Online Recognition with Postural Transition Awareness - Smartphone-Based Human Activity Recognition - page 122

Table 7.4 PTA-7A confusion matrix. Before and after filtering
Method

PTA-7A

Activity WK WU WD SI ST LD PT

WK 539 1 3 2 0 0 1

WU 28 513 2 1 1 0 14

WD 2 5 498 0 4 0 0

SI 0 3 0 486 68 0 0

ST 1 0 0 19 591 0 1

LD 0 0 0 0 0 604 0

PT 3 2 0 2 0 0 322

Method PTA-7A

Activity WK WU WD SI ST LD PT UA

WK 542 1 0 1 0 0 1 1

WU 26 510 1 0 1 0 16 5

WD 0 0 506 0 3 0 0 0

SI 0 0 0 499 57 0 1 0

ST 0 0 0 3 607 0 2 0

LD 0 0 0 0 0 604 0 0

PT 0 0 0 2 0 0 327 0

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

Table 7.5 Sensitivity and specificity
Activity Sensitivity Specificity

PTA-6A (%) PTA-7A (%) PTA-6A (%) PTA-7A (%)

WK 99.82 99.27 99.11 99.18

WU 91.95 91.23 99.65 99.97

WD 99.21 99.41 99.97 99.97

SI 90.48 89.58 99.68 99.81

ST 99.67 99.18 99.67 98.03

LD 100.0 100.0 100.0 100.0

PT - 99.39 - 99.41

The sensitivity and specificity measures for each activity were also estimated
and presented in table Table 7.5 . In this way, we can analyze separately how the
actual positives and negatives are correctly classified. For the 6 BAs we obtained a
high specificity, also something similar occurred with the classes sensitivity except
for the classes walking-upstairs and sitting which have a smaller value. They are
being mostly misclassified as walking and standing respectively which are similar
activities. It is also visible that the detection of PTs is working well as the sensitivity
and sensitivity are above 99%. This is only applicable to the PTA-7A as in the other
method PTs are not learned.