Advancing Hand Gesture Recognition with

High Resolution Electrical Impedance Tomography

Yang Zhang Robert Xiao Chris Harrison
Carnegie Mellon University, Human-Computer Interaction Institute
5000 Forbes Avenue, Pittsburgh, PA 15213
{yang.zhang, brx, chris.harrison}@cs.cmu.edu
Tomo provided a glimpse into the applicability of EIT for
ABSTRACT
Electrical Impedance Tomography (EIT) was recently em- input sensing. To further expand the feasibility of this
ployed in the HCI domain to detect hand gestures using an promsing technique, we sought to explore technical
instrumented smartwatch. This prior work demonstrated improvements to low-cost, worn, EIT sensing. The most
great promise for non-invasive, high accuracy recognition obvious parameter that can be varied is the number of elec-
of gestures for interactive control. We introduce a new sys- trodes. Intuitively, more electrodes will produce a denser
mesh of sensed paths (Figure 1), which should yield a supe-
tem that offers improved sampling speed and resolution. In
rior reconstructed image.
turn, this enables superior interior reconstruction and ges-
ture recognition. More importantly, we use our new system In addition to varying the number of electrodes, EIT sys-
as a vehicle for experimentation we compare two EIT tems can also choose between two very different types of
sensing methods and three different electrode resolutions. EIT sensing (see e.g., [1,11] for more details). The first
Results from in-depth empirical evaluations and a user method is two-pole sensing, in which impedance meas-
study shed light on the future feasibility of EIT for sensing urements are captured from each pair of skin electrodes
human input. one acting as an emitter and the other as a receiver. This
Author Keywords method is affected by skin impedance [1,8,20] and so larger
Electrical Impedance Tomography; EIT; hand gestures; electrodes are typically used for greater contact area with
smartwatch; bio-impedance; biometrics; input. the skin, precluding dense electrode arrays. However, this
approach is popular due to its technical simplicity.
ACM Classification Keywords
H.5.2. [Information interfaces and presentation]: User inter- Alternatively, EIT systems can use a more sophisticated
faces Input devices and strategies. four-pole scheme, which excites an adjacent pair of elec-
trodes with an AC signal and measures the voltage between
INTRODUCTION another pair of electrodes (Figure 5). This process is repeat-
Tomography is an imaging technique that estimates the ed for all possible emission and measurement pairs. This
cross-sectional interior structure of objects through the use differential measurement approach makes four-pole sensing
of an external, penetrating signal [27]. In this work, we use less sensitive to contact conditions at the skin [1,15].
Electrical Impedance Tomography (EIT) [10,22], which
uses surface electrodes and high frequency alternating cur- We built our EIT system such that it could toggle between
rent (AC) to measure internal electrical impedance. By two- and four-pole sensing schemes, as well as 8-, 16- and
placing many electrodes around an object, it is possible to 32-electrode configurations (i.e., six EIT configurations in
reconstruct the internal impedance distribution and infer the
interior structure [10]. EIT is safe for long-term continuous
operation (low voltage, no ionizing radiation), non-invasive
to the wearer (rests on skin without the need for e.g., con-
ductive gel) and can be made inexpensive (~$50). For these
reasons, we recently adapted the technology for use in a
gesture-sensing smartwatch called Tomo [43].

Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distribut-
ed for profit or commercial advantage and that copies bear this notice and the full
citation on the first page. Copyrights for components of this work owned by others
than ACM must be honored. Abstracting with credit is permitted. To copy other-
wise, or republish, to post on servers or to redistribute to lists, requires prior specif-
ic permission and/or a fee. Request permissions from permissions@acm.org.

UIST 2016, October 16-19, 2016, Tokyo, Japan Figure 1. The number of sensed paths (grey lines) dramatically
2016 ACM. ISBN 978-1-4503-4189-9/16/10$15.00 increases as electrode count grows (red dots). For reference,
DOI: http://dx.doi.org/10.1145/2984511.2984574 Tomo [43] uses a two-pole, 8-electrode scheme (upper left).

