Online Recognition with Postural Transition Awareness - Smartphone-Based Human Activity Recognition - page 120

Table 7.2 System error
basedonfilteringstageand
type of activity

Filter BAs (%) PTs (%) Overall (%)

PTA-6A
No filtering

4.46 41.34 7.72

Probability 3.45 18.24 4.76

Discrete 3.10 5.77 3.34

PTA-7A
No filtering 4.60 2.12 4.39

Probability 3.63 0.61 3.36

Discrete 3.51 0.61 3.25

is possible to have an idea how the different stages of the algorithm are progressively
affecting the overall classification performance. The three stages are: no filtering
(SVM output), probability filtering, and discrete filtering. On the table, every row
represents these stages of the algorithm and the columns each HAR method.
From the table, it can be also noticed that the error in the PTA-6A method without
filtering is the highest achieved (7.72%). As we decompose this, we can see that
this is mainly due to a large error of 41.34% in the classification of PTs. BAs error
instead remains much lower with a 4.46%. We can also observe that the temporal
activity filters widely improve the classification of PTs reaching a minimum error
of 5.77%. BAs instead improve only slightly after filtering. The final error of the
PTA-6A method is 3.34% which nearly matches the one achieved with L-HAR and
D 2 T that did not take into account PTs.
Alternatively, the PTA-7A method presents a different behavior. As PTs are
learned, its recognition error is much lower from the first beginning when they
are classified by the SVM (2.12%) instead of the value obtained with the previ-
ous method which was much larger (41.34%). However, the classification error of
BAs is always slightly higher when compared to the previous method. Primarily
because the addition of the extra class in the learning stage causes some BAs to get
misclassified as PTs as it would be expected. The temporal filtering does help to
improve all BAs but its effect is minor. This method is showing that the learning of
postural transitions can be helpful to the classification of activities. The final error
of both methods is similar, being slightly lower for the PTA-7A by only 0.09%.

7.5.2 Activity Classification Performance

The confusion matrices for the first HAR method PTA-6A are shown in Table 7.3 .
They depict the classification results of the system before and after the activity
temporal filters. The first noticeable difference between them is the matrix size as
the unknown-activity appears only after filtering. Additionally, it is evident that the
number of false negatives for the PT class is quite large before filtering. In particular,
dynamic activities such as walking-upstairs provide most of these misclassifications