The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
4. EXPERIMENT 
Road intersection detection experiment was conducted by using 
the double circle. 
4.10bject Image Used in This Experiment 
The proposed road intersection detection algorithm was applied 
for QuickBird image. The object image is shown in Figure 1. 
This image is obtained on 7 June 2002. This area is included in 
Kodaira city that is located at the west of Tokyo in Japan. There 
are many houses, many roads, many bare grounds , many trees 
and many glass lands in this area. These roads have the variety 
of width. 
4.3 Road Intersection Extraction 
Multi spectral image has 4 band images that are composed of 3 
visible band images and one infrared band image. The ground 
resolution is 2.44m. On the other hand, panchromatic image has 
high ground resolution that is correspond to 0.61m. Multi 
spectral image can be used for land cover classification. But, 
resolution is not so high. Panchromatic image can be used for 
object detection. Road area has many land covers, for example, 
cars, central lines, pedestrian crossings and stop lines. Color of 
car has a many variety. Most of the road has characteristics of 
asphalt. When road is classified by using ordinary land cover 
classification method, cars, central lines, pedestrian crossing 
and so on becomes obstacles, in order to remove these obstacles, 
two methods are selected. One method is to calculate the NDVI. 
Another method is to extract the saturation image from HIS 
transformed image. These obstacles are removed in both images. 
Some roofs have same characteristics of asphalt. But, these 
areas do not continue to some direction. Road intersection has 
three or four continuous direction. By using this difference, road 
intersection can be detected. Peripheral circle was introduced to 
detect the roads from road intersection. 
4.4 Generation of Mask Image 
Mask image is generated from NDVI and saturation images. 
The image is shown in Figure 3. Black area shows removed 
pixels which have vegetation or high saturation. White area 
is candidate pixels of roads. 
Figure 1. QuickBird image used in this experiment 
4.2 Double Circle Used in This Experiment 
The size of main road in target image is about 12m width which 
is correspond to 20 pixels in the object image. Main road 
intersection is surrounded by pedestrian crossing in actual 
image shown in Figure 1. The radius of core circle should be 10 
pixels. Peripheral circle radius is selected 20 pixels. Proposed 
double circle is shown in Figure 2. 
T 
Figure 3. Mask Image 
4.5 Supervised classification 
Supervised classification method was applied masked image 
of NDVI, saturation Image and panchromatic image. Classified 
result of road is shown in Figure 4. 
Figure 2. Example of double window 
506