Chen Yuen Teoh 
  
  
  
  
  
  
  
| Attribute | Property of Junction Model Addressed | 
Number of Pairs There might be different pairs of road 
segments that join to form a junction 
Intensity Neighbouring road segments should have similar 
Difference intensity value. 
Width Difference Neighbouring road segments should have similar 
width, or minimum width difference. 
Minimum/Maximum | Junction should have a minimum/maximum gap 
  
  
  
  
  
  
Gap between the closest two edges from the two pairs. 
Minimum/Maximum | Junction should have different angles, and minimum/ 
Angle’s Type maximum angle for a list of angles measured. 
List of There exist different types of angle 
Angles’ Type between two connecting road segments. 
  
Table 2: Level 3 Attributes List 
  
Rule 1: 
Max Gap <= 101.5 
=> class Junction [79.4%] 
Rule 3: 
Intensity Difference > 17.1 
Min Gap <= 2 
"> class Junction [63.0%] 
Rule 4: 
Min Gap > 2 
Max Gap > 101.5 
-> class Not_Junction [91.7%] 
Rule 2: 
Intensity Difference <= 17.1 
Max Gap > 101.5 
=> Class Not_Junction - [33.2%] 
Default class: Not_Junction 
Evaluation on training data (28 items): 
Rule Size Error Used Wrong Advantage 
1 1 20.6% 6 0 (0.0%) 5 (510) Junction 
3 2 37.0% 2 0 (0.0%) 2 (210) Junction 
4 2 8,3% 16 0 (0.07) 0 (010) Not Junction 
2 2 11,8% 4 0 (0.0%) O0 (010) Not Junction 
Tested 28, errors 0 (0.0%) << 
fa) (b) «-classified as 
8 (a): class Junction 
20 (b): class Not Junction 
The original and preprocessed edge image on which experiments based is shown in figure 5. The output image from 
level 1, T-junction extraction result at level 3, the junction recognition level, is shown in figure 6. Figure 6(b) shows the 
result based on supervised learning using the learned rules shown previously, while figure 6(c) is the result of clustering 
approach. In general, the number of edges at this level is about 180 (for this image), reduced from about 260 after level 1 
recognition. The difference between the two approaches is minimal. 
5 CONCLUDING REMARKS 
For junction extraction, the system manages to provide similar results using both supervised learning and clustering 
technique. However, since the supervised learning approach is conducted with very limited examples, the result is far 
from ideal and can be improved upon. 
  
886 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.