Wy, = xD (3.3)
Vig = 0.501 X M, - 0.510812, 1
+ Log(W, 4)
120M 10.
Here >, denotes extremal pixels
correlation matrix. Uy denotes l and J
regions average and "y denotes Hy, a
priory probability.
Let's find Hoc region.
Filter impuls response and frame
difference fild_ parametrs _ allows to
define easily 3.4 and Hag (Jens,
1989). Fig. 1 shell be used for Wi
defining. Apriory probabilities of any
pixel belonging to one of the regions
followed Fig.1 are:
=
Il
=
N
=
AA AB BB BA
=
|
2
IW, (1 -W. ))
Wie = Wye = Wop = Wop = WW,U1-W, (2-W.)1
- 5 - - ge - 2
Wan d Wap v WA - Won = WW. (1 V.)
Wc = L1-WAR-W)3®
Wen = Wc = WW, [1-W, (2-W,)]
2 2
Win = (W,W.)
Here moving object pattern area is
W. about frame area and pattern has not
moved to LA about his area for time
distance T.
We supposed in what follows that W,-0. In
this case it's easy to make sure (see
(Jess, 1989)) that regions in which
hypothesises are true are placed as it
a= X W
1#C
J#C
B= Woo Poo (RIZE)
+ P TIX
13 PB, 4 (XI Xen)
P„=a+ß,
where Qc is hypothesis E C acceptence
region. 7
nf
g ac
N Z
pod | Ar
BA , 'CA T
PT e E Tm rr cert
i Q u^
| ec zz |
^ »
s m Y ob
m m.
oy rien re
LE
e :
S MEAT d
AS
ar Magi Mog | "AB
Fig.2 Hypothesises aooeptenoe
regions. Admissible hypothesises
are above heavy line.
h 1comj ce =
As Cin £ ob admissible hypothesises
are above heavy line in Fig.2. Then H,_,
HAC and HE hypothesises are not
admissible. This makes some easily total
error calculating. Nevertheless it
remains very tradicus.
As previous resercher (Holben, 1980) we
have examined some particular cases which
are typical for airborne Sensors.
Examined range of the pattern and
background parameters is shown in table
4.
Table 1. Range of examined parameters
Field type
Parameters
Background Pattern Noise
Average 0.240-0.400 | 0.320-0.550 0.000
Standart 0.036-0.265 | 0.042-0.200 | 0.0I4-0.044
deviation
Correlation| 0.523-2.0I3
constant
I.043-5.996 | 25.56-3I.I2
shown in Fig.2.
As it follows from Eg.(3.1), Eq.(3.2) and
Eq. (3.3) hyperquadrios is defined by
hypothesises parameters and may have
different forms in some cases. In Fig.2.
we show the simplest case in which random
filds are not correlated.
It's easy to make sure that with above
restrictions total error of detector is:
430
Here average and standart deviation is
defined about maximum pixel value.
Correlation constant is defined about
characteristic pattern size.
The dependence of the total error on the
background displacement for typical scene
in some seasons is shown in Fig.3. We
supposed that W,-0.001.
In Fig.3 displacement value is defined
about characteristic pattern size.