Signal Processing - March 2016 - 90

Time-slice Acc.

perform different activities at the same
Similar to data set 1 with a one-perAutomatic classification
time and at different locations, dissimilar
son household, FKL shows the highest
of ADLs enables automatic
activities such as Sleep and Meal Prepaperformance in terms of average accuration can be also confused.
racy. random forest and SVM perform
monitoring of the ability of
slightly worse than FKL, while HMM
an elderly person to live
provides the worst average accuracy,
independently in his house Data set 3
mostly because of the complete failure
For the third data set, we evaluated two
and can allow for early
on the Watch TV class. A confusion
scenarios. In the first scenario (one
detection of diseases
matrix for FKL is provided in Figure 7.
household) training and testing were
such as Alzheimer
The classes Work and Meal Preparation
done only using events coming from a
can be discriminated the best among all
single household. In the second scenario
and dementia.
other ADLs in this data set, while Per(two households) training and testing
sonal Hygiene and Sleep demonstrate
were performed on data taken from both
moderate recognition accuracy, and Watch TV performs the
households. All events from both households were shuffled
worst. Since two people are living in the apartment and can
in a way that the event sequence within one ADL is kept to
preserve the structure in the data (non-IID assumption).
Evaluation results on data set 3 for random forest, SVM,
Table 6. Time-slice and class-average accuracy for SVM, RF, HMM,
HMM, and FKL on one and two households are provided
and FKL on data set 3 for the one household scenario.
in Tables 6 and 7, respectively.
SVM
RF
HMM
FKL
Also, for this data set, FKL outperforms the other methods,
No ADL
0.12
0.31
0.17
0.65
both for the one household and the two households scenario
Continence
0.39
0.90
0.75
0.70
(see Tables 6 and 7). However, the performance of random
Hygiene
0.93
0.81
0.63
0.79
forest and HMM are very close to FKL for the two houseShowering
0.89
0.73
0.55
0.94
holds scenario. The HMM even has an improved performance
Food preparation 0.96
0.93
0.98
0.75
compared to its performance for the one household scenario,
Class-average accuracy 0.66
0.74
0.62
0.77
probably because (for this data set) the improvement that it can
gain with more data is bigger than the potential loss because of
increased variation in the class representation.
Table 7. Time-slice and class-average accuracy for SVM, RF, HMM,
The confusion matrices of FKL for the two scenarios are
and FKL on data set 3 for the two households scenario.
shown in Figures 8 and 9. In the one household scenario, FKL
SVM
RF
HMM
FKL
can achieve high performance, with the only noteworthy conNo ADL
0.31
0.19
0.62
0.63
fusion appearing for very similar activities (see Figure 8). For
Continence
0.87
0.91
0.64
0.54
example, Food Preparation and No ADL (which is defined as an
Hygiene
0.72
0.74
0.57
0.78
activity that is not described by a formal ADL, e.g., standing in
Showering
0.64
0.44
0.80
0.87
the kitchen, reading at a table, etc.) have an overlap while ConFood preparation 0.88
0.94
0.84
0.74
tinence and Hygiene also are partly difficult to discriminate in
Class-average accuracy 0.68
0.64
0.69
0.71
the absence of clear signature detections from the sensors on

0.06

0.29

0.79

Showering

0.06

0.14

0.07

0.87

0.5
0.4

0.4
Showering 0.05

0.3

0.94
0.75

Food 0.24
Preparation

0.1

L
oA
D

0.2
0.74

0.1
0

ar Fo
at od
io
n

0.02

0.3

Hy
gi
en
e
Sh
ow
er
in
g

Co
nt
in
en
ce

0.78

0.5

0.25

No
AD
L

0.06

0.8
0.6

0.6

0.2
Food
Preparation

Hygiene 0.02

0.27

ep F
ar o
at od
io
n

0.07

0.43

in
g

0.14

0.54

Hygiene

0.04

Continence 0.03

0.05

0.7

ie
ne

0.22

0.02

Sh
ow

0.70

0.03

0.04

NoADL 0.63

0.8

Continence

0.9

0.65

tin
en

NoADL

figure 8. The confusion matrix for FKL on data set 3 for one household.
90

figure 9. The confusion matrix for FKL on data set 3 for two households.

IEEE SIgnal ProcESSIng MagazInE

March 2016

Table of Contents for the Digital Edition of Signal Processing - March 2016