Hardware-Friendly Activity Recognition with Fixed-Point Arithmetic - Smartphone-Based Human Activity Recognition - page 75

Chapter 5
Hardware-Friendly Activity Recognition
with Fixed-Point Arithmetic
5.1 Introduction
Exploiting SVMmodels for HAR on smartphones requires a multitude of operations
to be carried out per second: despite not being an issue from a theoretical point
of view, this could lead to battery discharge after few hours of continuous opera-
tion, making this approach unfeasible to allow people's mobility. In this chapter,
we explore a fixed-point arithmetic based reformulation of the conventional SVM,
targeted towards multiclass classification. Up to date we have no knowledge of other
research works that have incorporated fixed-point arithmetic into the learning algo-
rithms for the classification of human activities. However, extensive research on
fixed-point arithmetic has been developed to integrate ML models on hardware with
limited resources (e.g. Wawrzynek et al. 1993 ). This idea was initially motivated
some years ago because the assemble of devices with floating-point units was infea-
sible. Moreover, limited devices are usually preferred for specific-purpose applica-
tions if they demonstrate similar performance to traditional processing units as their
production (and/or acquisition) costs are generally lower. Nowadays, it has become
particularly interesting to retake these approaches and apply them in the development
of software applications for portable devices such as smartphones which are highly
demanding in terms of energy consumption and system resources management.
The term Hardware-Friendly SVM (HF-SVM) was first presented in (Anguita
et al. 2007 ). This method was designed for binary classification problems by emp-
loying fixed-point arithmetic in the FFP of the SVM classifier, with the purpose of
allowing its use in hardware-limited devices. In thiswork, we adapt themodel to target
HAR on smartphones through a modified multiclass HF-SVM learning algorithm
MultiClass HF-SVM (MC-HF-SVM). It aims to provide faster predictions and better
preserve the battery lifetime of these portable devices with respect to conventional
floating-point formulations while maintaining comparable system accuracy levels.
The introduction of this ML algorithm for HAR gives origin to the HF-HAR system
which is described throughout this chapter.
The organization of this chapter is as follows: Sect. 5.2 describes the methodology
of the hardware-friendly approach including its mathematical formulation from the