State of the Art - Smartphone-Based Human Activity Recognition - page 39

thewavelet transforms are also applicable. Once obtained, they can be further reduced
using feature selection (e.g. exhaustive search, or wrappers, filters and embedded
methods (Guyon and Elisseeff 2003 )) and extraction approaches (e.g. using (PCA)
(Bishop 2006 )), or a combination of both. In Sect. 4.3.3 the features selected for the
development of our HAR dataset are presented. Moreover, in Sect. 6.4 we describe
the feature selection approaches used in this work.
3.2.4 Machine Learning
Several ML approaches have been developed throughout the years for HAR. It
has mostly been targeted through supervised learning algorithms although semi-
supervised and unsupervised methods have also been proposed (Stikic et al. 2008 ;
Wyattetal. 2005 ).
Frequentist and Bayesian models have been well covered throughout HAR
literature. They involve rule-based models such as DT and RF (Coley et al. 2005 ;
Ermes et al. 2008 ), geometric approaches including k-NN, ANN and SVMs (He and
Jin 2009 ; Khan et al. 2010a ; Maurer et al. 2006 ), and probabilistic classification
methods as for example NB classifiers, and Hidden Markov Models (HMM) (Tapia
et al. 2007 ; Zhu and Sheng 2009 ).
Many of these ML approaches have demonstrated comparable performance in
different works (e.g. (Mannini and Sabatini 2010 )) though suggesting that the effec-
tiveness and right selection of the algorithms can be linked to other aspects such as
data structure and application (Wolpert and Macready 1997 ). Other relevant aspects
for ML algorithm selection include: energy consumption, memory requirements,
interpretability and computational complexity, etc. As a matter of example, decision
trees could be preferred when the model interpretability is required and SVMs for
high performance applications. In Sect. 3.3.3 we provide more details regarding HAR
systems which have employed different ML algorithms.
3.3 Related Work in HAR Systems
Several approaches have been previously proposed in literature for the recognition
of human activities. They cover diverse application domains such as healthcare,
smart homes, UbiComp, AAL, surveillance and security (Cedras and Shah 1995 ;
Choudhury et al. 2008 ; Poppe 2010 ; Turaga et al. 2008 ). In general, these proposed
HAR systems are able to sense, monitor, and learn from our actions in order to
provide useful information which can be used to decide better about our future needs
or behavior (Cook and Das 2012 ).
These approaches can be categorized according tomany different criteria. Some of
the most relevant are: by sensor type , which is reliant on the class of signals measured
to extract activity information (e.g. inertial, vision-based and physiological (Lara and