Human Sensing Using Visible Light Communication
Tianxing Li, Chuankai An, Zhao Tian, Andrew T. Campbell, and Xia Zhou
Department of Computer Science, Dartmouth College, Hanover, NH
{tianxing, chuankai, tianzhao, campbell, xia}@cs.dartmouth.edu
ABSTRACT
We present LiSense, t he first-of-its-kind system that enables both
data communication and
fine-grained, real-time human skeleton re-
construction
using Visible Light Communication (VLC). LiSense
uses shadows created by the human body from blocked light and
reconstructs 3D human skeleton postures in real time. We over-
come two key challenges to r ealize shadow-based human sensing.
First, multiple lights on the ceiling lead to diminished and complex
shadow patterns on the floor. We design light beacons enabled by
VLC to separate light rays from different light sources and recover
the shadow pattern cast by each individual light. Second, we design
an efficient inference algorithm to reconstruct user postures using
2D shadow information with a limited resolution collected by pho-
todiodes embedded in the floor. We build a 3 m × 3 m LiSense
testbed using off-the-shelf LEDs and photodiodes. Experiments
show that LiSense
reconstructs the 3D user skeleton at 60 Hz in
real time wit h
10
◦
mean angular error
for five body joints.
Categories and Subject Descriptors
C.2.1 [Network Architecture and Design]: Wireless communica-
tion
Keywords
Visible light communication; sensing; skeleton reconstruction
1. INTRODU C TION
Light plays a multifaceted role (e.g., illumination, energy source)
in our life. Advances on Visible Light Communication (VLC) [30,
59] add a new dimension to the list: data communication. VLC
encodes data into light intensity changes at a high f requency im-
perceptible to human eyes. Unlike conventional RF radio systems
that require complex signal processing, VLC uses low- cost, energy-
efficient Light Emitting Diodes (LEDs) to transmit data. Any de-
vices equipped with light sensors (photodiodes) can recover data by
monitoring light changes. VLC has a number of appealing proper-
ties. It reuses existing lighting infrastructure, operates on an un-
regulated spectrum band with bandwidth 10K times greater than
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 distributed
for profit or commercial advantage and that copies bear this notice and the full cita-
tion on the first pa ge. Copyrights for components of this work owned by others than
ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or re-
publish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from Permissions @acm.org.
MobiCom’15, September 7–11, 2015, Paris, France.
c
2015 AC M. ISBN 978-1-4503-3619-2/15/09 ...$15.00.
DOI: http://dx.doi.org/10.1145/2789168.2790110.
Figure 1: Shadow cast by varying manikin postures under an LED
light (CREE XM-L). In this scaled-down table-top testbed, the manikin
is 33 cm in height, and the LED light is 22 cm above the manikin.
the RF spectrum, and importantly, is secure (i.e. , does not pene-
trate walls, resisting eavesdropping), energy-efficient, and free of
electromagnetic interference.
In t his paper, we push the envelope further and ask: Can l ight
turn into a ubiquitous sensing medium that tracks what we do and
senses how we behave? Envision a smart space (e.g., home, office,
gym) that takes the advantage of the ubiquity of light as a medium
that integrates data communication and human sensing [71]. Smart
devices (e.g., smart glasses, smart watches, smartphones) equipped
with photodiodes communicate using VLC. More importantly, light
also serves as a passive sensing medium. Users can continuously
gesture and interact with appliances and objects in a room (e.g.,
a wall mounted display, computers, doors, window s, coffee ma-
chine), similar to using the Kinect [57] or Wii in front of a TV, but
there are no cameras (high-fidelity sensors with privacy concerns)
monitoring users, neither any on-body devices or sensors that users
have to constantly wear or carry [27, 62], just LED lights on the
ceiling and photodiodes on the floor.
The key idea driving light sensing is strikingly simple: shadows.
Any opaque object (e.g., human body) obstructs a light beam, re-
sulting in a silhouette behind the object. Because the wavelength
of visible light is measured in nanometers, any macroscopic object
can completely block the light beam, much more significant than
the radio frequencies [16, 50, 65, 70]. The shadow cast on the floor
is essentially a two-dimensional projection of the 3D object. As
the object moves and changes its shape, different light beams are
blocked and the projected shadow changes at light speed – the same
principle as the shadow puppet. Thus, by analyzing a continuous
stream of shadow s cast on the floor, we can infer a user’s posture
and track her behavior. As a simple illustration, Figure 1 shows the
shadow shapes for varying manikin postures under a single light.
