Online Recognition with Postural Transition Awareness - Smartphone-Based Human Activity Recognition - page 123

7.6 Summary

In this chapter, we presented a fully operational HAR system for the recognition of
activities using smartphones (PTA-HAR). We showed through two different methods
how to achieve online recognition while taking into account the effect of postural
transitions in the overall system classification performance. Although, these transi-
tory events are usually disregarded in most applications, they become relevant when
their incidence is high (e.g. sports activities and housekeeping), as well as the need
to explore and evaluate them.
The two proposed methods (PTA-6A and PTA-7A) exploit correlation between
contiguous activities to reduce misclassifications, in particular during transitory
events. This is achieved through the filtering of activities which are interpreted as
probability signals that change over time.
Results have shown that the two proposed methods of the PTA-HAR system have
similar classification performance and are therefore suitable options for HAR appli-
cations. They have also confirmed to be more appropriate for online classification
than L-HAR which does not consider activity temporal filtering. Moreover, here
we provide some considerations in order to guide the selection of the most suitable
method depending on its application:

• When the precise detection of PTs is required, the PTA-7A approach is the only

one that can be applied, though most HAR applications are only concerned about
the detection of BAs. The PTA-6Amethod instead avoids learning these events but
still prevents problems that could arise in the presence of PTs during classification.

• The method selection is also a trade-off between having a simpler learning algo-

rithm and system accuracy. The PTA-6A method is in particular easier to imple-
ment as the learning stage does not require PTs. For example, in applications with
a larger number of activities (e.g. by adding activities such as bent, reclined, lying
down facing up/down, etc.), the recording of transitions between basic activities
becomes more complex as the number of possible PTs increases quadratically in
a proportion of υ(υ − 1 ) , where υ is the number of studied BAs ( υ = 3 SPs in our

case). This, therefore affects the learning on the PTA-7A method.

• The method selection can be also associated with the number of PTs that occur

with respect to the time extent of other activities. If they do not occur too often or
the time between PTs is rather large, then we can be less rigorous about learning
transitions and use the PTA-6A approach.

• Some datasets only include information about BAs and do not have transitions

labeled. This limits the study to use only the PTA-6A method.

The adaptation of the unknown-activity class into the system was useful for the
performance of the PTA-6A approach as PTs were detected as unfamiliar events.
Furthermore, this concept can be extended to real life applications such as in the
monitoring of activities where there are chances to perform different activities that
are not known in advance. It is preferable to have a system that notifies that an activity
seems unknown rather than always classifying it only as one of the studied group of