Online Recognition with Postural Transition Awareness - Smartphone-Based Human Activity Recognition - page 109

Online smartphone-based HAR systems have not yet studied PTs along with other
activities. However, in cases such as in (Zhang et al. 2010 ) an offline HAR system
combines various PTs as a single class along with other activities for a daily monitor-
ing application. Moreover, in (Rednic et al. 2013 ) an approach for performing posture
classification is proposed. Even though this is done using a multi-accelerometer BSN
instead of smartphones, they also investigate the effects of PTs in their system and
introduce exponentially weighted voting transition filters in order to improve by 1%
the accuracy of their recognition system. They explore 7 activities usually performed
during explosive ordnance disposal operations (e.g. kneeling, crawling and sitting).
In this work, we present, in contrast, a method that aims for its real-time execution
on a single smartphone and deals with activity information in the vicinity of each
event occurrence.

7.3 HAR with Postural Transitions Awareness

This section presents the two proposed posture-aware methods that deal with the
occurrence of PTs in an online activity recognition system. Their difference relies
on the way they handle PTs and the number of activities that are learned in the ML
algorithm. These are:

• PTA-6A : This first method, as its name designates, only takes into account the

6 studied BAs (standing, sitting, lying-down, walking, walking-downstairs and
walking-upstairs) for the ML learning process with SVMs, while PTs are initially
disregarded. Following the ML algorithm, a temporal filtering stage is introduced.
It deals with PTs and misclassifications of BAs based on the classifier output of
contiguous window samples.

• PTA-7A : The secondmethod takes into account seven classes in theML algorithm:

the original 6 BAs plus an additional class which represents all the 6 PTs at
once. The PTs are: StSi, SiSt, SiLi, LiSi, StLi, and LiSt (Refer to Fig. 7.1 for an
illustration). Similarly to the previous method, temporal filtering is also applied
after prediction, however, PTs are handled differently provided that they are one
of the ML algorithm possible outputs.

The entire recognition algorithm is composed of three main stages which are
depicted in Fig. 7.2 . Moreover, a pseudocode of the entire recognition process is
presented in Algorithm 3.

Fig. 7.1 Postural transitions from the 3 studied static postures: standing, sitting and lying-down