843
total). We used this setup as a vehicle for experimentation, to obtain the conductivity image of the interior of an ob-
allowing different EIT schemes to be readily and directly ject. The interior is discretized using a finite element meth-
compared. As we will discuss, we used a variety of con- od to generate a mesh, and the conductivity at each mesh
trolled baselines to compare and discuss performance element is computed. Generally, finer meshes provide high-
tradeoffs and accuracy gains. We also replicated the user er resolution output images but increase the computational
study in Tomo [43] to see how hand gesture recognition requirements.
accuracy varies across our sensing conditions.
Early algorithms were based on linear back-projection (e.g.,
RELATED WORK [9]), which is commonly used for PET and CT image re-
Our work broadly intersects with two large research do- construction (using gamma and X-rays respectively). This
mains: 1) hand gesture sensing from a wrist- or arm-worn technique assumes that the electrical current travels approx-
device, and 2) electrical impedance tomography. imately along certain fixed equipotential lines. However, as
Wrist Gesture Sensing this does not accurately capture the complete 3D movement
Humans naturally gesture with their hands, which can form of the electric signals, accuracy can be limited [42]. Tomo
many static poses and kinetic gestures. For this reason, they [43] employed this method to obtain computationally inex-
have long been studied for controlling interactive systems. pensive and straightforward image reconstruction.
For example, Digits [28] reconstructed a 3D model of the More recent methods use least-squares optimization to find
hand using a wrist worn camera, while Way et al. [40] used the best-fit image [38]. To solve this nonlinear problem,
a time-of-flight camera to sense small free-hand gestures. systems often employ Newton-Raphson or Gauss-Newton
Similarly, zSense [39] used infrared sensors deployed on a iteration, which are computationally expensive. For real-
ring or smartwatch to detect both static and dynamic poses. time use, sophisticated single-iteration methods, such as the
Other systems have leveraged IMUs on the wrist to detect maximum a posteriori estimator [2], are able to produce
dynamic hand motions such as finger rubbing and hand acceptable images rapidly after performing significant pre-
waving [29,41]. Researchers have also utilized arm contour computation. We use the latter approach.
changes that occur when performing different hand ges-
tures, including sensing capacitance [32], pressure [13,26], Applications of Electrical Impedance Tomography
and proximity [17]. Electrical Impedance Tomography has been applied to
many application domains. For example, due to its safe and
More similar to our work are systems that use bio-sensing. non-invasive nature, it has been widely used in clinical
For example, electromyography (EMG) senses the electri- applications such as sensing lung ventilation [35], brain
cal signals produced by muscle activation. Saponas et al. function [25] and blood flow [19]. EIT has also seen use in
[33,34] built an EMG-based system that supported 4 pinch geophysics (e.g., monitoring mountain permafrost [23]),
gestures at 79% accuracy. Another popular approach is bio- environmental science (e.g., underground pollutant
acoustics, which use vibrations that propagate through the detection [12]), biology (e.g., tree inspection [36] and
body upon performing hand gestures. For example, Ham- industrial monitoring (e.g., measuring liquid flow [16]).
bone [14] used contact microphones to detect 4 gestures
with ~90% accuracy. The Sound of One Hand [3] used a Our previous Tomo system [43] showed that two-pole EIT
similar setup to detect finger gestures such as rub, tap, and could be made compact and low cost, making it more ame-
flick. Skinput [21] also leveraged bioacoustics to detect 8 nable for use in consumer electronics. Tomo used an off-
flicking/pinching gestures at 87.3% accuracy. the-shelf AD5933 impedance converter chip for measuring
bio-impedance, which takes 3.6 milliseconds to perform
EIT Image Reconstruction one measurement. Sensing all combinations of its 8 elec-
Tomographic image reconstruction is well studied in the trodes (28 pairs) to generate a single image frame takes
signal processing literature, and a number of popular algo- roughly 100 ms, yielding a frame rate of 10 Hz. Unfortu-
rithms exist. A good overview of current approaches can be nately, this hardware cannot easily scale to 16 and 32 elec-
found in [38]. The basic goal of EIT image reconstruction is trode configurations, which have 120 and 496 electrode
pairings respectively (see Table 1).
As we will describe in greater detail in the next section, our
new EIT system uses both custom hardware and software,
granting us tight control over the entire sensing pipeline. In
contrast to Tomo, our system can complete a single meas-
urement in just 0.33 ms, which allows us to achieve
100 frames per second in a two-pole, 8-electrode configura-
tion. More interestingly, it enables us to scale to greater
Table 1. Performance characteristics of Tomo and our new numbers of electrodes (Table 1). This, in turn, offers supe-
setup. We extrapolate hypothetical performance (grey region) rior sensing accuracy, as we discuss in our evaluation.
for 16 and 32 electrode versions of Tomo.

