Online Recognition with Postural Transition Awareness - Smartphone-Based Human Activity Recognition - page 121

Table 7.3 PTA-6A confusion matrices. Before and after filtering
Method

PTA-6A

Activity WK WU WD SI ST LD PT

WK 542 0 3 1 0 0 0

WU 32 523 2 0 2 0 0

WD 3 4 498 0 4 0 0

SI 0 4 0 481 71 1 0

ST 3 3 0 18 588 0 0

LD 0 0 0 0 0 604 0

PT 12 101 1 18 4 0 193

Method PTA-6A

Activity WK WU WD SI ST LD PT UA

WK 545 0 0 0 0 0 0 1

WU 28 514 1 0 1 0 0 15

WD 0 0 505 0 3 0 0 1

SI 0 0 0 504 52 0 0 1

ST 0 1 0 1 610 0 0 0

LD 0 0 0 0 0 604 0 0

PT 0 10 0 9 0 0 310 0

Note The bold diagonal highlights the most important part of the confusion matrix.

indicating that the system was incorrectly predicting them during the occurrence of
PTs. Therefore the temporal activityfilters have helped tominimize this error. It is also
noticeable the reduction of interclass misclassifications between similar activities
such as in the static postures sitting and standing, and also between walking and
walking-upstairs. After filtering, the false negatives of the standing class produced
by sitting samples become nearly zero (from 18 to 1), however, the opposite case,
which has a reduction of 26% in the number of misclassification, still preserve some
errors.
Table 7.4 shows the results for the PTA-7A method. The addition of the PT class
into the learned SVM model shows how most of the PT samples are correctly clas-
sified even before applying temporal activity filtering as opposed to the previous
method which had a large number of false negatives of the PT class. Moreover,
after filtering, small improvements are still evident: for instance, the number of false
negatives for the PT class is further reduced (from 7 to 2), and also the interclass
misclassification of static postures is also diminished. The SVM output is, however,
generating some false positives of the PT class. This means that some actual BAs are
being confused with PTs. Therefore, the addition of an additional activity (PT) into
the model is producing an increase on the BAs error which can be unfavorable in
applications where the occurrence of BAs is larger than PTs. This explains why the
BAs error with the PTA-7A method is higher than with the PTA-6A. Notice also that
in our model, transitions are considered as a single class but for some other applica-
tions it might be needed to learn them separately. This can decrease even further the
recognition performance of BAs as the number of classes of the classifier increases.