Information Technology Reference
In-Depth Information
robust against scene illumination changes. Hence, the extracted set of features should
be invariant to a set of image transformations. Five general types of such changes
related to different lighting conditions are defined in literature [
49
]. The general pixel
color transformation model is defined according to Eq.
12.1
⊡
⊤
⊡
⊤
⊡
⊤
⊡
⊤
R
G
B
a
00
0
b
0
00
c
R
G
B
o
R
o
G
o
B
⊣
⊦
=
⊣
⊦
⊣
⊦
+
⊣
⊦
(12.1)
where RGB represent image values in corresponding channels,
o
is the
offset and
a
,
b
and
c
denote scaling factors. From the relation presented in Eq.
12.1
,
five different types of illumination changes can be listed:
[
R
,
G
,
B
]
•
Type 1:
a
=
b
=
c
and
o
R
=
o
G
=
o
B
=
0—light intensity change
•
Type 2:
a
=
b
=
c
=
0 and
o
R
=
o
G
=
o
B
—light intensity shift
•
Type 3:
a
=
b
=
c
and
o
R
=
o
G
=
o
B
—light intensity change and shift
•
Type 4:
a
=
b
=
c
and
o
R
=
o
G
=
o
B
=
0—light colour change
•
o
B
—light colour change and shift
To show the usefulness of the particular image descriptors they should be tested
against the enlisted transformations. Most extensively utilized descriptors in the topic
of multi-camera object tracking include SIFT-like features [
5
,
23
,
47
].
In this work, a set of feature extraction techniques is chosen, to show efficiency
of other solutions. Besides the regular image colour histogram, which is calculated
for comparison reasons, each of the descriptors shows resistance to at least one of
the illumination transformations. The related information is presented in Table
12.1
.
Another important feature property is its geometric invariance. Object dimen-
sions can vary due to different camera placement and perspective. Additionally, for
non-levelled cameras and in case of lens distortions, objects visible in the scene
can be rotated. Some of the descriptors listed in Table
12.1
would suffer from
these conditions. To overcome this problem, during image preprocessing, objects
are rotated accordingly and resized to defined dimensions to introduce rotation and
scale invariance.
Type 5:
a
=
b
=
c
and
o
R
=
o
G
=
12.4.1 Colour Histogram
Two types of colour histograms have been selected as visual image features. The first
one is full RGB histogram of an object image (
HistFull
descriptor—Fig.
12.3
).
The second one is a two-channel histogram using a chromatic space of
R
c
G
c
colours
(
Hist
descriptor). The values of
R
c
and
G
c
in the chromatic space may be derived
from a RGB pixel through the following formulas:
R
G
R
c
=
B
,
G
c
=
B
.
(12.2)
R
+
G
+
R
+
G
+
Search WWH ::
Custom Search