844
 Figure 3. Our EIT sensing boards.
Figure 2. Our electrode band (left) and our EIT sensor
worn on a users arm (right). voltage between the VCCS output and the receiver elec-
trode. In this case, the voltage measurement pair is the same
IMPLEMENTATION as the signal projection pair. In the four-pole scheme, the
Our setup has two main components: a wristband (Figure voltage measurement pairs are different from the current
2), which features an array of electrodes to be worn by a projecting pair. Note the system captures the voltage differ-
user, and a custom-designed data capture and processing ence between electrodes (i.e., we do not compare to ground).
board. Our board is built around a Cortex-M4-based Teensy
3.2 system-on-chip (SoC), augmented with a voltage con- Analog Sampling
trolled current source (VCCS), direct digital synthesis We first amplify our signal with a preamp. The gain value
(DDS) IC, and ADC preamp (Figure 3). The board also is adjusted to maintain a consistent dynamic range when
features multiplexers that allow for dynamic electrode se- switching between four-pole and two-pole measurement
lection, enabling different EIT configurations. A schematic schemes. We also use a high pass filter with a 15.6 kHz
view of our system is shown in Figure 4. The total cost of cutoff frequency to remove ambient EMI (e.g., from power-
our hardware is $80, which could be made both smaller and line noise). The input signal is then biased by AVDD/2
less expensive in a high volume commercial application. (1.65 V) and sampled by our microprocessors ADC at
2 MHz with 12-bit resolution.
Wristband
We made a leather wristband with 32 evenly spaced stain- Data Acquisition
less steel electrodes (Figure 2). Each electrode measures Once the multiplexer has selected the appropriate electrodes,
3.4 15.3 mm. The average human forearm diameter is we wait 100 s for the DC bias on AC coupling capacitor to
roughly 190 mm in circumference, suggesting an upper- stabilize. We then collect 250 samples, or roughly five pe-
bound electrode width of ~6 mm in a 32-electrode configu- riods of the 40 kHz excitation signal (collecting multiple
ration. The band is secured to the user with a Velcro strap. periods to reduce noise). The root-mean-square (RMS) of
the samples is calculated to form a single measurement. The
Excitation Signal sensor then moves to the next measurement, reconfiguring
We use an AD5930 [7] DDS IC and an AD8220-based the multiplexers accordingly. After it collects all values for
VCCS [5] to generate the EIT excitation signal. The the current frame, it sends the RMS measurements to a lap-
AD5930 is configured to output 40 kHz sinusoidal waves top over Bluetooth. The number of measurements and our
(the same frequency used in Tomo [43]). This signal is then systems frame rate can be found in Table 1.
fed into the VCCS to output a constant 300 A AC current
(06 Vpp depending on the load impedance). Two-Pole and Four-Pole Measurement Schemes
As noted in the introduction, the simplest EIT setup uses a
Multiplexing two-pole scheme. In each measurement, one pair of elec-
Two 32-to-1 multiplexers (ADG732 [6]) connect the VCCS trodes is used for both signal emission and voltage meas-
terminals to any two electrodes, forming the signal- urement, and all pairs are tested for a total of Ne choose 2 =
projection pair. Two more multiplexers connect the preamp Ne(Ne 1)/2 measurements.
buffer terminals to two electrodes to form the voltage-
measuring pair. In two-pole EIT sensing, we measure the

Figure 5. Two projection rounds in four-pole measure-

Figure 4. A schematic view of our system, ment scheme with 8 electrodes. Higher voltage difference
illustrated with 8 electrodes. is shown with brighter color.

