Information Technology Reference
In-Depth Information
(2007) (Lane et al. 2010 ). Gyroscopes, on the other hand, made a late appearance a
few years later (2010), though less approaches have considered them for HAR (e.g.
(Altun and Barshan 2010 ; Lee and Mase 2002 )). Recently, in Wu et al. ( 2012 ), a
hybrid accelerometer and gyroscope approach was used for the classification of 9
activities using an iPhone 4. They showed insights of the benefits of adding gyroscope
signals into the recognition system achieving improvements ranging from 3
.
1to
13
.
4% in classification accuracy.
3.3.3 Machine Learning
Machine Learning approaches that have already been applied for the recognition
of activities include: NB (Jatoba et al. 2008 ), HMM (Mannini and Sabatini 2010 ),
DT, and SVM (Maurer et al. 2006 ). Some HAR works have also compared various
ML methods to find the most suitable approach in their applications. However, they
do not coincide with a specific best solution and have found instead heterogeneous
approaches (e.g. DT, GMM, k-NN, etc.). This finding shows a possible dependency
of the algorithm selection on the type of application and data used. Moreover, meta-
classifiers, which predict based on the output of a series of base-level classifiers, has
also been used in HAR systems. In Ravi et al. ( 2005 ), for instance, plurality voting
is preferred as the best approach for the classification of activities from the output of
standard classifiers including SVM, NB, k-NN and DT. The use of these classifiers
is therefore interesting in order to improve the overall recognition performance of a
HAR system at the expense of carrying out more complex operations which can lead
to limitations with respect to battery consumption, shared resources, and real-time
operation.
Our approach exploits SVMs for the classification of activities similarly to other
works which have successfully employed them (He and Jin 2009 ; Khan et al. 2010a ;
Maurer et al. 2006 ). Furthermore, SVMs have shown to be effective in heteroge-
neous types of recognition such as in handwritten characters (LeCun et al. 1995 ) and
speech (Ganapathiraju et al. 2004 ). In this thesis, we show in Sect. 4.4.2 that they
outperform other ML approaches when our HAR dataset is used. SVMs provide a
good compromise between accuracy and training time while also count with a vari-
ety of publicly available learning tools for experimentation such as LIBSVM (Chang
and Lin 2011 ). However, although characterized by several appealing attributes, one
of its drawbacks consists in its naïve two-class nature, that makes generalization
to multiclass problems (as in the typical case of HAR) not straightforward. Dif-
ferent approaches have been explored for targeting this issue (LeCun et al. 1995 ).
The two most commonly used methods are: OVA and one-vs-one (OVO) as seen in
Sect. 2.5.3.2 . We selected, in particular, the OVA method for our research.
Adaptive HAR systems have also been explored in the literature. Their purpose is
to take classification algorithms and adapt them to a particular user in order to person-
alize it to the user characteristics making it more robust and accurate. Although, in
this thesis we do not performadaptation, some approaches that have already proposed
 
Search WWH ::




Custom Search