Linear SVM Models for Online Activity Recognition - Smartphone-Based Human Activity Recognition - page 97

Table 6.1 Confusion matrices for the MC-GK-SVM algorithm with D 2 (top), MC-L2-SVM) with
D 2 (center), and with only time domain features D 2 T (bottom)
Activity

WK WU WD SI ST LD Sensitivity (%) Specificity (%)

MC-GK-SVM— D 2
WK 486 6 4 0 0 0 97.98 99.31

WU 12 458 1 0 0 0 97.24 98.59

WD 5 27 388 0 0 0 92.38 99.80

SI 0 2 0 450 39 0 91.65 99.71

ST 0 0 0 7 525 0 98.68 98.39

LD 0 0 0 0 0 537 100.00 100.00

Accuracy 96.50

MC-L2-SVM— D 2
WK 493 1 2 0 0 0 99.40 99.10

WU 20 450 1 0 0 0 95.54 99.56

WD 2 6 412 0 0 0 98.10 99.88

SI 0 4 0 435 51 1 88.59 99.43

ST 0 0 0 14 518 0 97.37 97.89

LD 0 0 0 0 0 537 100.00 99.96

Accuracy 96.54

MC-L2-SVM— D 2 T
WK 494 2 0 0 0 0 99.60 98.82
WU 20 451 0 0 0 0 95.75 99.80
WD 8 0 412 0 0 0 98.10 100.00
SI 0 3 0 428 60 0 87.17 99.39
ST 1 0 0 15 516 0 96.99 97.52
LD 0 0 0 0 0 537 100.00 100.00
Accuracy 96.30
Note The bold diagonal highlights the most important part of the confusion matrix.

6.4.2 Selection of Subsets of Features

The second set of experiments consisted of evaluating the available features in the
dataset aiming at a significant reduction in their number. This was firstly achieved
by separating the inputs in groups with respect to: (i) the type of sensor employed,
namely Accelerometer ( A ) and Gyroscope ( G ); and (ii) the feature domain, namely
Time ( T ) and Frequency ( F ). The analysis aims to gather some evidence of the benefits
the addition of gyroscope signals bring into the HAR system, and also assesses the
need of frequency domain features for improving the recognition performance.
In practice, we expect to balance the trade-off between the addition of meaning-
ful features and the removal of the ones that are redundant or that require expen-
sive computations for their estimation. For such purposes, we tested the possible
combinations of feature groups and computed the system accuracy performed by a