845
In four-pole sensing, separate pairs are used for signal pro- cally, we downsampled the reconstructed images to 16 16
jection and voltage measurement. The Adjacent Drive pixels, and used the resulting 256 data points as raw fea-
method [11,24] is the most common projection pattern the tures. We used a support vector machine (SVM) implemen-
signal is applied through one pair of adjacent electrodes and tation provided by the Weka Toolkit [18] for classification
the voltage difference is measured between the Ne 3 other (SMO; polynomial kernel with default parameters).
adjacent electrode pairs. This is repeated for all emitter
FIDELITY EXPERIMENTS
pairs, resulting in a total of Ne(Ne 3) measurements.
We ran a series of experiments to evaluate the fidelity of
Figure 5A illustrates the first signal projection round for the our six different EIT configurations. These were designed
four-pole scheme. The signal is emitted using electrodes 1 to elicit fidelity characteristics of the tomographic output.
and 2, and the voltage differences V1, V2, V5 are meas- Apparatus
ured sequentially with five electrode pairs: 3-4, 4-5, 7-8. In order to analyze the quality of our tomographic recon-
In the second projection round (Figure 5B), the next pair (2- struction, we needed a stable and known reference input.
3) is used for emission, and the voltage differences are For this, we used a 20 cm diameter acrylic cylinder filled
measured sequentially with another five electrode pairs. with 3 cm of saline water (9000 ppm NaCl, approximating
This process is repeated sequentially until a full loop as the conductivity of human tissue). We used acrylic shapes
been completed (8 (8 3) = 40 measurements). to make different EIT phantoms [15,19,30] for testing;
Baseline Calibration the impedance of these shapes is approximately 1017 times
Although we endeavored to make our sensing board and that of saline water at 25C.
wristband as consistent as possible, there are nonetheless Number of Interior Features
small variations in the pathways to different electrodes. To We varied the number of the cylinders (2.25 diameter) in
prevent this from impacting our tomography, we must ob- the bath from one to five to see how well different EIT con-
tain a baseline measurement of the impedance between all figurations can resolve distinct objects (Figure 6). As one
pairs of electrodes. With 928 electrode pairings in our four- would expect, as resolution increases, the ability to discern
pole configuration, it was impractical to perform these cali- dense constellations of objects improves. For example, in
brations manually (as was done for Tomos [43] 28 pairs). both two- and four-pole sensing schemes, five objects are
Instead, we calibrated our setup by wrapping the wristband not discernable when using 8 electrodes. Even at 16 elec-
around a homogenous, electrically conductive material and trodes, it is hard to separate them. However, with 32 elec-
capturing one frame of data using our standard sensing trodes, all five objects can be readily segmented. Overall, at
pipeline to serve as a baseline. We experimented with a each electrode level, four-pole offers improved fidelity
variety of materials: ground beef in Saran wrap, lean pork (particularly with respect to background uniformity) com-
chop, cylinder of ice, saltwater bath and Jell-O, with the pared to two-pole.
latter producing the best results.
Size of Interior Features
Interior Image Reconstruction Resolving the size of features is also important, as the hu-
We performed all interior image reconstruction on a 13 man body has bones and muscles of varying cross-sectional
2015 MacBook Pro with a 2.7 GHz Intel Core i5 processor. size. Additionally, we wished to see if small features were
The reconstruction algorithm was derived from the detectable. To explore this, we tested our EIT configura-
EIDORS toolkit [37], which provides a large library of dif-
ferent solvers. After preliminary experimentation, we chose
the nodal one-step Gauss-Newton iterative solver, which
produces a maximum a posteriori (MAP) estimate [2] of the
conductivity at each node of the finite element mesh.
This algorithm is parameterized by a single hyperparameter
, which controls the smoothing of the output. In all of our
experiments, we fix the hyperparameter value at =0.03 for
the two-pole configuration and =0.001 for the four pole
configuration. The solver requires a precomputation step,
which takes about three seconds on our laptop; subsequent
image reconstruction is carried out as a single matrix multi-
plication, taking 2.4 ms per image (numbers provided for
32 electrodes, the most computationally expensive case).
Hand Gesture Classification
Since interior image reconstruction is sufficiently fast to be
used in real-time system, we chose to derive machine learn-
ing features purely from the reconstructed images. Specifi- Figure 6. Reconstructed images of bath
with different numbers of objects

