Content uploaded by Kai Kunze
Author content
All content in this area was uploaded by Kai Kunze on Oct 04, 2014
Content may be subject to copyright.
Smarter Eyewear– Using Commercial
EOG Glasses for Activity Recognition
Shoya Ishimaru
Graduate School of Engineering
Osaka Prefecture University
Sakai, Osaka, Japan
ishimaru@m.cs.osakafu-u.ac.jp
Yuji Uema
Graduate School of Media Design
Keio University
Yokohama, 223-8526 Japan
uema@kmd.keio.ac.jp
Kai Kunze
Graduate School of Engineering
Osaka Prefecture University
Sakai, Osaka, Japan
firstname.lastname@gmail.com
Koichi Kise
Graduate School of Engineering
Osaka Prefecture University
Sakai, Osaka, Japan
kise@cs.osakafu-u.ac.jp
Katsuma Tanaka
Graduate School of Engineering
Osaka Prefecture University
Sakai, Osaka, Japan
ishimaru@m.cs.osakafu-u.ac.jp
Masahiko Inami
Graduate School of Media Design
Keio University
Yokohama, 223-8526 Japan
inami@kmd.keio.ac.jp
Permission to make digital or hard copies of part or all of this work for personal
or classroom use is granted without fee provided that copies are not made or
distributed for profit or commercial advantage and that copies bear this notice
and the full citation on the first page. Copyrights for third-party components
of this work must be honored. For all other uses, contact the Owner/Author.
Copyright is held by the owner/author(s).
UbiComp’14 Adjunct, September 13-17, 2014, Seattle, WA, USA
ACM 978-1-4503-3047-3/14/09.
http://dx.doi.org/10.1145/2638728.2638795
Abstract
Smart eyewear computing is a relatively new subcategory
in ubiquitous computing research, which has enormous
potential. In this paper we present a first evaluation of
soon commercially available Electrooculography (EOG)
glasses (J!NS MEME) for the use in activity recognition.
We discuss the potential of EOG glasses and other smart
eye-wear. Afterwards, we show a first signal level
assessment of MEME, and present a classification task
using the glasses. We are able to distinguish of 4 activities
for 2 users ( typing, reading, eating and talking) using the
sensor data (EOG and acceleration) from the glasses with
an accuracy of 70 % for 6 sec. windows and up to 100 %
for a 1 minute majority decision. The classification is done
user-independent.
The results encourage us to further explore the EOG
glasses as platform for more complex, real-life activity
recognition systems.
Author Keywords
Smart Glasses, Electrooculography, Activity Recognition,
Eye Movement Analysis
ACM Classification Keywords
I.5.4 [PATTERN RECOGNITION Applications]: Signal
processing
Introduction
With wearable computing receiving increasing interest
from industry, we believe that especially smart eyewear is
a fascinating research area. In this paper, we show that
Figure 1: The EOG Glasses used
for the experiments. The second
picture shows the 3 electrodes
touching each side of the nose
and the area between the eyes.
The last picture shows a user
wearing JINS MEME.
the sensor data quality obtained by the EOG glasses
seems good enough for activity recognition tasks.
The contributions are two fold. First, we want to motivate
that smart eye wear is interesting for ubiquitous
computing applications, as it enables to track activities
that are hard to observe otherwise, especially in regard to
cognitive tasks.
Second, we evaluate specific smart glasses, a prototype of
J!NS MEME (available to consumers next year) for their
use for activity recognition tasks. We show a signal level
evaluation and a simple classification task of 4 activities (
2 users 2 x 5 min. per activity). Both indicate that the
device can be used for more complex scenarios.
In the end we discuss application scenarios and limitations
for smart eyewear.
Toward Using Smarter Eyewear
Since the release of Google Glass, smart eyewear gains
more and more traction for a wide range of applications
(e.g. oculus rift for virtual reality). This new class of
devices proves to be an interesting platform for ubiquitous
computing, especially for activity recognition. As we
humans perceive most of our environment with senses on
our head (hearing, smell, taste and most dominantly our
vision), the head is very valuable position for sensors.
