Biomedical Engineering Reference
In-Depth Information
Furthermore, Rubine's algorithm involves a covariance matrix for gesture c and
all features i, j defined in
E
c
−
1
ci j
=
f
cei
−
f
ci
f
cej
−
f
cj
e
=
0
where
f
cei
is feature i, for gesture c, of the e sample (ranging from 0 to total samples
E
c
)
,
f
cej
is feature j for gesture c and sample e, and
f
ci
,
f
cj
are the mean feature
vectors for both features being compared. The covariance matrix is then averaged
with
C
−
1
c
0
ci j
ij
=
=
+
C
−
1
c
=
0
E
c
where
ci j
is the covariance matrix for features i, j and gesture c, divided by total
gestures C plus the summation of all sample totals
E
c
for each gesture c.
With all of these calculations done on the training data, the weights for the main
Rubine algorithm can be calculated with
−
C
F
−
1
ij
f
ci
w
cj
=
i
=
1
where
−
1
ij
is the inverse covariance matrix for gestures i, j and
f
ci
is the mean
feature, i for gesture c, and
F
is the total number of features in the feature vector.
Finally
w
c
0
is calculated with
F
1
2
f
ci
w
c
0
=−
w
ci
i
=
1
f
ci
is the mean vector of feature
where
w
ci
is the weight of feature i for gesture c and
i for gesture c.
The algorithms show the linear classifier is heavily dependent upon the feature
set, which is purposely left to you to define depending on your device and gestures
to recognize. For example, in Hoffman et al. [
6
] the accelerometer and gyroscope
data had features such as the length of time for the gesture, their centroid, average
accelerometer data, etc. Being able to create an application-specific feature vector
allows you to tailor the classifier to selections between similar gestures. With good
training data, a classifier can be quite accurate and work well with raw data. However,
it does not performwell on streaming data, which requires identifying when a gesture
starts and stops (data segmentation) as well as performing recognition. It is also, in the
simplest form, limited to discrete actions. For example, it is capable of determining
“the user jumped” but not necessarily how high or where to.
