1OMICS,
r a variety of
ation industry
acting features
e feasibility of
long corridors.
em has proved
four-parameter
g an improved
d an analytical
e twofold: 1) a
. Subclasses of
wly developed
ests have been
ric features of
lues. First, the
, which can be
ues, such as
its not à
y corrupt the
t.
feature space,
irried out. The
'ameters of all
ed a data set
| sets obtained
‘the Principal
|. 2003a. As a
ger car, multi-
separated and
ws the vehicle
for automated
ed techniques
| (Duda et al.
jas 1993) are
cles + 30 test
al. 2003b).
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004
LE
6 I=
5 L^ s: ; f Trucks (against)
=
: +
‘©
Lg
I *
= FH Trucks (along)
3 =
+
rur E
Lu: x
2 = "
s
o à
S *
(ET
o
$e
o E
© C
S
0r o )
o
©
a. Passanger cars (along)
-1 il ] l J p EN, ] ] I l J
0 5 10 15 20 25 30 35 40 45 50
Figure 1. Applying rule-based classifier based on PCA.
2.1 PCA REFINEMENT: INTENSITY VALUES
Applying more features could improve the performance of Yength width DL! ho hs m4 intl int dine) "ined
PCA-based classification. Since our latest dataset also contains 3 a7 33 05 29 0.25.0 à
; T ; an 1,33 Q.9. 1,1. 3,2. Idi 87,9. 59, 2 7
intensity values, our next idea was to use them as additional
parameters. Comparing the dataset with spatial coordinates and 5,72 a IRC) ROME na to Tm
the one with X, Y and intensity values, it can be clearly seen
8,18 1, 0,9 i 2 14 55,1 8,1 81,1 5,0
how the vehicles differ from the road surface (Figure 2).
oo ai
Figure 2. Elevation (red) and intensity (blue) road map
Supposedly, different vehicle categories produce different
reflection intensity. For example, an MPV or light truck has
very steep rear window (if any), hence the rear sections of the
car theoretically reflect more points, or points with higher
intensity values, so does a lorry or an eighteen-wheeler.
We stored the intensity values associated with the segmented
vehicles in a database management system. The input matrix
previously contained the following parameters: length, width
and four mean height values (hl-h4) along the vehicle. In the
next approach, the input matrix is extended with four further
columns corresponding to the four mean intensity values (intl-
Int4), sec Table 1.
O2
Un
Table 1. The structure of the input matrix
As opposed to the striking positive results of the PCA method
which is based on geometric parameters, the enhanced
algorithm did not result in good distinguished categories: the
deviation between the intensity values turned out to be too
high..
3. MODEL-BASED CLASSIFICATION
Our next approach is also based on geometric parameters of
vehicles, this method uses height values and the length/width of
the vehicle, and its objective is deriving the shape, i.e., the
profile of the vehicle.
The dataset used as training data for the PCA was acquired over
Route 35 (Dayton, OH). In order to investigate the effect of the
point density, we used a dataset obtained in Toronto (in 2004)
for our shape and profile investigations (Table 2). The dataset
collected in Ohio has relatively modest point density. but, still
we could apply it for classification purposes. However. the data
from Toronto had a higher point density that we were able to
use for developing a refined technique.