Tracking eye movements can give us great insights about
the context of the user, from recognizing what documents
a user is reading, over recognizing memory recall to
assessing expertise level [8, 6, 1, 4, 7].
Hardware
To evaluate the potential of smart eyewear for activity
sensing, we are using an early prototype from J!NS
MEME. The glasses are not a general computing
platform. They are a sensing device. They can stream
sensor data to a computer (e.g. smart phone, laptop,
desktop) using Bluetooth LE. Sensor data includes vertical
and horizontal EOG channels and accelerometer +
gyroscope data. The runtime of the device is 8 hours
enabling long term recording and, more important, long
term real-time streaming of eye and head movement.
They are unobtrusive and look mostly like normal eyewear
(see Figure 1).
Before recording with the device the first time, the
electrodes should be adjusted a bit to the user’s nose/eyes
to get an optimal EOG signal. This is a one-time
adjustment due to the early prototype stage.
Initial Signal Level EOG Evaluation
A manual inspection of the data recorded by the
prototype reveals that detecting blinks and reading
activity seems feasible. A signal example from a user
blinking 7 times is given in Figure 2. We depict the raw
vertical EOG component before any filtering, even then
the blinks are easy recognizable. Another signal example
from the horizontal EOG component is shown in Figure 3.
In this case, the user was reading. Again we depict the
”raw” horizontal component of the EOG signal.
Simple Blink detection –We apply a very simple peak
detection algorithm on data from 2 users. We consider a
point being the a peak if it has the maximal value, and
was preceded (to the left) by a value lower by a constant.
2 users with wearing the smart glasses sitting in front of
the stationary eye tracker and blink 30 times naturally.
We can detect 58 of the 60 blinks with this very simple
algorithm applied to the ”raw” vertical EOG component
signal.
Figure 2: The vertical EOG
component (raw signal), while
the user blinks seven times.
Figure 3: The horizontal EOG
component (raw signal), while
the user reads.
Classification Task
For a first impression, we evaluate if the sensor data from
the glasses can distinguish more complex activity
recognition tasks. We assume that modes of locomotion
etc. can easily be recognized by the motion sensors alone.
Therefore we concentrate on tasks performed while sitting
in a common office scenario. We include 4 activities:
typing a text in a word processor, eating a noodle dish,
reading a book and talking to another person.
Method
We use a simple classification method: windowed feature
extraction with a K-Nearest Neighbor classifier (k = 5)
and majority decision. 7 features are calculated over a 6
sec sliding window (2 overlapping): the median and
variance of the vertical and horizontal EOG signal and the
variance for each of the 3 accelerometer axes. The
features are used to train the nearest neighbor classifier.
On top of the classification we apply a 1 minute majority
decision for smoothing.
Experimental Setup
For the experimental setup, we record data using the J!NS
MEME prototype connected over Bluetooth to a Windows
laptop for 2 participants 4 activities, each activity for 2 x
5 min. We asked them to perform the activities naturally
while sitting at a desk.
Before starting to record with a participant, we need to
adjust the electrodes on the current prototype towards the
facial features of the user to be sure to capture a clean
EOG signal. This initial setup step needs to be done only
once per user.
Initial Results and Discussion
We apply the windowed feature extraction and
classification method on the data, performing a user
independent classification, training with the data of one
user and evaluating with the other user.
For the frame-by-frame classification we reach a correct
classification rate of 71 % on average for a 6 sec. window
(2 sec. overlap). The confusion matrix is given in
Figure 4. Applying the majority decision window of 1 min.
we reach 100 % discrimination between classes.
Strengthened by the good performance distinguishing the
4 activities for 2 users in a user independent way, we will
evaluate the platform to see if the detection of specific
activities is possible in real life situations during long term
deployment. Being able to detect food intake behavior or
learning tasks (e.g. reading) are of particular interest to
us.