846
 Figure 7. Reconstructed images of bath
with objects of varying sizes.

Figure 8. Reconstructed images of bath

tions with circular reference shapes of three different sizes
with different object shapes.
(1.5, 2.25 and 3). We used all combinations of these
three objects as reference inputs. Output is shown in Figure Biological Reference
7, showing the significant fidelity improvement offered by While the various reference shapes used above are useful
four-pole sensing. Specifically, both relative and absolute for controlled testing, they do not approximate biological
object sizes are clearly distinguishable in the 16 and 32 tissue. Unfortunately, no one at our institution was willing
electrode versions of our four-pole setup, while the two- to give us a cross-sectional slice of their arm to use as a
pole setup has difficulty even separating the two objects. ground truth. Instead, as a biological reference, we used a
Shape of Interior Features cross-cut lamb shoulder (Figure 9, left). This closely emu-
Unlike directional waves (e.g., X-rays), electric current is lates human tissue, and further offers reference features
not confined to a path ray or a plane, and thus a change in with more subtle impedance differences (unlike salt water
impedance anywhere in the domain (not just along the path and plastic, which are polar opposites with respect to con-
ray) will affect the measurement. Furthermore, this soft ductivity). As can be seen in Figure 9, two-pole sensing
field problem [30] causes the electrical path to curve away struggles with reconstruction at any electrode resolution, in
from objects along the path, resulting in incomplete occlu- line with our previous results. In particular, two-pole has
sion (unlike the perfect occlusion from e.g. an optical mesh trouble localizing the high impendence bone when it is
system [31]). Thus, shape recovery is particularly challeng- roughly in the middle, yielding a bulls-eye-like impedance
ing for EIT systems. distribution. Four-pole does better, and with 32 electrodes,
the bone and fat are both visible.
To test this property in our EIT setup, we used paired small
USER STUDY
and large squares, triangles, cylinders and rectangles (small
The ultimate objective for our EIT system is interactive
size 1.5; large 3), as seen in Figure 8. As with our size
control. To evaluate the feasibility of the different EIT con-
experiment, two-pole sensing only discerns that two objects
figurations we support, we ran a hand gesture recognition
are present, but little other detail is visible. Four-pole per-
user study. In order to compare our results with Tomo [43],
forms substantially better, with steady improvements in
fidelity as electrode resolution grows. However, even four- we adopted its experimental procedure and gesture set
pole reconstruction with 32 electrodes does not resolve (which were themselves chosen to be comparable to prior
sharp edges (e.g., our triangle reference shape). We are very work). Specifically, we used a finger pinch gesture set:
likely approaching the upper limit of what can be recon- Index Pinch, Middle Pinch, Ring Pinch, and Little Pinch,
and a hand gesture set, designed around coarse hand mo-
structed using low-cost EIT.
tions: Fist, Stretch, Right, Left, Thumbs Up, Spider-Man
and Index Pinch. A relax gesture was included in both sets
as the neutral state, resulting in a total of five pinch gestures
and eight hand gestures.
Participants
We recruited 10 participants (four female), with a mean age
of 23. The study took approximately thirty minutes to com-
plete and paid $20. All of our participants were right hand-
ed, so the sensor wristband was fitted to participants left
arm (where a watch would usually be). Participants were
seated for the duration of the study with their left elbow
Figure 9. Reconstructed images of a cross-cut lamb shoulder resting on a table. After participants felt comfortable, data
with different EIT configurations. collection began.

847
Procedure
We first collected data using four-pole sensing. Participants
were asked to perform all 11 gestures, one gesture at a time,
in a random order. While holding a gesture pose, the system
automatically switched between 8, 16 and 32 electrode con-
figurations, with ten trials recorded for each (for a total of
30 trials). This process took 15 seconds and our laptop
beeped to indicate the participant could relax. We then Figure 10. Accuracies for three hand gesture sets (left: Hand,
switched to two-pole sensing mode, and repeated the same middle: Pinch, right: Combined) across different EIT configu-
procedure. Thus, a round of data collection captured 10 rations. Numbers below bars are electrode count.
trials 11 gestures 3 electrode resolutions 2 EIT allow for even higher resolution arrays. Our results show a
schemes = 660 gesture trials. clear upward and sustained improvement in accuracy as
In total, we ran five rounds of data collection, yielding electrode count grows, suggesting that moving to e.g., 64
33,000 gesture trials (660 trials per round 5 rounds 10 electrodes could offer further improvements in accuracy.
participants). To estimate the gesture recognition accuracy However, new technical insights would be needed to
of our EIT system, we performed a post hoc, leave-one-out achieve interactive speeds. Additionally, given that EIT has
train/test experiment. Specifically, we trained our classifier a soft-field problem, there are bottlenecks on sensing reso-
on four rounds of a users data, and tested on the fifth (all lution that cannot be resolved by simply increasing elec-
trode count. For example, some gestures involve similar
combinations). This ensured that any two data points close
muscle groups and movements, and thus are hard to differ-
in time (which naturally tend to be more similar) were ei-
ther in the test set or the training set, but never both. entiate irrespective of resolution.

