Online Recognition with Postural Transition Awareness - Smartphone-Based Human Activity Recognition - page 115

Smoothing Filter This filter targets the probability signals during the occurrence
of BAs. It helps to stabilize signal fluctuations when their probability values are
greater than a threshold (0.2) within the activity sequence. This is aimed to make
evident small differences between activities with high interclass misclassification
e.g. standing and sitting or walking and walking-upstairs. Oscillations are smoothed
using a linear interpolation.

7.3.3.2 Discrete filtering

The next step after the probability signals have been filtered is to define c ∗ ,themost
likely activity for each window sample. This is done using MAP over the probability
vector ( p

P ( s − 1 , : )

) extracted from the filtered activity matrix. From this one of
the classes is selected as the predicted activity.
However, under some circumstances, the entire probability vector contains small
values. This indicates that none of the classes seems to be representative of the current
activity. To this end, we have defined a minimum activity threshold which is used to
label samples as undefined (unknown-activity) when none of the classes reaches this
value. This is particularly useful during PTs in the PTA-6A method as they are not
learned by the SVM model. But in general, this approach can be beneficial in real
life situations when the HAR system is used while activities outside the studied set
occur. Consequently, these will not be categorized as any of them, instead the system
will show that an unknown event has occurred.
This filter removes sporadic activities that appear for a short time and are unlikely
to happen for only a window sample. It also includes cases when the unknown-
activity is detected and its contiguous activities belong to the same class. The filter
allows to relabel them as their neighbors. The final predicted activity is the result of
this discrete filter

=

c ˆ = ( z ) , where z in the buffer containing the last 3 predicted

activities c ∗ .

7.4 PTA-HAR Experiments

This section presents a collection of experiments carried out for the evaluation of our
PTA-HAR system. It starts with an evaluation of the

D 3 T dataset focused on PTs
which are analyzed in terms of duration and then compared against BAs. Moreover,
we review the system error metric proposed for this work which takes into account
the possible detection of the PT and unknown-activity classes. Finally, we describe
the main features of the smartphone application which has been developed for the
online recognition of activities.