Related Work
We follow the early pioneering work from Bulling et al.
and Manabe et al. in using EOG for activity
recognition[9, 3]. Bulling et al. described an approach to
recognize different visual activities using EOG prototypes,
including reading, solely from gaze behavior using
machine learning techniques in stationary and mobile
settings [3, 2].
There is some work to use Google Glass as activity
recognition platform. This work is complementary to the
approach in this paper, as Google Glass is a very different
device (a full fledged wearable computer) with different
sensing modalities [5]. Most of the related work uses
dedicated research prototypes, often attaching electrodes
directly to the skin above or below the eye.
Conclusion
We presented an initial evaluation of a smart glasses
prototype for activity recognition. Both signal level
analysis and 4 activity classification task show favorable
results. 58 of 60 blinks for 2 users can be detected by
straight forward peak detection. The 4 activities, typing,
eating, reading and talking can be distinguished perfectly
over a 1 minute window.
Smart glasses like J!NS MEME are very unobtrusive and
can be easily confused with ”normal” glasses. Yet, the
question is if this type of devices can produce a high
enough signal quality to be used for complex activity
recognition systems. Of course, the verdict is still out, yet,
our initial results are very positive, indicating the potential
of smart glasses for ubiquitous computing applications.
Figure 4: The confusion matrix
in percent for the frame-by-frame
classification using a 6 sec.
sliding window (accuracy 70 % ).
Acknowledgements
We would like to thank the research department of J!NS
for supplying us with prototypes. This is work is partly
supported by the CREST project.
References
[1] Bulling, A., and Roggen, D. Recognition of visual
memory recall processes using eye movement analysis.
In Proceedings of the 13th international conference on
Ubiquitous computing, ACM (2011), 455–464.
[2] Bulling, A., Ward, J. A., and Gellersen, H. Multimodal
Recognition of Reading Activity in Transit Using
Body-Worn Sensors. ACM Trans. on Applied
Perception 9, 1 (2012), 2:1–2:21.
[3] Bulling, A., Ward, J. A., Gellersen, H., and Tr¨oster, G.
Eye Movement Analysis for Activity Recognition Using
Electrooculography. IEEE Trans. on Pattern Analysis
and Machine Intelligence 33, 4 (Apr. 2011), 741–753.
[4] Eivazi, S., Bednarik, R., Tukiainen, M., von und zu
Fraunberg, M., Leinonen, V., and J¨a¨askel¨ainen, J. E.
Gaze behaviour of expert and novice
microneurosurgeons differs during observations of
tumor removal recordings. In ETRA, ACM (New York,
NY, USA, 2012), 377–380.
[5] Ishimaru, S., Kunze, K., Kise, K., Weppner, J.,
Dengel, A., Lukowicz, P., and Bulling, A. In the blink
of an eye: combining head motion and eye blink
frequency for activity recognition with google glass. In
Proceedings of the 5th Augmented Human
International Conference, ACM (2014), 15.
[6] Kunze, K., Iwamura, M., Kise, K., Uchida, S., and
Omachi, S. Activity recognition for the mind: Toward
a cognitive quantified self. Computer 46, 10 (2013),
0105–108.
[7] Kunze, K., Kawaichi, H., Yoshimura, K., and Kise, K.
The wordmeter – estimating the number of words read
using document image retrieval and mobile eye
tracking. In Proc. ICDAR 2013 (2013).
[8] Kunze, K., Utsumi, Y., Shiga, Y., Kise, K., and
Bulling, A. I know what you are reading: recognition
of document types using mobile eye tracking. In
Proceedings of the 17th annual international
symposium on International symposium on wearable
computers, ACM (2013), 113–116.
[9] Manabe, H., and Fukumoto, M. Full-time wearable
headphone-type gaze detector. In CHI’06 Extended
Abstracts on Human Factors in Computing Systems,
ACM (2006), 1073–1078.