Introduction - Smartphone-Based Human Activity Recognition - page 2

In this thesis, we explore the feasibility of using smartphones to perform the
automatic recognition of physical activities while also addressing some of these
current HAR limitations. These modern devices, which are a new generation of
mobile phones, are provided with enhanced computing capabilities and embedded
sensors that make them highly suitable for their application in HAR. We aim to
provide a technological tool that can be employed as an activity information resource
to other systems for better decision making (e.g. for the remote monitoring of elderly
patients who live alone or without permanent caretaking and themeasurement of their
personal autonomy).
The development of a smartphone-based HAR system introduces new challenges
linked to the incorporation of the recognition system on everyday use devices. In
this work, we propose various approaches in order to recognize, in real-time, human
activity using the smartphone inertial sensors ( accelerometer and gyroscope ) in con-
junction with the implementation of supervised Machine Learning (ML) algorithms
(specifically Support Vector Machine (SVM)) on on the device. We focus particularly
on the study of a group of six locomotion BAs: standing , sitting , lying-down , walk-
ing , walking downstairs and walking upstairs ; and also examine six PTs that occur
between its static postures: stand-to-sit , sit-to-stand , sit-to-lie , lie-to-sit , stand-to-lie
and lie-to-stand . Sometimes these PTs are disregarded in HAR systems (Lara and
Labrador 2012 ). But they are significant for certain applications where their inci-
dence is high and overall duration is comparable to other activities. For example,
in the design of activity monitoring systems for the disabled during rehabilitation
practices. The study of these PTs is also a central element of our research.
This thesis addresses the following questions: (i) How can we performHAR using
existing smartphones? (ii) How to exploit inertial sensors (accelerometer and gyro-
scope) to develop smartphone-based HAR systems? (iii) Which ML algorithms are
suitable for efficient HAR implementations in battery-limited smartphones? (iv) How
to achieve real-time HAR using smartphones?

1.2 Main Contributions

The main contributions of this thesis are presented as follows:

￿ We present a smartphone-based HAR system for the online recognition of human

activities from inertial data (PTS-HAR). It consists of the combination of four
elements: the device-embeddedmotion sensors, a signal processing unit for feature
extraction, a multiclass linear SVM algorithm and an activity filtering module for
dealing with recurrent PTs. We demonstrate the system operation in real-time
and show the improvements achieved when considering the detection of postural
transitions (Chap. 7 ) .

￿ We propose an ML algorithm for activity classification based on a one-vs-all

(OVA) SVMwith L1- and L2-Norm regularization. Its advantage relies on its faster
prediction when compared against non-linear SVM approaches while also allow-