We present LiSense, the first-of-its-kind system that enables fine-
grained, real-time user
skeleton reconstruction
1
at a high frame
rate (60 Hz)
using visible light communication. LiSense consists
1
We define skeleton reconstruction as calculating the vectors of the
skeleton body segments in the 3D space.
of VLC-enabled LED lights on the ceiling and low-cost photodi-
odes on the floor.
2
LiSense aggregates the light intensity data from
photodiodes, recovers the shadow cast by individual LED light,
and continuously
reconstructs a user’s skeleton posture in real time.
LiSense’s ability of performing 3D skeleton reconstruction in real
time puts l ittle constraints on the range of gestures and behaviors
that LiSense can sense, which sets a key departure from existing
work that eit her targets a limited set of gestures [3, 17, 46] or only
tracks user’s 2D movements [2, 4, 10]. More importantly, by i nte-
grating both data communication and human sensing into the ubiq-
uitous light, LiSense also fundamentally differs from vision-based
skeleton tracking systems (e.g., Kinect) t hat are built solely for the
sensing purpose. In addition, these systems rely on cameras to cap-
ture high-resolution video frames, which bring privacy concerns as
the raw camera data can be leaked to the adversary [52, 64]. While
prior vision methods [55, 56] have leveraged shadow to infer hu-
man gestures, they work strictly under a single light source and do
not apply in a natural indoor setting with multi ple light sources.
LiSense overcomes two key challenges to realize shadow-based
light sensing: 1) Shadow Acquisition: Acquiring shadows using
low-cost photodiodes is challenging in practice. In the presence of
multiple light sources, light rays from different directions cast a di-
luted composite shadow, which is more complex than a shadow cast
by a single light source. A shadow can also be greatly i nfluenced
by ambient light (e.g., sunlight). Both factors limit the ability of
photodiodes detecting the light intensity drop inside a shadow. To
address this challenge, LiSense leverages the fact that each light is
an active transmitter using VLC and designs light beacons to sepa-
rate light rays from individual LEDs and ambient light. Each LED
emits light beacons by transmitting (i.e., flashing) at a unique fre-
quency. LiSense transforms the light intensity perceived by each
photodiode over time to the f requency domain. By monitoring fre-
quency power changes, LiS ense detects whether the photodiode is
suddenly blocked from an LED and aggregates the detection results
from all photodiodes to recover the shadow map cast by each light.
2) Shadow-based
Skeleton Reconstruction: Shadow maps mea-
sured by photodiodes are 2D projections w ith a limited resolution
(constrained by t he photodiode density). Such low-resolution, im-
perfect shadow images pose significant challenges to
reconstruct a
user’s 3D skeleton
. Existing computer vision algorithms [11, 19,
21, 24, 42, 66, 51] cannot be directly applied to this problem be-
cause they all deal wit h video frames in a higher resolution and are
often augmented with the depth information. LiSense overcomes
this challenge by combining shadows cast by light sources in dif-
ferent directions and infers the 3D vectors of key body segments
that best match shadow maps. LiSense fine-tunes the inferences
using a Kalman filter to take into account movement continuity and
to further reduce
the skeleton reconstruction errors.
LiSense Testbed . We build a 3 m × 3 m LiSense testbed (Fig-
ure 9), using five commercial LED lights, 324 low-cost, off-the-
shelf photodiodes, 29 micro-controllers, and a server. We i mple-
ment light beacons by programming the micro-controllers that mod-
ulate LEDs. We implement blockage detection and 3D
skeleton
reconstruction
algorithms on the server, which generates a st r eam
of shadow maps and continuously tracks user gestures. The
recon-
struction
results are visualized in real time using an animated user
skeleton
(Figure 11). We t est our system with 20 gestures and seven
users in diverse settings. Our key findings are as follows:
• LiSense reconstructs a user’s 3D skeleton with the average an-
gular error of
10
◦
for five key body joints;
2
Engineering photodiodes on the floor sounds labor-intensive to-
day, but it can be eased by smart fabric [1, 45] (see more in § 7).
