State of the Art - Smartphone-Based Human Activity Recognition - page 45

Considering the classification of activities based on complexity, most of the works
have been concentrated on short events and basic activities. For example, in Ravi et al.
( 2005 ), 8 basic activities related to locomotion and ADL are classified using four dif-
ferent learning algorithms. Additionally, transitions and basic activities are classified
simultaneously in Najafi et al. ( 2003 ). Not much work has been done on complex
activities or activity sequences. Primarily because of the difficulty of developing
automatic sensor signal segmentation into activities and the high computational cost
involved which takes into account past events. In Lukowicz et al. ( 2004 ), a method
for continuous recognition of activities based on large decision windows after small
window classification is proposed. It is a first step in the study of more complex sets
of activities. They apply Linear Discriminant Analysis (LDA) and HMM over audio
and inertial signals for developing their classification model.
Some systems have focused on the detection of specific events that occur to people.
For example, in healthcare, falls are commonly studied activities. Their detection is
essential in order to assist people with any type of limitation such as the elderly (e.g.
(Li et al. 2009 )). Other activities are directly related to particular illnesses. In the case
of PD patients, symptoms such as bradykinesia, freezing of gait, and on/off states are
also relevant and therefore their detection have been previously investigated such as
in Sama et al. ( 2012 ), Bachlin et al. ( 2010 ) and Takaˇcetal.( 2013 ).
Most of the HAR systems do not consider PTs as part of their activity set, even
if these frequently occur during the transitions from one activity to another (Refer
to Sect. 7.1 for further information). Among the works which have studied PTs we
have the following: (Khan et al. 2010b ), which chose 3 SPs, 4 AAs and 7 PTs for
their study. Furthermore in Najafi et al. ( 2003 ), a comprehensive study of signals
occurring during PTs for healthy and elderly people is used for the classification
of activities. In Salarian et al. ( 2007 ), the detection of sit-to-stand and stand-to-sit
transitions were crucial for the distinction between standing and sitting. This was
achieved through a fuzzy logic classifier which required past and future transition
information for this task. In Rodríguez-Martín et al. ( 2013 ), a hierarchical structure
of classifiers was employed to distinguish stand-to-sit and sit-to-stand PTs in patients
with PD through measurements from a triaxial accelerometer located on the waist.
Lastly, in the medical field, the detection of PTs is sometimes necessary such as in
the work of (Mellone et al. 2012 ) where the Timed Up and Go test, used for the
assessment of balance and mobility in patients with motor problems, is automated
using smartphones. The objective measure of this time required the detection of
stand-to-sit and sit-to-stand transitions. In this thesis we first explore methods for
the detection of 6 BAs linked to locomotion, and then we also propose an approach
which considers PTs as their detection can also enhance the correct prediction of the
first set of activities.