State of the Art - Smartphone-Based Human Activity Recognition - page 43

For comparative purposes, we assumed in these cases a balanced dataset with the
same number of positive and negative samples in order to approximate thismeasure as
accuracy

1

. In the same manner, we present in Table 3.4
the most relevant online HAR systems. In the tables, it can be noticed the large
diversity of sensors, types of activities, number of subjects in the experiment and
learning approaches that has been employed to achieve the recognition of activities.
These differences clearly produce variations in the algorithms performance. For
example, the number of people involved in an experiment is an important factor during
learning: the higher this number is, the more realistic representation of the overall
population (or target group) is obtained. This can therefore increase the generalization
ability of a recognition system in order to correctly classify the activities of new
unseen persons (e.g. the work in Lukowicz et al. ( 2004 ) performed experiments with
a single individual. Even though it achieved a high accuracy it is possible that their
system performance would degrade when tried on new people). Likewise, the system
accuracy decreases when more activities are added into the system for classification.
The HAR systems proposed in this thesis are included at the bottom of the
Tables 3.3 and 3.4 . They are identified by their corresponding system name. We use
smartphones as a wearable device located on the waist which contains two inertial
sensors (the accelerometer and gyroscope) for the extraction of activity information.
With regard to human activities, we classify 6 different BAs and also consider the
effects 6 PTs. Our learning approach is based on supervised SVMs which is a data-
driven model as it is built from the HAR dataset. Lastly, we explore both online and
offline approaches for the recognition of activities.
Various surveys regarding HAR systems have been presented in the literature
covering general approaches (Chen et al. 2012b ), or more specific ones such as
focusing on wearable sensors (Lara and Labrador 2012a ), on-body accelerometers
(Mannini and Sabatini 2010 ), smart environments (Cook and Das 2007 ). In the
following subsections, we focus on some of these HAR systems attributes in order
to analyze them independently against our proposed approaches.

=

2 (

sensitivity + specificity )

3.3.1 Human Activity Type

Human activities can be categorized based on complexity and area of application
as described in Sect. 3.2.1 . Related HAR works have directed efforts in these two
directions. Most of them, however, have focused on the study of locomotion activities
as it can be seen in Tables 3.3 and 3.4 . Within this group, we have found that static
postures such as sitting and standing are commonly investigated along with some
dynamic activities such as walking and climbing stairs. In Karantonis et al. ( 2006 ),
for instance, the user posture is detected using a waist-mounted accelerometer and
it is then used for early stage decisions in a classifier of dynamic movements (e.g.
walking, running, cycling). Moreover, (Allen et al. 2006 ) developed a classifier which
combines 3 static postures and 5 movements using a rule-based heuristic system and
a Gaussian Mixture Models (GMM).