C omp u t ing Pr ac t ice s
CASAS: A
Smart Home
in a Box
Diane J. Cook, Aaron S. Crandall, Brian L. Thomas, and Narayanan C. Krishnan
Washington State University
The CASAS architecture facilitates the
development and implementation of future
smart home technologies by offering an
easy-to-install lightweight design that
provides smart home capabilities out of
the box with no customization or training.
I
n the past decade, machine learning and pervasive
computing technologies have matured to the point
where they provide integrated and automated context-
aware support in our everyday environments. A smart
home is a physical embodiment of such a system. Computer
software uses sensors and artificial intelligence techniques
to perceive and reason about the state of the home’s physi-
cal environment and its residents, and then initiates action
to achieve specified goals.
During perception, sensors embedded in the home gen-
erate readings while residents perform their daily routines.
The sensor readings are collected by a computer network
and stored in a database that an intelligent agent uses to
generate useful knowledge such as patterns, predictions,
and trends. On the basis of this information, a smart home
can select and automate actions that meet the goals of the
smart home application.
Many researchers view the potential uses of smart home
technology for applications such as health monitoring and
energy-efficient automation as “extraordinary.”1 However,
62 computer Published by the IEEE Computer Society 0018-9162/13/$31.00 © 2013 IEEE
Authorized licensed use limited to: Eskisehir Teknik Universitesi. Downloaded on August 29,2021 at 20:35:39 UTC from IEEE Xplore. Restrictions apply.
to date most implementations are somewhat narrow and
are performed primarily in controlled laboratory settings.
These limitations are due in large part to the difficulty
of creating a fully functional smart home infrastructure.
In fact, although realistic smart home prototypes
exist,2,3 implementing these designs is so cumbersome
that meetings have been organized to discuss ways to
scale such pervasive computing systems and to share
valuable data that have been successfully captured in
such settings. A recent example of such a meeting is the
2012 National Science Foundation Workshop on Pervasive
Computing at Scale (http://sensorlab.cs.dartmouth.edu/
NSFPervasiveComputingAtScale).
The goal of the Washington State University’s CASAS
project is to design a smart home in a box (SHiB). This smart
home kit is small in form, lightweight in infrastructure,
extendable with minimal effort, and ready to perform key
capabilities out of the box.
ARCHITECTURE
Figure 1 shows the CASAS architecture. The physical
layer contains hardware including sensors and actuators.
The architecture utilizes a ZigBee wireless mesh which com-
municates directly with hardware components. A publish/
subscribe manager governs the middleware layer. The
manager provides named broadcast channels that allow
component bridges to publish and receive messages. The
middleware provides valuable services, such as adding
time stamps to events, assigning universally unique identi-
fiers (UUIDs), and maintaining site-wide sensor state. Every
component of the CASAS architecture
communicates via a customized Exten-
sible Messaging and Presence Protocol Activity Activity
recognition discovery Energy
(XMPP) bridge to this manager. Ex-
amples include the ZigBee bridge; the
Scribe bridge, which archives messages Application bridge
in permanent storage; and bridges
Scribe
for each application-layer software bridge
component. Publish/subscribe manager
The CASAS architecture is easily Archive
maintained because the communica- storage
tion bridges use lightweight APIs that ZigBee bridge
support a wide variety of free-form
messages. As a result, the middleware
is compact and stable—it has had only ZigBee wireless mesh
one update in five years. CASAS is ex-
tendable because users can configure
and integrate new bridges without
changing or even restarting the middle- Sensor/
ware. In addition, we have designed actuator
Actuator
bridges that link multiple smart homes Sensor
together, allowing CASAS to scale to
communities of smart homes.
As Figure 2a shows, all of the CASAS Figure 1. CASAS smart home components. During perception, control flows
up from the physical components through the middleware to the software
components fit into a single small box. applications. When the smart environment initiates an action, control moves from
The physical components in the cur- the application layer to the physical components that automate the action. The
rent box consist of sensors that are goal is that each layer is lightweight, extensible, and ready to use.

(a)

(b)

Figure 2. CASAS smart home in a box: (a) SHiB kit and (b) smart home installation site. The sensors come prelabeled with
their intended location.

