# The DarkLight Rises: Visible Light Communication in the Dark
Tian, Wright, Zhou, ACM MobiCom '16

## Motivation/Introduction
- Most VLC systems involve bright light, easily visible
- **Goal:** Low-rate, low-power, human _imperceptible_ VLC
	- Suitable for mobile devices' limited battery capacity
	- Avoids safety issues with IR, avoids illluminating environment with VLC if this is not desired
- **Main Idea:** Encode information into light pulses that are short (and probably faint) enough to be imperceptible to the human eye, yet long enough to communicate data reliably
	- Human eye response time: 10 ms for cones (color, high light requirement), 100 ms for rods (B&W, low light requirement)
- Authors get credit for building a working system
- Transmitter side (LED)
	- Need rise time **less than** pulse duration so that LED reaches peak intensity
		- