which is a 
method of ‘energy minimization’ (Trinder J, Li H H, 1995, Gruen A,Li H H., 1996). According to former ways of 
interactivity, one gives start point and tracking direction (Quam L H, 1985). However, sometimes the result is 
unpredictable. By active contour, operator give some seed points as the global constraint of shape of the object, if the 
extraction fails, one has to do it again. It is not so good interactivity. 
Our algorithm and interactive operation of the extraction are different with above-mentioned methods. We do not derive 
a uniform mathematical model of the road feature. Our approach is a step-by-step extraction. Firstly, a pair of seed point 
limits the searching scale in which some possible points are attained by correlation of local radiometric template. Then 
the segmented curve fitting eliminates the blunder points, finally a Hopfield neural network is built to optimize the 
result. By this approach, extraction algorithm is running in a pair of point. The interactive mode is ‘what you see is what 
  
you get’. That is to say, the new extraction result appears 
  
  
while the operator gives the new seed point. During the Input road width 
procedure, one can expediently cancel the extraction in the El 
  
‘segment’, furthermore, the manual positioning operation : 
: . Input new seed point 
could be inserted with no influence of entire extraction. At 
  
  
  
  
  
  
  
the end of entire extraction, all points of these segments i 
will be fitted by spline. Figure 1 is the procedure of the Image segment resampling 
algorithm. I 
3 STEPS OF THE ALGORITHM Template matching 
3.1 Input Road Width and Seed Point x 
  
  
Segmented curve fitting 
  
  
  
The road width should be input through interface, such as 
dialog input or manual measurement. When the width is | 
  
wider (larger than 10 pixels), operator can give two points Hopfield network optimization 
  
  
near two sides of the road, then the algorithm 
  
  
automatically finds the real edges and calculates the width. 
  
  
  
  
  
+ 9 
3.2 Image Segment Resampling Cancel? 
N 
A rectangle area between a pair of seed point is resampled 
| Input next seed point 
from the whole image. The range of the area is related to 
the road width and the distance between the two points | 
(see Figure 2). Figure 1. Procedure of semiautomatic road extraction 
3.3 Template Matching 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 995