Hardware-Friendly Activity Recognition with Fixed-Point Arithmetic - Smartphone-Based Human Activity Recognition - page 81

This result is similar to the one presented in (Neven et al. 2008 ). The important
outcome of this section is that the number of bits in the HF-SVM has a strong
regularization effect with an impact on the generalization ability of the classifier.
Between two classifiers with approximately the same performance, we have to choose
the one that can be representedwith less number of bits since it ismore energy efficient
and it has more capacity of performing well on previously unseen data.
Finally we want to highlight that the bound in Eq. ( 5.12 ) is very loose since it is
data independent and does not take into account the quality of the available samples
for its estimation. In the last few years, proposed data dependent bounds (Bartlett
and Mendelson 2003 ; Bartlett et al. 2005 ) are becoming tighter and providing better
interpretation of the generalization ability of classifiers. They have shown to work
well on the performance estimation of real world problems such as in (Anguita
et al. 2012 ). For these reason, the understanding of the influence of fixed-point
arithmetic approaches in the estimation of these bounds is an interesting topic of
research.

5.4 Results

The performance of the MC-HF-SVM was evaluated through a collection of exper-
iments using the HAR dataset D 1 described in Sect. 4.3.3 . The data samples were

composed of 17 features extracted from the smartphone triaxial accelerometer in the
time and frequency domain. These features had been previously suggested in the
HAR literature (Bao and Intille 2004 ; Lovell et al. 2007 ; Sama et al. 2010 ) including
measures such as SMA, mean, Standard Deviation (STD), entropy and signal-pair
correlation, when applied to the available processed inertial signals, provided the set
of features depicted in Table 5.1 .
The method evaluation is divided in two parts. The first one analyzes the sys-
tem recognition performance against a traditional approach (Multiclass LK-SVM
(MC-LK-SVM)). Then, in the second part, we focus on evaluating other algorithm
attributes more related with its use in hardware devices. These are the recognition
speed and battery discharge.

Table 5.1 List of measures for computing feature vectors
Feature vector
Measure

Applied to

, a ˄ ( , a (

SMA a ˄ ( t ) t ) , a ( t ) t )

Mean a ˄ , 1 ( t ) , a ˄ , 2 ( t ) , a ˄ , 3 ( t )

STD a ˄ , 1 ( t ) , a ˄ , 2 ( t ) , a ˄ , 3 ( t )

Correlation a ˄ , 1 ( t ) − a ˄ , 2 ( t ) , a ˄ , 1 ( t ) − a ˄ , 3 ( t ) , a ˄ , 2 ( st ) − a ˄ , 3 ( t )

a

Entropy A ˄ , 1 (ˉ) , A ˄ , 2 (ˉ) , A ˄ , 3 (ˉ) , A mag (ˉ)

a Capitalized letters represent the signal in the frequency domain

ˉ