• LiSense generates shadow maps in real time. It is able to produce
shadow maps of all LEDs every 11.8 ms, reaching the same level
of capturing video frames yet without using cameras;
• LiSense tracks user gestures in real time. It reconstructs the user
skeleton within 16 ms based on five shadow maps, thus generat-
ing 60 reconstructed postures, each of which consists of the 3D
vectors of five key body segments. The reconstructed skeleton is
displayed in real t ime (60 FPS), similar to playing a video at a
high frame rate;
• LiSense is robust in diverse ambient light settings (morning, noon,
and night) and users with different body sizes and shapes.
Contributions. We make the f oll owing contributions:
• We propose for the first time the concept of continuous user
skeleton reconstruction based on visible light communication,
which enables light to be a medium for both communication and
passive human sensing;
• We design algorithms to extract the shadow of each individual
light source and reconstruct 3D human skeleton posture continu-
ously using only a stream of low-resolution shadow information;
• We build the first testbed implementing real-time, human skele-
ton reconstruction based on VLC, using off-the-shelf, low - cost
LEDs, photodiodes, and micro-controllers in an indoor environ-
ment;
• Using our testbed, we test our system with diverse gestures and
demonstrate that it can reconstruct a user skeleton continuously
in real time with small reconstruction angular errors.
Our work takes the first step to go beyond conventional radio
spectrum and demonstrates the potential of using visible light spec-
trum for both communication and fine-grained human sensing. We
believe that with its unbounded bandwidth, light holds great poten-
tial to mitigate the spectrum crunch crisis. By expanding t he ap-
plications VLC can enable, we hope that our work can trigger new
radical thinkings on VLC applications. Our work examines the in-
terplay between wireless networking, computer vision, and HCI,
opening the gate to new paradigms of user interaction designs.
2. LIGHT SHADOW EFFECT
Shadow is a common phenomenon we observe everyday. It is
easily recognizable under a single light source by unaided human
eyes. Our goal is to understand whether off-the-shelf, low-cost pho-
todiodes can reliably detect the light intensity drop in the shadow. If
so, we can deploy them on the floor and aggregate their li ght inten-
sity data to obtain the shadow cast by a human body. In this section,
we first study the impact of a blocking object on light propagation
using low-cost photodiodes. We then examine the challenges of
shadow-based analysis in t he presence of multiple lights.
2.1 Experiments on Blocking the Light
Consider a single photodiode on the floor, we hypothesize that
if any opaque object stands in the direct path between the point
light source and the photodiode, the photodiode will not be able
to perceive any light coming from this point light source. Thus,
the photodiode observes a light intensity drop compared to the case
when there is no object blocking its direct path to the light source.
To confirm our hypothesis, we build a scaled-down table-top
testbed (Figure 2) using commercial LED lights (CREE XM-L) and
low-cost photodiodes (Honeywell SD3410-001). We set up a sin-
gle LED chipset as the point light source at 55 cm height and place
the photodiode directly below the light. By default we calibrate the
photodiode’s location using a plumb bob to ensure 0
◦
of light’s in-
cidence angle. The photodiode has 90
◦
field of vision (FoV), i.e., it
reector
PD
(a) Default setup ( left) and setup
w/ reflector (right)
LED
PD
resistor
(b) LE D and photodiode (PD) w/
Arduino boards
Figure 2: Experiment setup with an
LED and a ph otodiode (a), both at-
tached to micro-controllers (b).