JULY 2013 63
Authorized licensed use limited to: Eskisehir Teknik Universitesi. Downloaded on August 29,2021 at 20:35:39 UTC from IEEE Xplore. Restrictions apply.
 C omp u t ing Pr ac t ice s
Table 1. Summary of costs for CASAS smart home in a box (SHiB).
Component
Computer server
Infrared motion/light sensor
Door sensor
Relay
Temperature sensor
Total cost
Note: prices are accurate at time of publication.
prelabeled with the intended location. Additional sensors
and controllers can be included as needed. The middle-
ware, database, and application components reside on a
small, low-power computer with an ITX form-factor server.
Although this layout lets each smart home run indepen-
dently and locally, smart homes can also securely upload
events to be stored in a relational database or in the cloud.
Table 1 summarizes the prototype costs of the
components that comprise our SHiB design.
USABILITY
Our research group has installed 32 smart home
testbeds to date. Many of the corresponding datasets are
available on the project webpage at ailab.wsu.edu/casas.
A total of 19 datasets represent single-resident sites, four
represent sites with two residents, and the rest house
larger families or residents with pets.
Figure 2b shows a smart home installation site. The
CASAS smart home design minimizes installation costs.
We can install a new smart home in approximately two
hours and can remove the equipment in 30 minutes, with
no changes or damage to the home.
Once the smart home system is installed, the residents
must maintain the equipment. The CASAS SHiB includes a
software agent that alerts residents if sensor battery levels
are getting low or if a sensor suddenly stops reporting
events. In practice, this seldom happens as the batteries
typically last more than a year.
To test the kit’s usability, we conducted a study in an
on-campus three-bedroom apartment. We recruited
participants to visit the apartment, one at a time, and
install a CASAS smart home. The study included 20
participants (eight men and 12 women) aged 21 to 62 years
(a mean age of 33 years), with a variety of backgrounds and
technological familiarity.
We gave each participant a writ ten document
explaining the smart home parts and installation process
and a CASAS smart home kit. All of the participants
completed the installation without difficulty. The
average installation time was just over one hour. On a
scale of 1 (simple) to 10 (impossible), participants rated
the installation difficulty as 2.53 (σ = 1.07). They noted
64 computer
Authorized licensed use limited to: Eskisehir Teknik Universitesi. Downloaded on August 29,2021 at 20:35:39 UTC from IEEE Xplore. Restrictions apply.
Unit price Quantity Total price
$350 1 $350
$85 24 $2,040
$75 1 $75
$75 2 $150
$75 2 $150
$2,765
that the most difficult issue was determining the optimal
placement of sensors.
SYSTEM CAPABILITIES
Ideally, a smart home installation can be accomplished
using only a toolkit that works out of the box, with no
customization or training. We designed two core software
components and two applications to meet this goal.
Activity recognition
Intelligent systems that focus on human needs require
information about the human’s activities. At the core of
these systems, then, is activity recognition.4,5 Smart home
sensors generate events consisting of a date, time, sensor
identifier, and sensor message. Activity recognition maps
a sequence of sensor data to a corresponding activity label.
The CASAS activity recognition software, called AR,
provides real-time activity labeling as sensor events arrive
in a stream. To achieve this functionality, we formulated
the learning problem as one of mapping the sequence of
the k most recent sensor events to a label indicating the
activity corresponding to the most recent event in the
sequence. The preceding sensor events define the context
for this event. For example, the sequence of sensor events
consisting of
2011-06-15 03:38:23.271939 BedMotionSensor ON
2011-06-15 03:38:28.212060 BedMotionSensor ON
2011-06-15 03:38:29.213955 BedMotionSensor ON
could be mapped to a Sleep activity label.
We designed a support vector machine (SVM) method
for real-time activity recognition. We have tested other
machine learning models as well, including naïve Bayes
classifiers, hidden Markov models, and conditional random
fields. However, we found that SVMs achieve consistently
stronger performance than other approaches. In addition,
the model quantifies the degree of fit between the data
and provides an activity label that facilitates additional
capabilities such as anomaly detection.
To provide input to the classifiers, we define features
describing a data point i that corresponds to a sequence
 Table 2. Activity recognition confusion matrix.

