State of the Art - Smartphone-Based Human Activity Recognition - page 40

Labrador 2012a )); by sensor location , which depends on the position of the sensor
with respect to the user. Namely external sensing when sensors are located in fixed
positions in the environment and wearable sensing when they are body-attached
(Yang and Yacoub 2006 ); by human activity type , which groups the systems with
respect to the activities they are able to identify (e.g. locomotion and ADLs); by
modeling principle , which can be data- or knowledge-driven depending on whether
the HAR models are built given pre-existing datasets or from the exploitation of
prior knowledge regarding a particular domain (Chen et al. 2012a , b ); by learning
approach , in relation to the type of algorithms used for learning such as supervised,
semi-supervised or unsupervised methods (Kwapisz et al. 2011 ; Stikic et al. 2011 ;
Wyattetal. 2005 ).
The work presented in Bao and Intille ( 2004 ) was pioneer in developing a method
for the detection of a set of activities of daily living using five body-worn accelerom-
eters and employing well-known ML classifiers (e.g. decision trees, naive Bayes,
and nearest neighbor). They developed a method for the detection of a large set (20)
of ADL. Their approach was performed offline and, considering the large amount of
activities included, it achieved promising results regarding the possibility of extract-
ing activity information fromaccelerometer signals. They also suggested the potential
advantages of developing online systems.
Moreover, other wearable systems have particularly grabbed the attention of the
HAR research community (Lee and Mase 2002 ; Lukowicz et al. 2004 ; Mantyjarvi
et al. 2001 ) due to the ease of obtaining activity information (e.g. body motion,
temperature and heart rate) directly from the user, unobtrusively and virtually at any
location without the need of fixed infrastructure. For example, (Ravi et al. 2005 ), as
opposed to (Bao and Intille 2004 ), used only one body-worn triaxial accelerometer
in the pelvic region, and evaluated a set of base-level and meta-level classifiers (e.g.
boosting and bagging) for improving the recognition performance. Furthermore, in
Lukowicz et al. ( 2004 ), an approach to recognized workshop activities was proposed.
It used accelerometers in combination with microphones strategically located on
the user's dominant arm. They analyzed the intensity of the acoustic signals and
correlated them with the inertial data in order to infer user activities. This approach,
which achieved accuracy levels of 84.4%, showed that the fusion of different sensors
can greatly improve the classification accuracy of their system.
More recently in Kwapisz et al. ( 2011 ) six human activities such as jogging,
walking and sitting were classified using a smartphone-embedded accelerometer
carried in the pocket in an attempt to simplify the recognition process with a more
pervasive, practical and unobtrusive approach. Other approaches have also been
proposed targeting specific applications: for example, from the medical standpoint,
monitoring systems have been presented for the detection of different attributes in
elder PD patients such as gait parameters, motion disorders and falls using on-body
accelerometer (Herrlich et al. 2011 ; Sama et al. 2012 ).
In Table 3.3 , a summary of the most influential offline HAR works is presented.
It highlights the most important aspects about them such as type of sensor used,
activities identified and system accuracy levels. In some of these works, system
accuracy was not provided, but instead, measures such as sensitivity and specificity.