Information Technology Reference
In-Depth Information
to recognize or match to any of the learned activities based on probability estimates.
Dealing with them can help make systems more functional for a variety of applica-
tions where other activities are unaccounted for in the learning process. We interpret
UAs as an arbitrary subspace that contains unknown activities not learned by the ML
algorithm, similarly to the NULL class presented in (Bulling et al. 2014 ).
Following a similar framework to previous chapters, the proposed system exploits
a waist-mounted smartphone with embedded accelerometer and gyroscope. More-
over, it aims to provide near real-time activity prediction for monitoring applications
within the device or through wireless connectivity to others. This work is, to the best
of our knowledge, the first to evaluate the occurrence of PTs on online smartphone-
based HAR systems. The studied PTs in this work are: stand-to-sit , sit-to-stand ,
sit-to-lie , lie-to-sit , stand-to-lie and lie-to-stand .
The following sections are organized as follows: Sect. 7.2 gives an introduction
about postural transitions in HAR. This is followed by the description of the proposed
methodology chosen for the two HAR methods and the main stages of the online
system in Sect. 7.3 . In Sect. 7.4 , we present the experiments conducted regarding the
built HAR dataset, the selection of the system performance evaluation metric and
the developed smartphone app. Moreover, in Sect. 7.5 , we depict the achieved results
and the comparison against our previous system (L-HAR). Finally in Sect. 7.6 we
summarize the obtained results along with a critical view of the proposed approach.
7.2 Postural Transitions in HAR Systems
APT is an event with limited duration determined by its start and end times. Generally,
this duration varies among healthy individuals to some extent. PTs are bounded by
other two activities and represent the transition period between the two. Conversely,
BAs such as standing and walking can prolong for a longer term. Data collection for
these two types of activity is also different as PTs need to be executed repeatedly to
get separate samples and BAs, as they are continuous, allows many window samples
to be taken from a single test only limited by its time extent. Most of the proposed
HAR systems in the literature do not include PTs in the set of studied activities
(Sect. 3.3.1 ) . Some assume, for instance, that the detection of two different SPs
defines the occurrence of a PT (e.g. standing followed by sitting assumes a stand-to-
sit transition between them). In (Lara and Labrador 2012 ), it is pointed out that PTs
can be ignored in some situations if its occurrence is less frequent and duration is
much shorter against other activities.
However, as this assumption is application specific, PTs can directly influence
the system performance in the other cases. Therefore, they should be taken into
consideration during the design and selection of the right recognition algorithms.
HAR systems such as the one presented in (Khan et al. 2010 ), combine 7 BAs with
7 PTs in the classification. In cases like this, however, large multiclass problems
may arise which could eventually increase the prediction of false negatives of the
activities we are particularly interested in (e.g. BAs).
 
Search WWH ::




Custom Search