0
200
400
600
800
0 50 100 150 200 250
Arduino readings
Light intensity(lux)
(a) Light intensity vs Arduino
reading
0
100
200
300
400
500
20 40 60 80
Arduino readings
LED duty cycle (%)
w/o blocking
w/ blocking
Ambient light
(b) Light source light intensity
0
100
200
300
400
500
0.1 0.2 0.3 0.4 0.5
Arduino readings
Distance between blocking object and PD (m)
w/o blocking
w/ blocking
Ambient light
(c) Blocking object distance
0
100
200
300
400
500
0 10 20 30 40 50
Arduino readings
Light incident angle (degree)
w/o blocking
w/ blocking
Ambient light
(d) I ncident angle
0
100
200
300
400
500
Morning Noon Night
Arduino readings
Ambient light (lux)
w/o blocking
w/ blocking
Ambient light
(e) Ambient light
0
100
200
300
400
500
None Clothes Paper Plastic Foam Tin
Arduino readings
Reflector material
w/o blocking
w/ blocking
Ambient light
(f) Multi-path
Figure 3: Experiments on blocking the light using our scaled-down table-top testbed. (a) shows that the mea-
sured Arduino reading is directly proportional to the perceived light intensity, where the slope decreases after
the photodiode enters its saturation range. (b)-(f) show the impact of blockage on the measured Arduino reading
under varying settings.
can sense incoming light with incidence angle within 45
◦
. To fetch
signal continuously from the photodiode, we cascade the photodi-
ode and a resistor (10 KΩ) and measure the resistor voltage using a
micro-controller (Arduino DUE in Figure 2(b)). It maps the mea-
sured voltage to an integer between 0 and 1023. Since the photodi-
ode’s output current is directly proportional to the perceived light
intensity, resistor voltage (thus the Arduino reading) reflects the
perceived light intensity. Using a light meter (EXTECH 401036)
we have verified that the Arduino reading is directly proportional
to the perceived li ght intensity (Figure 3(a)). To understand the im-
pact of blockage, we place a 10 cm × 10 cm × 2 cm wood plate
between the photodiode and the LED, and compare the Arduino
readings before and after placing the wood plate
3
. We aim to an-
swer the following key questions:
Q1: Is the shadow dependent on t he light intensity of the light
source? We first examine how the light source brightness affects
the photodiode’s sensing data upon blockage. We connect the LED
to an Arduino UNO board to vary the LED’s duty cycle from 10%
to 90% (Figure 2(b)), resulting in light intensities from 5 to 30 lux
perceived by t he photodiode. For a given duty cycle, we record the
average Arduino reading before and after blocking the LED and
plot the results in Figure 3(b). We observe that upon blockage, the
Arduino’s reading reports only the ambient light in all duty cycle
settings, meaning that an opaque object completely blocks the light
rays r egardless of the brightness of the light source.
Q2: Does the distance between the blocking object and the light
source matter? Next, we test whether the relative distance between
the blocking object and the light source affects the photodiode’s
sensed light intensity. To do so, we fix the LED’s duty cycle to
90%, move the wood plate along the line between the LED and
the photodiode, and record the Arduino data at each distance. Fig-
ure 3(c) shows that as long as the object stays in the direct path
between the LED and the photodiode, the light beam is completely
blocked regardless of the r el at ive distance of the blocking object.
Q3: How does the light incidence angle come into play? Be-
cause a photodiode has a limited viewing angle and can perceive
3
We also measured the blockage impact using different body parts
(e.g., arms, hands) of a manikin and observed similar results.
incoming light only within its FoV, we further examine whether
it can detect blockage under varying light incidence angles. We
move the photodiode horizontally with 10-cm intervals and record
the Arduino’s reading before and after blockage. As expected, the
perceived light intensity gradually drops as t he photodiode moves
further away from the LED (Figure 3(d)). More importantly, at all
locations (incidence angles), the light beam blockage result in a sig-
nificant drop in the Arduino’s reading. The drop is less significant
when the incidence angle approaches half of the photodiode’s FoV.
This is because the photodiode can barely sense any light coming
at its FoV and thus blocking the light beam has a negligible impact.
Q4: What is the impact of ambient light? We also perform our
measurements during different time of a day as the ambient light
varies. In Figure 3(e), we plot the Arduino reading before and af-
ter blockage as the ambient light intensity increases from 2 to 100
lux. In all conditions, we observe a significant drop in the Arduino
reading. Because the photodiode senses a combination of the am-
bient light and the light from the LED, its perceived light intensity
increases as the ambient light intensity increases.
