Introduction - Smartphone-Based Human Activity Recognition - page 3

ing to fine tune the trade-off between dimensionality reduction and classification
accuracy. At the same time, we present a novel unified approach for training the
three possible instances of this algorithm: L1-SVM, L2-SVM, L1-L2-SVM. Its
significance relies on its flexibility and suitability to be implemented with well-
known solvers (Chap. 6 ) .

￿ We propose the first hardware-friendly SVM based on fixed-point arithmetic for

the prediction of human activities (MC-HF-SVM). This approach allows to vary
the fixed-point number representation (number of bits) to control over model accu-
racy and complexity, leading to improvements in terms of recognition accuracy,
speed and battery energy sparing with respect to conventional floating-point based
formulations (Chap. 5 ) .

￿ We have generated and made publicly available a HAR dataset (Reyes-Ortiz et al.

2013 ) composed of trials performed with a group of 30 participants which per-
formed a set of common daily activities while carrying a smartphone as a wearable
device. The dataset provides a collection of 10299 labeled activity instances which
include the raw inertial data from the smartphone accelerometer and gyroscope
along with the extracted activity features (Chap. 4 ) .

1.3 Thesis Outline

This thesis focuses on the design and implementation of smartphone-based HAR
systems. It has been divided in three main parts. Part I covers essential aspects on
the topic of HAR. First, a series of fundamental concepts required to contextualize
our research problem are portrayed. These are followed by an introduction to HAR
accompanied with its most influential works which are contrasted and discussed.
Part II concentrates on data collection and offline HAR. It presents the procedure
for the generation of the HAR dataset required for this research and also introduces
the first offline HAR method based on fixed-point arithmetic. Finally, Part III pays
particular attention in the description of the methods developed for the implemen-
tation of the online HAR systems and concludes with a summarization of all the
achievements of this work.
The remaining thesis chapters are briefly described here:

￿ Chapter 2 describes the main ideas about the areas of study relevant to the devel-

opment of HAR systems in order to develop a global perspective of our research
problem. These areas are divided in two groups: regarding our framework context
(Ambient Intelli-gence (AmI), Ambient Assisted Living (AAL)) and implemen-
tation mechanisms (sensors, smartphones and ML with emphasis on svms).

￿ Chapter 3 examines the current state of the art on the subject of HAR. It starts with

a general introduction regarding the HAR pipeline and then focuses on various
already implemented HAR systems relevant to our research. It also highlights
particular aspects of these systems such as sensing technologies, types of activities,
ML approaches and real-time computing.