IEEE Computational Intelligence Magazine - November 2019 - 35

speed of light, denoted as c, x k(t ) can be repreAccording to our empirical studies, we find that
sented as d (t )/c. Let m represent the wavelength,
where m = f /c . Thus, the phase shift can be
the channel response for the same motion differs
written as e -j2rd (t)/m. Therefore, when a subject
significantly for subcarriers with large sequence
appears at the boundary of the even/odd Fresnumber difference.
nel zone, the dynamic path phase shift Tp is
equal to r and 2r, respectively. As a result, the
combined signal amplitude should be degraded
the phenomenon, we select the 4th and 30th subcarriers for
in the even zones and enhanced in the odd zones, as shown
comparison, where we can observe that the channel response
in Fig. 2(b).
on the latter is too weak to seize the attenuation caused by
A gesture consists of directed motions. For example, a
gestures. In general, the larger the variance of the subcarrier
keystroke consists of two directed motions, i.e., finger movwaveform, the more sensitive it is to the action. Therefore,
ing down and up. We divide the directed motions into two
we select the subcarrier, which has the largest variance in
cases according to whether the motion crosses the Fresnel
our experiments.
zone boundary:
Butterworth Filter. The raw channel data selected may
Case 1: It happens within a single Fresnel zone. In this
contain abnormal samples caused by background noise or
case, the superposition signal should be either monotonically
hardware glitches and thus should be filtered. In the preproincreasing or decreasing.
cessing module, we choose the Butterworth filter to denoise
Case 2: It involves two Fresnel zones. In this case, the
the data. Under our experimental conditions, the speed at
superposition signal does not show such monotonicity.
which we type or move the mouse is usually between 2 and
Fig. 2(a) presents a case study, where one keystroke is cap60 cm/s. One second of the movement may pass through
tured under both cases. When it only happens in a single zone,
approximately 15 Fresnel zones, which will go through 15
i.e., the 3rd zone, the amplitude of the superposition signal is
peaks or nadirs; therefore, the cutoff frequency of the filter
monotonically decreasing. However, when it crosses two zones,
is preferably set to 7.5 Hz.
i.e., the 3rd and 4th zones, the amplitude of the superposition
Signal Segmentation. The major challenge of segmentasignal increases first and then decreases.
tion lies in the short duration of each designated gesture. For
Therefore, it is feasible to keep the micro-gestures within
example, a keystroke usually only lasts for approximately 0.7 s.
some single zones by an appropriate antenna deployment. ParTo ensure real-time recognition, the segmentation algorithm
ticularly, we select the 9th, 10th and 11th zones since their
should be fine-grained, adaptive and light-weight. Here, we
thickness is around 4 cm, close to our target micro-gestures like
leverage the variance of the channel data, which differs
keystrokes and mouse movements.
B. Processing Layer
Heat Map

Subcarrier ID

Amplitude (dBm)

The processing layer processes raw CSI data to point out signal
segments corresponding to micro-gestures. More specifically, it
first selects subcarriers with finer granularity, and then denoise
them using Butterworth filter. Lastly, a light-weight segment
algorithm is designed for finding out the signal segments of
micro-gestures.
Subcarrier Selection. According to [34], different subcarriers have different sensitivities to human motions. This is
because different subcarriers have different central frequencies
and wavelengths. Combining the effect of multipath/shadowing with different frequencies, CSI measurements for one
motion at different subcarriers have different channel responses.
Therefore, it is essential to choose proper subcarriers that can
better capture the designated gestures.
According to our empirical studies, we find that the
channel response for the same motion differs significantly for
subcarriers with large sequence number difference. Fig. 3
shows such an example in which we visualize one experiment involving both typing and mouse moving that lasts for
approximately 10 seconds. It can be clearly seen from the
heatmap that the first 10 subcarriers are much more capable
of preserving gestures than the last 10. To further illustrate

10
20
30

(a)

15
Typing
10
5
0

5
6
Time (s)

Subcarrier 4

Subcarrier 30

(b)
FIGURE 3 Subcarrier difference: In the above figure, the channel
response of the first ten subcarriers is more obvious than others; In
the below figure, for the same typing gesture, subcarrier 30 has a
weaker channel response than subcarrier 4.

NOVEMBER 2019 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE

IEEE Computational Intelligence Magazine - November 2019

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - November 2019