Q5: How significant is the light multi-path effect? Would it di-
minish the shadow? Visible light is diffusive in nature. While a
object blocks the direct path between the photodiode and the LED,
light rays can bounce off surrounding objects and reach t he pho-
todiode f r om multiple directions. Since the photodiode perceives
a combination of light rays coming in all directions, this multi-
path effect can potentially r educe the light intensity drop caused
by blocking the direct path. To examine the impact of the multi-
path effect, we place a flat board vertically close to the LED to
increase the r eflected light rays (Figure 2(a), right) and record the
Arduino’s reading with and without blocking the direct path to the
LED. Among all types of material we have tested, the significant
drop in the Arduino’s reading is consistent (Figure 3(f)). Thus, light
in the direct path dominates the perceived light intensity. The tin
has the highest light intensity because of its minimal energy loss
during reflection.
Overall, our experiment results confirm that opaque objects can
effectively block li ght in diverse settings and the blockage can be
detected by low-cost photodiodes under a single point light source.
(a) 2 LEDs (b) 3 LEDs (c) 4 LEDs (d) 5 LEDs (e) Setup w/ PD
0
200
400
600
800
1000
1 LED 2 LEDs 3 LEDs 4 LEDs 5 LEDs
Arduino readings
w/o blocking
w/ blocking
(f) Sensed li ght intensity
Figure 4: Shadow cast by multiple LEDs on the table-top testbed (a)-(d). We further measure the light intensity change caused by blockage using
off-the-shelf photodiode (e). Light intensity drop caused by blockage is less significant under more LEDs (f).
2.2 Where is My Shadow?
Detecting a shadow is relatively st raightforward under a single
point light source. However, when there are multiple light sources
present, shadow detection becomes much more challenging. This
is because light rays fr om different sources result in a composite
shadow, which comprises shadow components created and fused
by multiple light sources. Figure 4 illustrates the resulting shadow
of the manikin as we switch on more LED lights in our table-top
testbed. We make two key observati ons. First, the shadow is grad-
ually diluted as more LEDs are switched on. This is because there
are more light rays coming from different directions, and hence
blocking a beam from one LED does not necessarily block t he
light beams from other LEDs, leaving a fading shadow. Second,
the shadow shape becomes more complex under more LEDs as a
result of superimposing different shadows cast by individual LEDs.
As a result, it becomes harder to infer the manikin’s posture based
upon the r esulti ng tangled shape pattern.
While visually less noticeable, a shadow is also much harder
to detect under these conditions using off-the-shelf photodiodes.
In our experiments, we place the photodiode in a shadow region
caused by the manikin’s posture, gradually switch on more LEDs,
and compare the Arduino’s reading with and without the manikin’s
blockage (Figure 4(e)). We observe that as more LEDs are switched
on, more light rays coming in different directions hit the shadow re-
gion and thus the perceived light intensity level ri ses. Furthermore,
once three or more LED lights are switched on, the photodiode
enters the saturation region (Figure 3(a)), thus blocking light rays
from a single LED has a negligible impact on the Arduino reading.
As a result, detecting shadow using these low-cost photodiodes is
very challenging in practice under multiple lights.
In the next two sections, we introduce LiSense, which disam-
biguates composite shadow s using VLC and continuously tracks
user posture in real time.
3. DISAMBIGUATING SHADOWS
To disambiguate composite shadows created by multiple lights,
LiSense recovers the shadow shape, referred to as the shadow map,
resulting from each individual light source. Specifically, the shadow
map associated with light source L
k
is the shadow if only light
source L
k
is present. We describe this as disambiguating a com-
posite shadow. The key challenge is that each photodiode perceives
a combination of light rays coming from different light sources and
cannot separate light rays purely based on the perceived light inten-
sity (Figure 5(a)(b)).
To overcome this technical barrier, we leverage the fact that each
LED light is an active transmitter using VLC. We instrument each
light source to emit a unique light beacon, implemented by modu-
lating the light intensity changes at a given frequency. By assigning
a different frequency to each light source, we enable photodiodes
to differentiate lights from different sources. This allows LiSense
to recover the shadow cast by each individual l ight.
In this section, we first describe our design of light beacons, fol-
lowed by the mechanism to detect blockage (shadow) and infer
shadow maps.
