![](https://csdnimg.cn/release/download_crawler_static/87895234/bg5.jpg)
86:4 • S. Yue et al.
We measure the exact location of both subjects with respect to the radio using the Bosch GLM50 laser distance
measurement tool [
5
]. We then make the FMCW radio and antenna array zoom in on the two voxels in space that
are centered on the chest of each subject. This gives us the best estimates of the subjects’ breathing signals based
on past work, under the assumption that past work is able to perfectly localize the subjects. We plot those RF
signals in the middle column in Fig. 1. The gure shows that by looking at RF reections from the two subjects, it
seems that person 2 has experienced an apnea event – i.e., a period during which he did not breath.
Now let us look at the actual breathing signals from the two belts on the subjects, which we show in the third
column in Fig. 1. The ground truth shows that both subjects have regular breathing and neither of them had an
apnea event, which contradicts the behavior shown in the RF signals reected o their chests. This can lead to
serious errors in reporting a person’s health condition.
This experiment shows that even if one zooms in on the signal from the exact location of the person’s chest,
one cannot eliminate the eect of the signal reected o a nearby person. The reason we cannot completely
separate the two signals is that antenna arrays and FMCW are lters over space. However, no lter is perfect. In
particular, both antenna arrays and FMCW lter signals using the Fourier transform [
1
,
23
]. Since the Fourier
transform is applied over a nite window, it will result in a sinc in the frequency domain. This causes nearby
signals to mix with each other due to the sinc tail, as shown in Fig. 2. The closer the two people are, the closer
the frequency response after taking the Fourier Transform, and the more the corresponding sinc functions leak
into each other.
(a) Two sine waves with dierent frequency (b) Windowed Fourier Transform result
Fig. 2. A sinc example. Fig. 2a shows two sine waves with slightly dierent frequency. The black box on top of the signals
indicates the Fourier Transform window. Fig. 2b shows the result of the Fourier Transform. The two signals are sincs in
frequency domain, and when the frequencies of the signals are similar, their sincs are close to each other.
3 RELATED WORK
Our work is related to past work on RF-based sensing and the general literature on blind source separation.
3.1 RF-Based Sensing
Recent years have witnessed much interest in using RF signals to monitor people’s movements and vital signs
without any on-body sensors. Researchers have developed systems that transmit low power RF signal and monitor
its reections. They used these reections to infer a person’s location [
1
,
18
], gait [
15
,
44
], breathing [
2
,
9
,
14
,
43
],
heart rate [2, 9] and even emotions [48].
Our work is closest to past work that uses RF signals to monitor breathing. Studies using Doppler Radio [
31
],
FMCW [
2
,
14
,
49
], millimeter waves [
9
,
45
], and WiFi signals [
21
,
43
] all demonstrate accurate monitoring of
a single person’s breathing. Respiration monitoring using radio signal is also closely related to monitoring
breathing using acoustic signals, which tends to use FMCW techniques and demonstrates good accuracy for a
Proc. ACM Interact. Mob. Wearable Ubiquitous Technol., Vol. 2, No. 2, Article 86. Publication date: June 2018.