Ground truth Automatically generated activity label

activity label
Bed-toilet Enter Leave Personal
transition Cook Eat home home hygiene Phone Relax Sleep Work Accuracy

Bed-toilet 18,288 143 261 0 0 22,233 0 3 5,866 38 0.39

transition
Cook 3 370,300 1,616 11 11 172 4 140 28 1,917 0.99
Eat 53 20,528 9,871 4 0 41 1 979 118 27,052 0.17
Enter home 0 195 0 1,606 107 3 0 4 57 126 0.77
Leave home 0 5 0 59 316 3 0 0 1 4 0.81
Personal 15,769 928 81 3 3 295,616 0 77 1,216 921 0.94
hygiene
Phone 0 21 2 0 0 4 8 34 73 1,072 0.01
Relax 6 1,282 322 13 0 178 8 2,030 1,459 2,735 0.25
Sleep 33,900 66 33 1 0 279 0 60 65,189 306 0.65
Work 37 2,875 10,544 66 17 489 20 497 237 71,684 0.83
Note: the diagonal entries indicate the activities that were correctly categorized; the last column shows the accuracy for each individual activity.

of sensor events. This fixed dimensional feature vector xi Activity discovery

includes the time of day for the first and last sensor events
(discretized into four equal-length bins), the time span of
the k-event sensor window, and the number of events for
each sensor within the window.
Each vector x i is tagged with the label y i , which
corresponds to the activity label associated with the last
sensor event in the window. Although we could identify
a fixed window size k that works well for a given dataset,
this approach requires additional user customization. To
increase the approach’s generalizability, AR dynamically
adjusts the window size k based on the most likely activities
that are being observed and their typical duration.
To evaluate AR’s ability to recognize activities out of
the box, we collected sensor data in 18 separate smart
apartments, each housing one resident and each using
the CASAS smart home kit. We manually annotated one
month of data to provide ground truth activity labels. We
evaluated performance as the percentage of sensor events
that were correctly labeled across all of the apartments,
with no additional customized training for each apartment.
Table 2 shows the confusion matrix we generated from
this experiment.
As the matrix indicates, it is easier to recognize
some activities than others. This occurs because
some activities, such as cooking, have a fairly unique
spatiotemporal signature. Other activities are more
challenging because they overlap with other activity
classes or not enough training data is available to learn
the model. The weighted average accuracy is 84 percent,
which indicates that the models are fairly robust even
when they are used out of the box in new, distinct home
settings.
Recognizing activities from streaming data introduces
new challenges because AR must process data that does
not belong to any of the targeted activity classes. One way
to handle unlabeled data is to design an unsupervised
learning algorithm to discover activities from unlabeled
sensor data. Segmenting unlabeled data into smaller
classes improves activity recognition performance because
the “other” class is no longer the largest, as frequently
happens in activity recognition datasets. Another
important reason to discover activity patterns from
unlabeled data is to characterize and analyze as much
behavioral data as possible, not just predefined activity
classes. Researchers must examine and model such
unlabeled data to get a complete view of everyday life.6
Like earlier approaches to sequence mining, our activity
discovery algorithm, called AD, searches the space of
candidate sensor event sequences ordered by increasing
sequence length. Because the space of candidate patterns
is exponential in the input data size, we use a greedy
search to find the sequence pattern that best compresses
the input dataset. During discovery, AD scans the entire
dataset to create initial patterns of length one.
After this initial pass, AD extends the patterns
discovered in the previous iteration by considering events
occurring before and after instances of the previous
pattern. AD stores the patterns in a beam-limited open
list and value-orders them. Once the search terminates and
the activity discovery algorithm reports the best pattern
found, AD compresses the sensor event data using the
best pattern. The compression procedure replaces all
instances of the pattern with single event descriptors
representing the pattern definition. AD can then invoke

