Introduction - Smartphone-Based Human Activity Recognition - page 4

￿

Chapter 4 presents the collection of stages required for the acquisition of the
experimental HAR data used in this thesis. It includes aspects such as smart-
phone selection, trials with volunteers, signal conditioning and feature selection. It
also describes the procedures concerning the dataset validation; internally through
experimentation and externally via a HAR contest in which other research groups
were encouraged to propose novel solutions to the recognition problem.

￿ Chapter 5 explains the proposed hardware-friendlyHAR system (HF-HAR) and its

core ML algorithm based on fixed point arithmetic (MC-HF-SVM). Initially moti-
vated by current limitations of mobile devices regarding battery life, this method
aims to predict activities relying on a modified SVM formulation with adjustable
fixed-point number representation. In this way, we show through experiments
how system accuracy levels can be maintained while increasing the speed during
the prediction of activities. This is explained under the light of Statistical Learn-
ing Theory (SLT) which shows how this implementation brings advantages with
respect to the generalization ability of the algorithm.

￿ Chapter 6 studies linear SVM algorithms and its application to an online system

for the recognition of activities on smartphones (L-HAR). The algorithms differ
on the norm of their formulation's regularization term (whether it is the L1-,
L2- or L1-L2-Norm). They allow to control over dimensionality reduction and
classification accuracy while increasing the prediction speed when compared with
kernelized SVM algorithms. Moreover, this chapter presents a novel approach for
training these classifiers (EX-SMO) withminimal effort usingwell-known solvers.
To conclude, the benefits of adding smartphones gyroscope into the recognition
system are presented along with another feature selection mechanisms that use
subsets of features in the time and frequency domain.

￿ Chapter 7 introduces an online HAR system for the classification of human activ-

ities using smartphones which deals with recurring postural transitions while
sequences of activities are carried out (PTA-HAR). For its implementation, the
linear SVMs presented in Chap. 6 are combined with temporal filters that use
activity probability estimations within a limited time window. The benefits of
these approaches are presented through experimentation with the HAR dataset
and compared against the previously presented HAR systems.

￿ Chapter 8 summarizes the accomplishments of this work and also proposes future

research directions

References

A.T. Campbell, S.B. Eisenman, N.D. Lane, E. Miluzzo, R.A. Peterson, H. Lu, X. Zheng,
M. Musolesi, K. Fodor, G.-S. Ahn, The rise of people-centric sensing. IEEE Internet Comput.
12 , 12-21 (2008)
L. Chen, J. Hoey, C.D. Nugent, D.J. Cook, Z. Yu, Sensor-based activity recognition. IEEE Transac-
tions on Systems, Man, and Cybernetics, Part C: Applications and Reviews 42 , 790-808 (2012)