Signal Processing - March 2016 - 91

Finally, we note that there is still progress
which the events are based (such as water
There is still progress
to be made such that research in ADL clasrunning in the sink, toilet flushing, etc.).
to be made such
sification can be reliably applied in a practiThe experiment with two households is
that research in
cal solution. This covers the optimization of
more challenging and also more interesting,
ADL classification
sensor setups for cost effectiveness, adaptive
first of all because the sensor setups between
classification algorithms that allow tracking
the households are similar but not identical,
can be reliably applied
changing behavior over time and robustness
and a BoW approach is used to map them in a
in a practical solution.
with respect to context changes such as hancommon feature space for learning and inferdling of visitors, caregiving personnel, or
ence. Furthermore, the households have difpets. One of the biggest challenges from the signal processing
ferent layouts and the persons are doing certain activities in very
and machine-learning side remains the generalizability over
different ways. For example, in one of the households, cooking
households. While training for an individual household is easalways involves opening and closing multiple kitchen cupboards
ily possible in a lab setup, this approach is not scalable to a
while in the other household this is only done sporadically. Nevreal-world scenario with thousands of households and more. A
ertheless, the classification scheme based on Fisher kernel is able
successful approach for generalizability has to consider envito do learning and inference in the joined space, with the only
ronmental/climate parameters, building layout, sensor placesignificant overlap between classes appearing between the
ment, and the behavior of the elderly.
Continence and Hygiene, which is to some extent attributable
to the different sensor setups and bath layouts.

Authors

Conclusions
Automatic classification of ADLs is a crucial part of assisted living
technologies. It enables automatic monitoring of the ability of an
elderly person to live independently in his or her house and can
allow for early detection of diseases such as Alzheimer's and
dementia. ADL classification involves the whole chain from a
plethora of wearable and nonwearable sensors, deployment
options, and signal processing and machine-learning algorithms.
Our study concludes that the recent developments in hybrid generative/discriminative methods, relying on kernel metric distances, are
superior over traditional generative methods such as HMM and its
variants. Specifically, FKL showed the best performance in a variety of data sets covering different activity types, sensors, and setups.
We expect continuing improvements on all aspects of the
aforementioned chain ranging from improvement of existing
sensor technologies, addition of new sensors, the acceptability of
certain technologies up to various algorithmic aspects such as the
generalizability and adaptiveness that are briefly detailed next.
Sensor technologies require improvements in several
directions including size, accuracy, energy efficiency, and
reliability. Establishing a common communication protocol
would allow to create an unified framework, providing a
significant speed up in infrastructure deployment. Reusability of sensors from smart home applications would be
another boosting factor helping to reduce infrastructure and
installation costs. Several legal issues related to the data
ownership and data security have to be addressed to get
acceptance for using ADL monitor systems.
While sensor technology improves, leading to higher-quality measurements and lower costs and maintenance, in practical
applications the (elderly) user needs to be taken into account
as well (see Figure 1). Specifically, user-centered design and
transparency can help to increase the acceptance of users of
technology in their home and their perceived privacy [15]. Furthermore, users are more and more exposed to sensor technology in other aspects of their lives, increasing their understanding
and, thereby, acceptance.

Christian Debes (cdebes@agtinternational.com) received B.Sc.,
M.Sc., and Ph.D. (with highest honors) degrees from from
Technische Universität Darmstadt, Germany, in 2004, 2006, and
2010, respectively. He currently holds a position as lead data scientist at AGT International and lecturer at Technische Universität
Darmstadt, Germany. He is a recipient of the IEEE Geoscience
and Remote Sensing Society 2013 Data Fusion Best
Classification Award and has authored more than 30 journal and
conference papers in target detection, classification, and image
processing. He is a member of the editorial board of Digital
Signal Processing (Elsevier) and area editor of IEEE Signal
Processing Magazine. He is a Senior Member of the IEEE.
Andreas Merentitis (andreas.merentitis@ieee.org) received
B.Sc., M.Sc., and Ph.D. degrees from National Kapodistrian
University of Athens in 2003, 2005, and 2010 respectively. In
2011, he joined the research center of AGT International,
Darmstadt, Germany, where he is now working as a senior data
scientist. He has more than 30 publications related to different
aspects of data analysis and machine learning. He was awarded
an IEEE Certificate of Appreciation as a core member of the
team that won the first place in the Best Classification Challenge
of the 2013 IEEE Geoscience and Remote Sensing Society Data
Fusion Contest. He is an associate editor of IEEE Signal
Processing Magazine and a Member of the IEEE.
Sergey Sukhanov (ssukhanov@agtinternational.com)
received B.Sc. and M.Sc. degrees from Saint-Petersburg State
Electrotechnical University, Russia, and Technische Universität
Darmstadt, Germany, respectively. Since 2015, he has been
working as a data scientist at AGT International, Darmstadt,
Germany. His research interests include data analysis and
machine learning. He is a Student Member of the IEEE.
Maria Niessen (mniessen@agtinternational.com) received
M.Sc. and Ph.D. degrees from the University of Groningen, The
Netherlands. After two years of postdoctoral research at INCAS,
she joined AGT International, Darmstadt, Germany, in 2012,
where she is currently a senior data scientist. Her research interests include data analysis and acoustics.

IEEE SIgnal ProcESSIng MagazInE

March 2016

http://www.B.Sc http://www.M.Sc http://www.B.Sc http://www.M.Sc http://www.B.Sc http://www.M.Sc http://www.M.Sc

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