We can know that
the signals from PIR sensors vary according to the orientation
of the sensing elements in PIR sensors. It is already known
that signal amplitude (the difference between the maximum
and minimum value of the PIR output) and passage duration
(the time during while PIR output exceeds two thresholds
above and below Vdd/2) can be used as a feature set to
classify the distance of the body from the PIR sensors [14].
They also used a peak detection method (one positive and one
negative) to recognize the direction of movement. However,
the feature sets are chosen only considering detecting two
opposite directional motions based on a single PIR sensor.
We have therefore decided to first use the time series captured
from the PIR array as a feature set for classifying walking
directions, and then deployed another reduced feature set
composed of peak values from PIR sensor signals. We will
demonstrate the effect of the trade-off between the amount
of computation load and recognition accuracy in the result
sections below.
C. Classifiers
Among various ava