Information Technology Reference
In-Depth Information
work proposed in Nham et al. ( 2008 ), they are able to perform offline classifica-
tion of four modes of transportation (biking, running, walking and driving) using
an iPhone accelerometer. Similarly, in this thesis we propose the HF-HAR system
which is able to perform offline the recognition of 6 activities from data gathered
from smartphones using a hardware-friendly approach (Chap. 5 ) . The ML method
uses a modified multiclass SVM with fixed-point arithmetic prediction aiming to
obtain a fast implementation more suitable for battery operated devices.
More recently, other contributions have proposed online smartphone-based HAR
systems. A Nokia smartphone was used in Brezmes et al. ( 2009 ) for the online
recognition of 6 activities. In the same way, the work presented in Kose et al. ( 2012 )
used an Android OS smartphone with embedded accelerometer for the classification
of 4 activities. The training stage was also performed online through the option of
collecting live data from the users by following a predefined activity protocol. Lara
and Labrador ( 2012b ) proposed an improved version of the work in Lara et al. ( 2012 )
for the recognition of activities in real-time. To conclude (Riboni and Bettini 2011 )
proposed a HAR systemwhich combined the smartphone internal accelerometer with
an external one on the user's wrist for the classification of ADL. They proposed a
context-aware framework based on ontologies, reasoning and statistical inferencing
to solve scalability problems when the number of activities becomes large.
3.4 Summary
In this chapter we discussed essential aspects regarding the recognition of human
activities in the light of the existing research literature. It included common strategies
for the development and evaluation of HAR systems. Additionally, relevant up-to-
date HAR approaches are highlighted and compared against the methods proposed
in this thesis.
There are already several alternatives that deal with the problem of recognizing
activities which have similarities between them in various aspects (e.g. the types and
number of activities identified, sensors used, ML approaches and real-time predic-
tion). Nevertheless, we explore in this thesis some unique aspects in this field which
have not yet been fully covered in the literature. They include elements such as the
evaluation of hardware-friendly approaches for the recognition of activities using
fixed-point arithmetic, the study of hybrid linear SVM models for their application
in HAR, and the awareness of postural transitions in real-time recognition systems.
These will be thoroughly covered in the forthcoming chapters.
 
Search WWH ::




Custom Search