Game Development Reference
In-Depth Information
•
The match score of feature vector sequence
given the model
O
=
O
O
...
O
1
2
T
(
m
=1,2,…,
M
) is calculated as follows:
λ
(
A
,
B
,
)
m
m
m
m
o
We compute the best state sequence
Q
given the observation se-
quence
O
, using Viterbi's algorithm, i.e.:
*
Q
=
arg
max
{
P
(
Q
/
O
,
λ
)}
(2)
m
Q
o
The match score of observation sequence
O
given the state sequence
Q
is the following quantity:
*
*
(3)
P
=
P
(
O
/
Q
,
λ
)
m
It should be mentioned here that the final block of the architecture corresponds
to a hard decision system, i.e., it selects the best-matched gesture class.
However, when gesture classification is used to support the facial expression
analysis process, the probabilities of the distinct HMMs should be used instead
(soft decision system). In this case, since the HMMs work independently, their
outputs do not sum up to one.
Figure 23. Block diagram of the HMM classifier.
Search WWH ::
Custom Search