RESULTS AND DISCUSSION Currently, we only recognize discrete hand gestures. How-
There are several high level conclusions to be drawn from ever, we have analog data of muscle movement inside the
our results. First, increasing electrode count strictly im- arm, and so we believe it may be possible to build a contin-
proves recognition accuracy (roughly +3.8% when moving uous kinetic model using regression. In addition, with our
from 8 to 16, and +1.5% when moving from 16 to 32). Sec- improved sample rate and tomographic resolution, it may
ondly, four-pole handily out-performs two-pole sensing (on be possible to capture biometric data (e.g., smartwatches
average, 6.8% better across all electrode resolutions and that auto-authenticate when worn) and bio-factors (e.g.,
gesture sets). Interestingly, four-pole recognition accuracies bone density, pulse, and blood pressure). Medical-grade
do not dip when combining hand and pinch gesture sets, EIT systems have shown this to be possible [8,19].
which suggest they are complementary. Unsurprisingly, our To facilitate comparison to prior work, we deliberately
best performing configuration is four-pole sensing with 32 adopted the gesture set used in Tomo [43]. However, our
electrodes; on the 11-gesture hand+pinch set, it achieves a system is not limited to these gestures, and it is likely that
mean accuracy of 94.3% across our ten participants a other gestures can be detected. Alternate gesture sets may
67.8% reduction in gesture recognition error vs. Tomo [43]. even provide better recognition accuracies, and future work
Figure 10 provides an overview of these results. Overall, could explore a wider variety of gestures to find those with
our results suggest EIT is competitive with more mature distinctive EIT signatures. Further extending Tomo, future
bio-sensing input methods, including bioacoustics (e.g., [3, iterations could explore other sensing locations, including
14,21]) and electromyography (see e.g., [33,34]). the face, throat, legs and feet, which would require new
Although four-pole sensing with 32-electrodes yielded the electrode form factors.
best accuracy, at 3 FPS, it is less suitable for interactive Finally, there are also technical improvements we wish to
control. Instead, four-pole, 16-electrode sensing at 16 FPS explore. For example, using (or combining) different exci-
might be the right balance between accuracy and interac- tation frequencies could yield better reconstructions [35].
tivity. Of course, this framerate limitation is due to our pre- Swept frequency excitations may also be possible. We also
sent prototype, and is not inherent in the sensing approach. want to try the auto rotation approach proposed by Amma
Future EIT systems could achieve higher framerates by et al. [4], which demonstrated a high-density EMG sensor
using higher frequency excitation signals, faster ADC sam- band that automatically detected the orientation with re-
pling, higher bandwidth communication, and similar. spect to the arm. This orientation data was used to digitally
CONCLUSION AND FUTURE WORK rotate the signal in order to provide robust classification no
A significant consequence of increasing electrode resolu- matter how the user wore the device.
tion is that the electrodes must also be smaller, reducing the ACKNOWLEDGMENTS
contact area with the skin. This makes the electrodes more This research was generously supported by the David and
susceptible to variation in skin contact condition, which Lucile Packard Foundation and a Google Faculty Research
will impair tomography techniques. Fortunately, four-pole Award.
sensing is more robust to this effect, and would therefore

