Linear SVM Models for Online Activity Recognition - Smartphone-Based Human Activity Recognition - page 89

Chapter 6
Linear SVM Models for Online Activity
Recognition
6.1 Introduction
The exploitation of smartphones for HAR has been an active research area in which
the development of fast and efficient ML approaches is crucial for providing real-
time performance and preserving the device's battery life. In this chapter, we focus
on the exploration of three linear SVM algorithms for performing online HAR.
Their formulation differs only in the regularization term. It uses different norms,
namely L1, L2 and L1-L2, and gives the SVM different properties and behavior.
Specifically, these differences affect the trade-off between dimensionality reduction
and classification accuracy. Using the proposed learned linear models we implement
our first online HAR system: Linear HAR system (L-HAR). It provides in real-time
the classification of activities using a smartphone device using features extracted
from its embedded accelerometer and gyroscope.
Moreover, we present a novel algorithm for training L1-L2-Norm SVM (L1-L2-
SVM) classifiers. The proposed training approach allows the exploitation of all the
well-known effective and reliable tools (e.g. Quadratic Programming (QP) solvers)
already developed for solving the conventional L2-Norm SVM (L2-SVM), thus min-
imal effort is required by the user to implement L1-L2 model training. The proposed
method is flexible, as it also allows to train L1-Norm SVM (L1-SVM), L2-SVM.
The effectiveness of this approach is tested on our HAR dataset.
We also show the benefits of adding smartphones gyroscope signals into the
recognition system against the common approach which only uses accelerometer
data. MEMS gyroscopes made its entrance in the smartphones market a couple of
years after accelerometers and they have not been fully explored (Wu et al. 2012 ).
We also study two feature selection mechanisms for allowing a faster recognition:
the utilization of exclusively time domain features from the inertial data and the
adaptation of the L1-L2-SVM which controls over the number of non-informative
features as part of its model construction process.
This chapter is distributed in the following way: first we present the standard
algorithms based on L1 and L2 norms (Sect. 6.2 ) and show how they are adjusted in