Information Technology Reference
In-Depth Information
movement imageries. Details of experimental setup can be found on [10]. We used
the downsampled version; 100 Hz, with 118 electrode channels.
For the given data, each subject has different number of training set and test set, we
used first session as training set, and the remaining 3 sessions (thus, there are totally 4
sessions) were used as test set. This dataset was treated as single trial EEG for analy-
sis. Trials were classified into motor imaginary of right hand vs. foot. Evaluation is
based on classification result of each trial.
5 Experimental Results
This section showed results of our both selection approaches and the results of CSP
using different values of
m
. The selection results of BCI Competition IV; dataset 1,
can be seen on table 1, and BCI competition III; dataset IVa, can be seen on table 2.
Table 1.
Selection results for value of
m
, testing on BCI Competition IV dataset1. There are 3
pairs of classes: task1, task2 and relax.
Subject
Frequency and
Time window
Between
Auto selection
m
Semi-auto
selection
m
Task 1 vs. Task 2
2
2
7-28 Hz,
0.75-4.5 s
a
Task 1 vs. Relax
3
3
Task 2 vs. Relax
24
24
Task 1 vs. Task 2
1
1
11-14 Hz,
0.5-4.75 s
b
Task 1 vs. Relax
3
3
Task 2 vs. Relax
1
1
Task 1 vs. Task 2
2
3
7-26 Hz,
0.5-4.25 s
f
Task 1 vs. Relax
3
4
Task 2 vs. Relax
3
4
Task 1 vs. Task 2
3
3
7-26 Hz
0.5-4 s
g
Task 1 vs. Relax
6
6
Task 2 vs. Relax
6
10
Table 2.
Selection results for value of
m
, testing on BCI Competition III dataset IVa
Subject Frequency, Time window Auto selection
m
Semi-auto selection
m
aa
11-16 Hz, 0.5-3.5 s 7 9
al
11-26 Hz, 0.75-4 s 1 4
av
9-25 Hz, 0.75-3.5 s 7 7
aw
10-24 Hz, 0.75-3.75 s 1 3
ay
8-27 Hz, 0.25-3.5 s 2 2
Figures 2 and 3 showed the plots of accuracy against the value of
m
obtained from
for the first and second datasets respectively.