848
REFERENCES No.00CH37276) 1, May: 577579.
1. Ackmann, J.J. 1993. Complex bioelectric impedance http://doi.org/10.1109/IMTC.2002.1006906
measurement system for the frequency range from 5
Hz to 1 MHz. Annals of biomedical engineering 21, 2: 13. Dementyev, A. and Paradiso, J.A. 2014. WristFlex:
135146. http://doi.org/10.1007/BF02367609 Low-Power Gesture Input with Wrist-Worn Pressure
Sensors. In Proc. UIST '14, 161166.
2. Adler, A. and Guardo, R. 1996. Electrical impedance http://doi.org/10.1145/2642918.2647396
tomography: Regularized imaging and contrast detec-
tion. IEEE Transactions on Medical Imaging 15, 2: 14. Deyle, T., Palinko, S., Poole, E.S. and Starner, T. 2007.
170179. http://doi.org/10.1109/42.491418 Hambone: A bio-acoustic gesture interface. In Proc.
ISWC '07, 310.
3. Amento, B., Hill, W. and Terveen, L. 2002. The sound http://doi.org/10.1109/ISWC.2007.4373768
of one hand. In Proc. CHI EA '02, 724.
http://dx.doi.org/10.1145/506443.506566 15. Dowrick, T., Blochet, C. and Holder, D. 2015. In vivo
bioimpedance measurement of healthy and ischaemic
4. Amma, C., Krings, T., Ber, J. and Schultz, T. 2015. rat brain: implications for stroke imaging using electri-
Advancing Muscle-Computer Interfaces with High- cal impedance tomography. Physiological Measure-
Density Electromyography. In Proc. CHI '15, 929938. ment 36, 6: 12731282. http://doi.org/10.1088/0967-
http://doi.org/10.1145/2702123.2702501 3334/36/6/1273
5. Analog Devices Inc. 2010. JFET Input Instrumentation 16. Dyakowski, T., Jeanmeure, L.F.C. and Jaworski, A.J.
Amplifier with Rail-to-Rail Output in MSOP Package. 2000. Applications of electrical tomography for gas-
Retrieved July 10, 2016 from solids and liquid-solids flows - A review. Powder
http://www.analog.com/media/en/technical- Technology 112, 3: 174192.
documentation/data-sheets/AD8220.pdf http://doi.org/10.1016/S0032-5910(00)00292-8
6. Analog Devices. 2005. 32-Channel Analog Multiplex- 17. Fukui, R., Watanabe, M., Gyota, T., Shimosaka, M.
er. ADG732. Retrieved July 10, 2016 from and Sato, T. 2011. Hand shape classification with a
http://www.analog.com/media/en/technical- wrist contour sensor. In Proc. UbiComp '11, 311-314.
documentation/data-sheets/ADG726_732.pdf http://doi.org/10.1145/2030112.2030154
7. Analog Devices. 2005. Programmable Frequency 18. Hall, M., Frank, E., Holmes, G., Pfahringer, B.,
Sweep and Output Burst Waveform Generator. Reutemann, P. and Witten, I.H. 2009. The WEKA data
AD8220. Retrieved July 10, 2016 from mining software. ACM SIGKDD Explorations 11, 1:
http://www.analog.com/media/en/technical- 1018. http://doi.org/10.1145/1656274.1656278
documentation/data-sheets/AD8220.pdf
19. Halter, R.L., Hartov, A. and Paulsen, K.D. 2008. Imag-
8. Aristovich, K.Y., Packham, B.C., Koo, H., Santos, G. ing Forearm Blood Flow with Pulse-Ox Gated Electri-
S., McEvoy, A. and Holder, D. S. 2016. Imaging fast cal Impedance Tomography. In Proc. EMBS08, 1192
electrical activity in the brain with electrical impedance 1195. http://doi.org/10.1109/IEMBS.2008.4649376
tomography. NeuroImage 124: 204213.
http://doi.org/10.1016/j.neuroimage.2015.08.071 20. Harikumar, R., Prabu, R. and Raghavan, S. 2013. Elec-
trical Impedance Tomography (EIT) and Its Medical
9. Barber, D.C., Brown, B.H. and Avis, N.J. 1992. Image Applications: A Review. International Journal of Soft
Reconstruction in Electrical Impedance Tomography Computing and Engineering (IJSCE) 3, 4: 193198.
Using Filtered Back - Projection. In Proc. EMBS '92,
16911692. 21. Harrison, C., Tan, D. and Morris, D. 2010. Skinput:
http://dx.doi.org/10.1109/IEMBS.1992.5761992 Appropriating the Body as an Input Surface. In Proc.
CHI '10, 453-462.
10. Borcea, L. 2002. Electrical Impedance Tomography. http://doi.org/10.1145/1753326.1753394
Inverse Problems 18, 6: R99R136.
http://doi.org/10.1088/0266-5611/18/6/201 22. Henderson, R. P. and Webster, J. G. 1978. An imped-
ance camera for spatially specific measurements of the
11. Chen, J., Cao, Z. and Xu, L. 2011. Four-terminal thorax. IEEE transactions on bio-medical engineering
scheme used in a two-terminal EIT system. In Proc. 25, 3: 250254.
IMT '11, 16481652. http://doi.org/10.1109/TBME.1978.326329
http://doi.org/10.1109/IMTC.2011.5944264
23. Hilbich, C., Hauck, C., Hoelzle, M., Scherler, M.,
12. DAntona, G., Ferrero, A., Lazzaroni, M., Ottoboni, R. Schudel, L., Vlksch, I., Vonder Mhll, D. and
and Samarani, E. 2002. Active monitoring apparatus Musbacher, R.. 2008. Monitoring mountain perma-
for underground pollutant detection based on electrical frost evolution using electrical resistivity tomography:
impedance tomography. In Proc. IMTC '02, (IEEE Cat. A 7-year study of seasonal, annual, and long-term vari-

