The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008
The training samples X. for equation(3) in hyper planes
H 2 , are the most close points to the classified plane in the two
types of samples. Because H 2 looks like supporting the
optimized classified plane, so we can call X. support
vector(SV). In Equation(5), corresponding non-SV, /1. = 0, it
only need to sum up SV X, . The alphabet b is the classified
threshold.
If samples cannot classify in given input space, it should be
mapped to a high dimension space by kernel function 121 , thereby,
in the transform space the samples can be classified by
optimized classified plane. The classifier function getting from
SVM is similar with from neural network. The output of SVM
is the linearity combination of middle layer nodes, and every
middle layer node is dot metrix of an input sample and a support
vector. The process of classifying by SVM is showed as Figure.
3:
output
SV • Sample
mapping
suport vector
input:samples
Step 6
Step 7
Step 8
Step 9
Record every stars in S star ‘ as a sample;
Whether done with the samples?
YES: getting Samples, to
Step 8 ;
training SVM with samples
Classify all stars in the catalogue with trained SVM;
if the output is “1”,select the stars as navigation
stars; record the other stars as non-navigation stars.
NO: to Step 3;
Because SVM is iterative optimization process, so it is
necessary to verify continually what the right SVM parameters
is and select samples for training SVMs again and again until
the catalogue can support navigation applications perfectly.
5. TEST AND ANALYSIS
Taking SAO catalogue as basic catalogue, we designed some
tests for verifying the method mentioned in this paper. Firstly,
given threshold VMT=8.0, we got a temporary catalogue
include 46,136 stars by filtering basic catalogue with MFM.
Then, given FOVg =3°x3°, sampled the temporary catalogue
randomly by FOV§ for getting training samples data.
Then the samples data were separated into two parts, one part
was used as training samples for confirming SVM parameters,
the other part was used as inspection samples for testing the
performance of SVM and optimizing SVM parameters
according to classified results. At last, all stars in the temporary
catalogue were classified by trained SVM, and then we got
navigation-stars catalogue for determining attitude of satellites.
The classification was executed 1,000 times randomly and the
experiment results can be shown by Table 2. FOV is given field
of view when extracted training samples from temporary
catalogue.
Figure3. Operation of SVM
4. NAVIGATION STARS SELECTION BASED ON SVM
Navigation stars selection based on SVM include some steps as
Table.! shown:
Table. 1 Navigation stars selection based on SVM
g 1 Support VMT=8.0, initialize basic catalogue based on
CP MFM;
Select samples in the whole celestial sphere
^ te P ^ according to a given sample interval FOV Inlerva ,;
Select the stars in every sample FOVs FOV'
Step 3
(FOV;<FOV). getting SJ ,
Arrange the stars in S star ‘ from bright to dark
^ te P ^ based on their brightness (Vmag), only reserve
n(n>3) stars from the beginning of list;
Consider 3 of the most brightness stars in S star ‘ as
^ navigation stars, and label with “1” ;the other stars
labeled with “0”,as non-navigation stars;
Method
Capability
of
catalogue
Density
navigation-stars
FOV
of
in
0
<3
<5
>=5
MFM8.0
46136
-
-
.
-
SVM8.0
9117
0
0
46
954
MFM7.0
15914
0
0
0
1000
MFM6.5
9023
0
4
37
959
SVM6.5
(FOV=3)
6936
0
1
29
970
SVM6.5
(FOV=6)
7685
0
0
43
957
Table.2 Experiment results
From Table.2, we can find that the capability of catalogue based
on SVM is obvious smaller than that based on MFM and the
density of FOVs has notable change. For example, when
VMT=6.5, the capability of catalogue based on MFM is 9,023
and the probability of stars fewer than 3 in FOVS is 0.4 %. But
the capability and probability of catalogue based on SVM is
6,936 and 0.1%. When adopted MFM, in order to avoid empty
hole of field, the VMT must be 7.0 and the capability is 15,914.