JULY 2013 65
Authorized licensed use limited to: Eskisehir Teknik Universitesi. Downloaded on August 29,2021 at 20:35:39 UTC from IEEE Xplore. Restrictions apply.
 C omp u t ing Pr ac t ice s
11 12
10
9
11 12 1
2
3
10
9
8 4 8 4
7 5 7 6
6
(a) (b)
11 12 1
10
9
8
2
4
3 Sensor key:
7 6 5
(c)
Figure 3. Visualization of discovered patterns: (a) sequence of motions that
might represent a person getting ready for bed; (b) sequence of motions
that could indicate a person returning home; and (c) sequence of motions
that might represent a person getting ready to begin a daily routine.
the activity discovery process on the compressed data to
find additional activity patterns.
We evaluate candidate patterns based on their ability to
minimize the original dataset’s size when it is compressed
using the pattern definition. Because AD’s compression
replaces each pattern occurrence with a single event labeled
with a pattern identifier, it calculates the description length
of a pattern P, given input data D, as DL(P) + DL(D|P), where
DL(P) is the description length of the pattern definition, and
DL(D|P) is the description length of the dataset compressed
using the pattern definition. Because human behavior
patterns vary greatly, we employ an edit distance measure
to determine if a sensor sequence is sufficiently similar to
a pattern to be considered an instance of the pattern. This
measure counts the minimum number of add, delete, or
transpose operations needed to convert a sensor sequence
to one that is equivalent to the pattern definition.
Figure 3 provides a visualization of the three top
activity patterns that occur when we apply the AD activity
discovery algorithm to our combined dataset. The pattern
in Figure 3a shows a sequence consisting of motion in
the bedroom, followed by motion in the living room,
followed by more motion in the bedroom, around 10:20
p.m. Many of these events occur prior to sleeping and thus
can represent a person getting ready for bed.
The pattern shown in Figure 3b consists of a front
door closing, followed by a series of kitchen events, and
then a living room event, usually in the late morning or
66 computer
Authorized licensed use limited to: Eskisehir Teknik Universitesi. Downloaded on August 29,2021 at 20:35:39 UTC from IEEE Xplore. Restrictions apply.
midafternoon. This could represent a person’s
1
2
3
activities upon returning home, such as putting
away groceries or getting a drink.
5
The pattern shown in Figure 3c consists of
a sequence of events occurring in the morning
that alternate between the bedroom, a work
area, and the living room. This pattern might
represent gathering items needed for the
resident’s daily routine.
Other patterns represent transitions between
activities or activities that are recognizable
but do not appear on the predefined activity
: Motion list, such as spending extended time in a
: Motion (area) secondary bedroom that is used for guests
: Door or crafts. Using the AD activity discovery
: Temperature
: Light algorithm to find patterns in unlabeled data
: Water, burner that would otherwise be labeled as “other”
: Item increases activity recognition for our smart
home datasets by an average of 10 percent.7
Activity-aware applications
The world’s population is aging, with the
estimated number of people over age 85
expected to triple by 2050.8 Instead of deploying
healthcare reactively, we need innovative and
preventive healthcare methods that can be
automated and deployed within an individual’s own home.
We installed 20 smart homes at an assisted care facility
where residents’ average age is 85 years. Because the
CASAS smart home kit is simple to maintain, we can collect
data over multiple years, allowing us to monitor behavioral
changes that indicate variations in cognitive or physical
health. As Figure 4 shows, the monitored parameters
include overall activity level, sleep quality, and time spent
on individual activities of interest.
In addition, CASAS can provide users with activity-aware
health assistance in the form of prompting them to initiate
daily activities such as taking medicine, exercising, or
talking to their children. Although many reminder systems
exist, few take into account an individual’s behavioral
patterns to provide context-aware prompts, despite studies
indicating that they offer significant advantages over
traditional time-based prompts.9
In the CASAS software, a machine learning algorithm
is trained to identify when an individual performs an
activity as a function of wall-clock time such as “pick up
grandchildren at 2:00 p.m.” and as a function of other
activity occurrences such as “take medicine with breakfast.”
An additional application is supporting energy-efficient
behavior in the home. Over the past 40 years, energy con-
sumption has increased at a higher rate than population
growth, and buildings are now responsible for 40 percent
of total energy usage.10 By identifying activities occurring
in the home and concurrently monitoring whole-home
Figure 4. Activity trends for a smart-home resident. The graphs include a plot of number of sensor events over the time period
(top left); daily indicators for activity level, sleep quality, and time out of the home (top middle); comparison of socialization
and sleep quality parameters for a time period (top right); occurrences of tracked activities (bottom left); time spent in
different areas of the home (bottom middle); and health trend over a time period calculated as a function of activity level,
sleep quality, and socialization (bottom right).

