Signal Processing - March 2016 - 85

Table 3. A summary of data sets collected in smart home environments, with their name and reference, whether wearable and/or nonwearable
sensors were installed, approximately how long the (average) recording time was, and whether it was recorded in a (living) lab or a real home.
Data set
CASAS [70]
HIS [27]
[23]
[24]
TigerPlace [71]
[65]
[72]
[63]

Institution
Washington State University
Grenoble TIMC-IMAG Lab
University of Virginia
University of Amsterdam
University of Missouri
Intel Research Seattle
Staffordshire University/Chiang Mai University
TU Darmstadt/Fraunhofer IGD

Sensor Types
Wearable and nonwearable
Wearable and nonwearable
Nonwearable
Nonwearable
Nonwearable
Wearable
Wearable
Wearable

Recording Duration
Up to months
Hours
Weeks
Weeks
Year
Weeks
Days
Hours

Lab/home
Lab and home
Lab
Home
Home
Home
Home
Lab
Lab

is the possibility of running the installation for long unincooking, and eating, to be able to determine any changes in
terrupted periods. However, because these relatively simple
their patterns. The experimental setting in which human
sensors require a wide coverage, the initial set up requires
activity data can be collected is called a smart home. A smart
more effort. Moreover, wearable sensors may not be easily
home is a normal living environment augmented with techaccepted by elderly users.
nology to improve the comfort or security of its residents [69].
The ambient approach is usually applied in experiIn the domain of AAL, sensors installed in the smart home
ments of longer duration in real-life settings, either in a
can be used to monitor the behavior of people living in the
smart home where participants live in an apartment (days
home. For example, a team at Washington State University
or weeks, e.g., [70]) or in a real apartment (e.g., [71]). In
introduced the Center for Advanced Studies in Adaptive Systhe controlled environment of a smart home it is easier to
tems (CASAS) Smart Home to test machine-learning techgather detailed and balanced data and annotate them, for
niques for human activity recognition [70].
example, with cameras, making it suitable to gather data
Depending on the focus of a study, the experimental sceto test activity recognition algorithms. On the other hand,
nario and, consequently, the requirements on the smart home
data recorded in real environments is more representative
environment vary. The smart home can be a real home where
of normal behavior and therefore more suited to test algosensors are installed, but it may also be a lab in which a smart
rithms for behavior modeling. For example, in [24], ambient
home is built and where temporary residents can stay for
sensors such as door contact sensors, motion sensors, and
a shorter or longer period of time. In addition, some studies
a float sensor in the toilet were used to recognize patterns
use predefined scenarios to be able to systematically evaluate
of activities. This example was followed in [76] as part of
activity recognition algorithms, while others investigate patthe CASAS project to detect broad activities such as eating
terns of normal behavior. Finally, the type of sensors that are
breakfast, sleeping, and wandering.
installed vary, depending on the focus,
e.g., energy efficiency or privacy considerations. Table 3 lists a selection of smart
ADL classification
generative models estimate
home data sets and properties of the experThe signal processing and machine-learnthe joint probability
imental settings.
ing methods that are referenced in the literdistribution of observation
Related to the two types of sensors
ature on ADL classification span a broad
samples, which can be
described in the "State of the art in senrange of techniques, from simple heuristics
used to predict the most
sor technology to assess ADLs" secto more advanced machine-learning algolikely class to which a
tion-wearable and nonwearable-the
rithms such as hidden Markov models
experimental approaches can be sepa(HMMs) and conditional random fields
new sample belongs.
rated in in-situ and ambient approaches.
(CRFs). Most of the classical machineIn the in-situ approach, the goal is to corlearning algorithms such as support vector
rectly identify particular activities, and this is often tested in
machines (SVMs) and random forests assume input data that
a laboratory setting for a short period of time according to
is independent and identically distributed (IID). However,
predefined scenarios. The types of sensors used are mostthere are certain cases where the independence assumption of
ly low cost and low power, so that many can be installed.
each data point does not hold. This is true, for example, in
These include accelerometers [63], [73], both body-worn and
speech recognition (every syllable is dependent on the nearby
attached to objects; RFIDs [74], [75], also both body-worn
ones) but also for human behavior modeling and recognition:
and attached to objects; and door contact sensors. Although
What someone is doing at a specific point in time is not indewearable sensors allow experiments to include activities outpendent from what he was doing just before. The taxonomy of
side of a home, contrary to the ambient approach, most work
machine-learning algorithms that are used for structured
in the in-situ approach and are applied indoors and in living
learning when the IID assumption does not hold is presented
labs. The advantage of using low-cost and low-power sensors
in Figure 3.
IEEE SIgnal ProcESSIng MagazInE

March 2016

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