7A-3-5 
(b) Part 2 
Figure 4. Part of an AIMS image imposed with extracted regions. Exterior orientation parameters 
of the image are obtained by GPS and INS. 
4. TRUCK RECOGNITION USING TWO LAYER 
HOPFIELD NEURAL NETWORK 
Figure 4 illustrates an AIMS (Airborne Integrated Mapping 
System) image from which trucks moving on a freeway are to 
be recognized. The height of a truck is relatively small in 
comparison to the flying height so that the top of the truck 
appears as a relatively light region in the image. It is clear that 
the truck’s shadow plays an important role in telling whether or 
not a region belongs to a truck. Given the interior and exterior 
orientation parameters of the image along with an approximate 
elevation value of the freeway, we could back-project a 3-D 
model of the truck as well as it’s shadow to the image. The 
trucks and their shadows don not change much in terms of site 
and shape. This allows us to match extracted regions (truck 
top) with the back-projected model directly. The shadows, 
unfortunately, differ from one to another. Some trucks may not 
have complete shadows. A two-layer Hopfield network 
recognizes not only the regions similar to the back-projected 
model without shadows but also those not very similar to the 
model with strong support of shadows. The relationship 
between each pair of regions (potentially a truck top with its 
shadow) and the edges is examined by the two-layer Hopfield 
network simultaneously. 
Figure 5. Recognized regions in Figure 4. States of the 
neurons are listed in Table 1. 
Line pattem layer 
Line pattern layer 
Modelline ID 
Modelline ID 
line Id 
(Region 
Id) 
0 
1 
2 
3 
line Id 
(Region 
Id) 
0 
1 
2 
3 
296(74) 
0028 
0000 
1000 
0000 
164(41) 
loco 
0D00 
oooo 
oooo 
297(74) 
0000 
0031 
oooo 
10C0 
165 (41) 
oooo 
1000 
oooo 
oooo 
298 (74) 
1000 
ooco 
0028 
oooo 
166 (41) 
oooo 
oooo 
loco 
oooo 
299(74) 
0000 
1000 
OOOO 
0031 
167 (41) 
oooo 
oooo 
oooo 
1000 
Region pattern layer 
Region pattern layer 
Model region ID 
Model region ID 
Region 
Id 
0 
1 
Region 
Id 
0 
1 
74 
1.0» 
0.002 
41 
1.000 
0.245 
265 
0 000 
1000 
Table 1. Extracted neuron states of the two-layer network. The 
shaded cells mean that the corresponding line or region 
matches with the model feature. Note: Line Ids are not shown 
in Figure 5. 
The network recognized 10 out of 12 trucks in the entire image. 
Figure 5 illustrates two recognized trucks. In Figure 5(a) the 
extracted edges and regions are of high quality. Region 74 
describing the top of the truck matches with the model top and 
the adjacent region 265 describing its shadow matches with the 
shadow of the model (Table 1). The edges of Region 74 match