3.1 Separating Light Using Light Beacons
The design of light beacons is driven by the observation that
while the perceived light intensity represents the sum of all incom-
ing light rays, these light rays can be separated in the frequency do-
main if they flash at different frequencies. That is, if we transform a
time series of perceived light intensity to the frequency domain us-
ing Fast Fourier Transform (FFT), we can observe frequency power
peaks at the frequencies at which these light rays flash. Figure 5
shows an example with two LED lights flashing at 2.5 kHz and
1.87 kHz, respectively, in an unsynchronized manner. The light in-
tensity perceived by the photodiode is a combination of these two
light pulse waves, and yet FFT can decompose the light mixture
and generate peaks at the two flashing frequencies. Thus, a light
beacon can be implemented by programming each light source to
flash at a unique frequency.
Benefits. Light beacons bring three key benefits when consid-
ering blockage detection. First, by examining the resulting fre-
quency power peaks after applying FFT, we can separate light r ays
from different light sources. The frequency power at frequency f
i
is approximately directly proportional to the intensity of light rays
flashing at f
i
. Thus, the changes in power peaks allow the photo-
diode to determine which lights are blocked. Second, light beacons
also allow us to avoid interference from ambient light sources by
applying a high pass filter (HPF). This is because the change of am-
bient light is random and generates frequency components close to
zero in the frequency domain. Third, by separating light rays from
different sources, we observe a much more significant drop in the
frequency power caused by blocking a light, which is the key to
achieving robust detection of blockage, especially when the photo-
diode perceives a weak light intensity because of a long distance or
a large incidence angle.
Light Beacon Frequency Selection. Designing robust light bea-
cons, however, is nontrivial, mainly because selecting the flashing
frequency for each light source is challenging. Specifically, assume
an LED flashes as a pulse wave with a duty cycle of D and flashing
frequency of f , the Fourier series expansion of this pulse wave is:
f(t) = D +
∞
X
n=1
2
nπ
sin(πnD) cos(2πnft). (1)
It indicates that the power emitted by the pulse wave is decomposed
into the main frequency power, which is the first AC component
when n = 1, and an infinite number of harmonics (components
425
450
475
0 1 2 3 4 5 6
Time (ms)
Arduino readings
L
1
, 1.87 kHz
375
400
425
450
L
2
, 2.5 kHz
(a) A single LED on
400
420
440
460
480
500
0 1 2 3 4 5 6
Arduino readings
Time (ms)
L
1
+ L
2
(b) Two LEDs on
0
150
300
450
600
0 2 4 6 8 10
Frequency power
Frequency (kHz)
L
1
L
2
Harmonics
(c) After applying FFT
0
150
300
450
600
0 2 4 6 8 10
Frequency power
Frequency (kHz)
L
1
L
2
Harmonics
(d) After blocking L
2
Figure 5: Experiments with a photodiode (PD) and two LEDs (L
1
and L
2
, 50% duty cycle) that flash at different frequencies. (a)-(b) show the PD’s
readings when only one LED is on and when both are on. PD perceives a combination of light rays, which however can b e separated in the frequency
domain after applying FFT (c). The frequency power of the flashing frequency f
i
reflects the perceived intensity of light rays flashing at f
i
. Thus the
power peak at 2.5 kHz disappears after L
2
is blocked (d).
Algorithm 1: Selecting the candidates of flashing frequency for
light beacons.
input : 1) R, signal sampling rate; 2) A, the number of FFT points; 3)
f
f licker
, the threshold to avoid flickering; 4) f
interval
, the
minimal interval between adjacent flashing frequencies
output: f
candidate
, flashing frequency candidates for all LEDs
f
candidate
= {
R
A
× ⌈
f
flicker
×A
R
⌉}
for k ← ⌈
f
flicker
×A
R
⌉ + 1 to
A
2
do
f
k
=
R
A
× k
val i d = true
for f
s
∈ f
candidate
do
if (f
s
mod f
k
) == 0) OR (|f
s
− f
k
| ≤ f
interval
) then
val i d = false
break
end
if valid then f
candidate
← f
k
∪ f
candidate
end
with n > 1). Hence, an LED light L
i
flashing at frequency f
i
leads to not only a global power peak (main frequency power) at
frequency f
i
, but also small local power peaks at all the harmon-
ics frequencies (Figure 5(c)). In other words, if the perceived light
intensity from light L
i
changes, it will affect not only the main fre-
quency power at f, but also the power peaks at harmonics. To sepa-
rate out the light rays and avoid interference across lights, we need
to ensure that the harmonics do not overlap with the main frequen-
cies of other lights. Tracking all harmonics is infeasible. In our
design, we focus on the top-ten harmonics frequency components.