849
 ations at Schilthorn, Swiss Alps. Journal of Geophysi- Proc. CHI '08, 515-524.
cal Research: Earth Surface 113, 1: 112. http://doi.org/10.1145/1357054.1357138
http://doi.org/10.1029/2007JF000799
34. Saponas, T.S., Tan, D.S., Morris, D., Balakrishnan, R.,
24. Holder, D. S. 2015. Electrical Impedance Tomography, Turner, J. and Landay, J.A. Enabling always-available
Methods, History and Applications. Institute of Physics input with muscle-computer interfaces. In Proc. UIST
Publishing. CRC Press: 456. '09, 167-176. http://doi.org/10.1145/1622176.1622208
http://doi.org/10.1118/1.1995712
35. Teschner, E., Imhoff, M. and Leonhardt, S. Electrical
25. Holder, D.S. 1992. Electrical impedance tomography Impedance Tomography: The realization of regional
(EIT) of brain function. Brain Topography 5, 2: 8793. ventilation monitoring, 2nd Edition.
http://doi.org/10.1007/BF01129035
36. Treetronic. 2015. PiCUS: Treetronic, Electric Re-
26. Jung, P.G., Lim, G., Kim, S. and Kong, K. 2015. A sistance Tomography. http://www.argus-
Wearable Gesture Recognition Device for Detecting electronic.de/en/content/download/431/4060/.../ TreeT-
Muscular Activities Based on Air-Pressure Sensors. ronic+3+manual.pdf
IEEE Transactions on Industrial Informatics 11, 2:
37. Vauhkonen, M., Lionheart, W.R.B., Heikkinen, L.M.,
485494. http://doi.org/10.1109/TII.2015.2405413
Vauhkonen, P.J. and Kaipio, J.P. A MATLAB package
27. Kalender, W. 2000. Computed Tomograpy: Fundamen- for the EIDORS project to reconstruct two-dimensional
tals, System Technology, Image Quality, Applications. EIT images. Physiological Measurement 22.1
Publicis MCD Verlag. http://dx.doi.org/10.1088/0967-3334/22/1/314
28. Kim, D., Hilliges, O. and Izadi, S. 2012. Digits: Free- 38. Vauhkonen, P.J. 2004. Image Reconstruction in Three-
hand 3D Interactions Anywhere Using a Wrist-Worn Dimensional Electrical Impedance Tomography.
Gloveless Sensor. In Proc. UIST '12, 167-176.
39. Withana, A., Peiris, R., Samarasekara, N. and Na-
http://doi.org/10.1145/2380116.2380139
nayakkara, S. zSense: Enabling Shallow Depth Gesture
29. Laput, G., Xiao, R. and Harrison, C. 2016. ViBand: Recognition for Grater Input Expressivity on Smart
High-Fidelity Bio-Acoustic Sensing Using Commodity Wearables. In Proc. CHI '15, 3661 3670.
Smartwatch Accelerometers. To appear in Proc. UIST http://dx.doi.org/10.1145/2702123.2702371
'16.
40. Way, D. and Paradiso, J. A usability user study con-
30. Lionheart, W.R.B. 2004. EIT Reconstruction Algo- cerning free-hand microgesture and wrist-worn sen-
rithms: Pitfalls, Challenges and Recent Developments. sors. In Proc. BSN '14, 138-142.
Physiological Measurement 25.1 http://dx.doi.org/10.1109/BSN.2014.32
http://doi.org/10.1088/0967-3334/25/1/021
41. Wen, H., Rojas, J.R. and Dey, A.K. Serendipity: Finger
31. Moeller, J. and Kerne, A. 2012. ZeroTouch: an optical Gesture Recognition using an Off-the-shelf Smart-
multi-touch and free-air interaction architecture. In watch. In Proc. CHI '16.
Proc. CHI '12, 2165-2174. http://dx.doi.org/10.1145/2858036.2858466
DOI=http://dx.doi.org/10.1145/2207676.2208368
42. Wu, K., Yang, J. and Dong, X. 2012. Comparative
32. Rekimoto, J. 2001. GestureWrist and GesturePad: Un- study of reconstruction algorithms for electrical imped-
obtrusive Wearable Interaction Devices. In Proc. ISWC ance tomography. IEEE Transactions on Biomedical
'01, 21-27. Engineering, 51077127: 22962299.
http://dx.doi.org/10.1109/ISWC.2001.962092
43. Zhang, Y. and Harrison, C. 2015. Tomo: Wearable,
33. Saponas, T.S., Tan, D.S., Morris, D. and Balakrishnan, Low-Cost Electrical Impedance Tomography for Hand
R. 2008. Demonstrating the feasibility of using forearm Gesture Recognition. In Proc. UIST '15, 167-173.
electromyography for muscle-computer interfaces. In DOI=http://dx.doi.org/10.1145/2807442.2807480

850