CASAviz: Web-based Visualization

of Behavior Patterns

Figure 5. Snapshot of the CASAS activity visualizer. The visualizer renders sensor events on a computer or mobile device while
plotting usage of resources such as electricity.

JULY 2013 67
Authorized licensed use limited to: Eskisehir Teknik Universitesi. Downloaded on August 29,2021 at 20:35:39 UTC from IEEE Xplore. Restrictions apply.
 C omp u t ing Pr ac t ice s

1.2 0.5
1.0 0.4
Relative activity level

Relative activity level

0.8
0.3
0.6
0.2
0.4
0.2 0.1

0 0
1 3 5 7 9 11 13 15 17 19 21 23

ne

an
ep
lax

rk

Hy r

Eat

Be k
t

ve

ds
te

ile
o
Wo

Lea

Me
Co
gie
Sle

Cle
En
Re

dTo
(a) Hour of day (b) Activity level

Figure 6. Relative activity level (a) as a function of the hour of the day and (b) as a function of the activity class.

energy usage, we can predict an activity’s energy con-
sumption. In addition to providing this information to the
home’s resident, as Figure 5 shows, the smart home can
promote energy-efficient behavior11 and automate con-
trol of selected devices to support more energy-efficient
activities.
POPULATION-WIDE FINDINGS
One type of analysis not found in the literature is a pop-
Larger variances exist for the enter activity (which con-
siders time spent outside the home) and for bed-toilet
transitions, which vary dramatically by age, health, and
sleep quality.
Conducting such large-scale analyses will provide a
valuable tool for understanding behavior that is central to
many research fields, including sociology, psychology, and
technology development.

ulation-wide evaluation of resident behavior using smart
home data. Although analyzing behavioral features across
a larger demographic could benefit researchers in many
fields, gathering data at a significant scale has not yet been
a practical goal. However, researchers can use CASAS to
investigate questions that apply to demographic groups,
families, and communities.
As a first step, we consider behavioral properties for the
CASAS datasets we have collected. In particular, we want to
identify how activity levels vary throughout the day for an
entire cohort. We also want to determine how individuals
spend their home time in terms of individual activities and
how consistent the functions are across the group.
Figure 6 shows the results of these two analyses for the
18 smart apartments included in our study. As Figure 6a
indicates, a clear pattern exists for the entire group, with
low activity levels early in the day but increasing, peak-
ing at midmorning, midafternoon, and early evening. The
exact activity levels vary across the population. This vari-
ance might be due either to mobility differences or sensor
granularity within the home.
In contrast, the variance across the population for
time devoted to various activities is much smaller. As
Figure 6b shows, the most time is dedicated to sleep,
while other activities, such as taking medicine (which
is typically quick) and cleaning the home (which might
not happen as often as other activities), receive less time.
A
s a next step in our work, we plan to evaluate the
ease with which we can incorporate additional
sensor modalities such as radio frequency identifi-
cation and smartphones into the CASAS architecture and
to design applications that more extensively utilize device
controllers. We also anticipate expanding the data collec-
tion to include a greater diversity of resident demographics
so that we can perform longitudinal studies. Finally, our
future work will focus on designing home automation
strategies that provide safe and energy-efficient support
of a resident’s daily activities.
Acknowledgments
We thank Jim Kusznir, Allan Drassal, Leah Zulas, and all the
members of the CASAS team for their contributions to this
work. This material is based on work supported by the US
National Science Foundation under grant number 0852172,
the Life Sciences Discovery Fund, and National Institute of
Biomedical Imaging and Bioengineering (NIBIB) grant number
R01EB009675.
References
1. P. Hewitt, “Speech by the Rt Hon Patricia Hewitt MP,
Secretary of State for Health, 16 May 2007: Long-term
Conditions Alliance Annual Conference,” 2007; http://
webarchive.nationalarchives.gov.uk/+/www.dh.gov.uk/en/
MediaCentre/Speeches/DH_074812.