This is because the harmonics frequency power drops significantly
as n increases. We observe it becomes negligible once n > 10.
Furthermore, the flashing frequencies need to satisfy three addi-
tional constraints. First, since lights are also used for illumination,
the flashing frequencies need to be above a threshold f
f licker
(1
kHz in our implementation) to avoid the flickering problem [32,
36, 48]. Second, the highest flashing frequency is l imited by the
sampling rate of the micro-controller fetching data from photodi-
odes. The Nyquist Shannon sampling theorem [25] says that it has
to be no larger than R/2, where R is sampling rate. Finally, the
adjacent frequencies have to be at least f
interval
away to ensure
robust detection of frequency power peaks. We set f
interval
=
200 H z based on a prior study [32]. Algorithm 1 details the pro-
cedure to select all candidate flashing frequencies satisfying all the
above constraints.
We then assign the candidate flashing frequencies to all LED
lights, such that the lights within each photodiode’s viewing an-
gle (FoV) flash at different frequencies. Since photodiodes have a
limited FoV (90 degrees for Honeywell SD3410-001), each photo-
diode perceives only a small subset of all lights. Thus we do not
need a large number of candidate frequencies to cover all lights and
the system can easily scale up to more lights. Supporting denser
lights requires more candidate flashing fr equencies, which can be
achieved by increasing the signal sampling rate R.
Light Beacon Overhead. Light beacons can be seamlessly inte-
grated into existing VLC systems, enabling light to fulfill a dual
role of data communication and human sensing. For VLC sys-
tems [5, 14 , 36, 48] that use Frequency Shift Keying (FSK) to
modulate data, all the data packets serve as light beacons, as long
as LED lights within the FoV of a photodiode use different flash-
ing frequencies to modulate data. For VLC systems that use other
modulation schemes [34, 35], we can instrument each LED light to
emit light beacons periodically, in the same way that Wi-Fi access
points periodically transmit beacons. For an ADC sampling rate
of 20 kHz and modulation window of 128 points, a light beacon
lasting for 6.4 ms is sufficient for the photodiode to separate light
rays. Thus, the overhead of transmitting light beacons is negligible
given that a data packet typically lasts for hundreds of ms based on
the IEEE 802.15.7 standard [8].
3.2 Blockage Detection
We detect blockage by t r ansforming the time series of light in-
tensity values of light beacons to the frequency domain and ex-
amining the frequency power changes. Specifically, the intensity
of light r ays from light L
i
flashing at frequency f
i
is represented
by the frequency power of f
i
. If an opaque object blocks the di-
rect path from light L
i
to a photodiode, the frequency power of f
i
changes (Figure 5(d)).
To examine the impact of blockage on the f requency power, we
mount five commercial LED lights (Cree CXA25) on an office ceil-
ing (2.65 m in height), attach them to an Arduino UNO board,
which modulates the Pulse Width Modulation (PWM) of each light
to allow each LED to emit light beacons at a given frequency (Ta-
ble 1 by running Algorithm 1).
We place photodiodes (Figure 2(b))
at 324 locations in a 3 m x 3 m area on the floor. Each photodiode
can perceive light rays from all LED lights. We then measure the
readings of the Ardunio controllers connected to the photodiodes
for 6.4 ms before and after blocking each LED l ight.
Figure 6 shows the CDF of relative frequency power change.
Assume the P
ij
(t) is the frequency power of f
i
(the flashing fre-
quency of the light beacons from light L
i
) at time t perceived by
the photodiode at location p
j
, its relative frequency power change
∆P
ij
(t) is defined as:
∆P
ij
(t) = |
P
nonBlock
ij
− P
ij
(t)
P
nonBlock
ij
|, (2)
