Dynamic Time Warping-Based K-Means Clustering for Accelerometer-Based Handwriting Recognition - Developing Concepts in Applied Intelligence - page 24

Fig. 2. 26 capitalized English alphabet gestures are collected.

in 100hz. In the first phase, motion detection is performed for detecting the
duration in which a gesture is performed. The detection algorithm depends on the
variation of motion energy inside a sliding window. The details of the algorithm
is out of the scope of this paper. The output of the motion detection phase is
a segment of acceleration signals. It is then smoothed for noise reduction using
1-D Gaussian filter, resampled in 50hz for reducing the number of samples, and
finally normalized to the values between -5.0

5.0.
In our experiment, we used a 1-NN classifier with DTW as the distance mea-
sure. Training is in two-phases: pure instance-based learning in the first phase
and then DTW-based K-Means clustering for optimizing the reference patterns
in the second phase. As we have 5200 sample patterns, applying 10-fold cross
validation produces a 1-NN classifier with 4680 reference patterns and 520 test
patterns. Each class has 180 reference patterns after the first phase. We mea-
sured the performance of the classifier with all the 180 reference patterns, and
then we optimized the number of reference patterns by applying the DTW-based
clustering algorithm and compared the performance against the original classi-
fication performance. We tested with varying number of clusters and the result
is shown in table 1.

∼

Table 1. Recall and precision according to different number of clusters.

No. of clusters per class 10 20 30 40 50 180
Recall(%) 88.1 89.6 90.4 90.8 91.2 92.5

Precision(%) 88.1 89.7 90.5 90.9 91.2 92.6

The recall rate constantly decreases as the number of reference patterns is
reduced, but the degree of decrease is not so significant. With 10 reference pat-
terns per class, which is 5% of the non-clustered 180 reference patterns per class,
the recall rate stays only 4.8% less than that of the non-clustered case.
Table 2 shows how long it does take to classify a test pattern with different
number of reference patterns. It is noticeable that in the case of 10 reference
patterns per class, the classification took only 8% of the time for non-clustered
case with 180 reference patterns.