68 computer

Authorized licensed use limited to: Eskisehir Teknik Universitesi. Downloaded on August 29,2021 at 20:35:39 UTC from IEEE Xplore. Restrictions apply.
 2. S. Helal et al., “The Gator Tech Smart House: A
Programmable Pervasive Space,” Computer, Mar. 2005,
pp. 50-60.
3. B. Logan et al., “A Long-Term Evaluation of Sensing
Modalities for Activity Recognition,” Proc. 9th Int’l Conf.
Ubiquitous Computing (UbiComp 07), Springer, 2007,
pp. 483-500.
4. D.H. Hu, V.W. Zheng, and Q. Yang, “Cross-Domain Activity
Recognition Via Transfer Learning,” Pervasive and Mobile
Computing, vol. 7, no. 3, 2011, pp. 344-358.
5. T.L.M. van Kasteren, G. Englebienne, and B.J.A. Kröse,
“Hierarchical Activity Recognition Using Automatically
Clustered Actions,” Proc. 2nd Int’l Conf. Ambient Intelligence
(AmI 11), Springer, 2011, pp. 82-91.
6. T. Gu et al., “An Unsupervised Approach to Activity
Recognition and Segmentation Based on Object-Use
Fingerprinters,” Data & Knowledge Eng., vol. 69, no. 6, 2010,
pp. 533-544.
7. D. Cook, N. Krishnan, and P. Rashidi, “Activity Discovery
and Activity Recognition: A New Partnership,” to appear
in IEEE Trans. Systems, Man, and Cybernetics, Part B, 2013;
http://ee.wsu.edu/~cook/pubs/smc12.pdf.
8. G.K. Vincent and V.A. Velkoff, “The Next Four Decades—
The Older Population in the United States: 2010 to 2050,”
US Census Bureau, 2010; www.census.gov/prod/2010pubs/
p25-1138.pdf.
9. P. Kaushik, S.S. Intille, and K. Larson, “User-Adaptive
Reminders for Home-Based Medical Tasks: A Case Study,”
Methods of Information in Medicine, vol. 47, no. 3, 2008, pp.
203-207.
10. L. Pérez-Lombard, J. Ortiz, and C. Pout, “A Review of
Building Energy Consumption Information,” Energy and
Buildings, vol. 40, no. 3, 2008, pp. 394-398.
11. A. Faruqui, S. Sergici, and A. Sharif, “The Impact of
Informational Feedback on Energy Consumption—A
Survey of the Experimental Evidence,” Energy, vol. 35,
no. 4, 2010, pp. 1598-1608.
Call for Articles
Authorized licensed use limited to: Eskisehir Teknik Universitesi. Downloaded on August 29,2021 at 20:35:39 UTC from IEEE Xplore. Restrictions apply.
Diane J. Cook is the Huie-Rogers Chair Professor in the
School of Electrical Engineering and Computer Science at
Washington State University. Her research interests in-
clude artificial intelligence, machine learning, graph-based
relational data mining, smart environments, and robotics.
Cook received a PhD in computer science from the Univer-
sity of Illinois. She is an IEEE Fellow. Contact her at cook@
eecs.wsu.edu.
Aaron S. Crandall is an assistant research professor in the
School of Electrical Engineering and Computer Science at
Washington State University. His research interests include
the application of artificial intelligence, human factors, and
engineering principles to building better smart home sys-
tems. Crandall received a PhD in computer science from
Washington State University. He is a member of IEEE and
ACM. Contact him at acrandal@eecs.wsu.edu.
Brian L. Thomas is a PhD student and IGERT Fellow in the
School of Electrical Engineering and Computer Science at
Washington State University. His research interests include
artificial intelligence, home automation, and computer se-
curity. Thomas received a BS in computer science from
Washington State University. Contact him at bthomas@
eecs.wsu.edu.
Narayanan C. Krishnan is an assistant research professor
in the School of Electrical Engineering and Computer Sci-
ence at Washington State University. His research interests
include activity recognition, pervasive computing, pattern
recognition, and machine learning for pervasive computing
applications. Krishnan received a PhD in computer science
and engineering from Arizona State University. Contact
him at ckn@eecs.wsu.edu.
Selected CS articles and columns are available
for free at http://ComputingNow.computer.org.
IEEE Software seeks practical,
readable articles that will appeal
to experts and nonexperts alike.
The magazine aims to deliver reliable
information to software developers
and managers to help them stay on
top of rapid technology change.
Author guidelines:
www.computer.org/software/author.htm
Further details: software@computer.org
www.computer.org/software
JULY 2013 69