Information Technology Reference
In-Depth Information
Algorithm 1
. DTW-MeanForNPatterns(
n, P, N, M
)
M ← P
0
for
i
=2
to
N
do
W ←
WarpingPathByDTW(
n,M, |M|,P
i
, |P
i
|
)
for
j
=1
to
|M|
do
Q ←{P
i,W
x,
1
|W
x,
0
=
j,
1
≤ x ≤|W|}
|
Q
|
k
=1
Q
k
M
j
← M
j
+
1
|
Q
|
end for
end for
M ←
N
M
The key to the algorithm 2 is the
for
-loop where input vectors are incre-
mentally added along the warping path
W
. After adding all the
N
patterns, the
centroid pattern is calculated simply by dividing
. With the algorithm 2,
it is possible to apply K-Means clustering on non-uniform length patterns to op-
timize the reference patterns for each gesture class.
M
by
N
3 Experimental Result
We evaluated the DTW-based K-Means clustering algorithm by applying it on
a problem of recognizing gesture patterns.
We used Nintendo
TM
Wii remote to collect gesture pattern data. Wii remote is
a handheld device that embeds a 3-axis accelerometer. Users performed gestures
while holding the device by hand. The embedded accelerometer captures the
hand movements as a series of acceleration changes across the 3-axes of (
).
The target patterns for this experiment are uppercase English alphabets, each
stroke of which is defined as in Figure 1. 40 adult people participated to perform
gestures. Each person played every alphabet 5 times, so each alphabet has 200
pattern samples. The total number of patterns for all the 26 alphabets are 5200.
The signal preprocessing stage for feature extraction is shown in Figure 2.
The input to the preprocessing is a train of 3-axis acceleration signal sampled
x, y, z
1
2
1
2
1
3
3
2
1
2
3
2
2
1
2
2
2
2
2
1
1
3
1
3
1
3
1
2
1
1
1
2
1
1
2
2
2
2
2
1
1
2
2
3
Fig. 1.
26 capitalized English alphabet gestures are collected.
Search WWH ::
Custom Search