International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
136 
Industrial Area 
T T 
◄— part-of 
•4k- — con-of 
•*— is-a 
— temp-reI 
[3, 
oo] 
[1.00] 
Hall 
Parking Lot 
—W 
Figure 6. Simplified semantic net for the extraction of an industrial fairground introducing temporal relations. The more concrete representations of 
the objects (physical and sensor layer of the semantic net) are illustrated here by corresponding image regions. 
phase. Having found hints for all obligatory states, a complete 
instance of Industrial Fairground can be generated and the 
interpretation goal is reached. 
5.3. Results 
The presented approach is currently being tested for a sequence 
of five aerial images (three colour, two greyscale images) of the 
Hannover fairground - the future Expo 2000 exhibition ground. 
The images cover the years 1995 to 1998 and show a 
construction/dismantling phase in 1997 and an ongoing fair in 
1998. Furthermore, the construction of three new halls can be 
observed. Unfortunately, no continuous sequence of aerial 
images exists which depicts all phases of a single fair. But the 
given images are suitable to simulate the whole cycle. They were 
coregistered and resampled in resolution pyramids of 0.5 to two 
meters per pixel to permit the segmentation of both large halls 
and small vehicles with minimal processing effort. 
The multitemporal features described in chapter 5.1. were 
integrated in the AIDA system and the semantic net illustrated in 
Fig. 6 was implemented. Additionally, a number of special image 
processing algorithms are necessary to realize the aspired 
application. Halls and parking lots have to be segmented, trucks, 
persons and cars must be detected on the fairground and the 
parking lots respectively. Figure 7 shows preliminary results for 
a colour image of 1997. 
The aerial image in Fig. 7a was classified at a resolution of two 
meters per pixel using a Maximum-Likelihood operator which 
exploits all available image bands. The classifier considers the 
classes of the 3x3-neighbourhood to modify the prior probability 
of the current pixel. This results in a more homogenous 
classification result and suppresses small noisy regions. The 
training regions were defined manually in the first image of the 
sequence and stayed the same for the following images. Figure 
7b shows the classification result for the class Hall (marked 
grey). From all candidates the semantic net chose the white ones 
to be a hall using features like area, elongateness, compactness, 
luminance value, and variance of the corresponding image 
region. The expected feature values were defined prior to the 
analysis in the attributes of the concept net according to the 
human experience. 
To verify the states Fair Construction and Fair Dismantling, the 
system looks for trucks on the fairground which appear as small 
bright rectangles in the image. The characteristic width and 
length of a truck is stored (in meters) in the semantic net, which is 
transformed into pixel units and used as expectation. To limit the 
segmentation spatially to the immediate neighbourhood of the 
halls, the detected halls are used to define a valid search area in 
the current image of 0.5 m/pixel resolution. Assuming an 
accurate segmentation of the halls, it can be avoided to confuse 
trucks with the skylights of the halls. For a part of Fig. 7a the 
search area is shown in Fig. 7c as a dark region, the detected 
trucks are marked white. If the number of detected trucks is larger 
than a given threshold defined in the semantic net, the hypothesis 
Trucks near Halls is regarded as verified. Consecutively, the 
construction or dismantling phase can be instantiated. A final 
distinction between construction and dismantling becomes 
possible, only if the preceding state of the fairground is known.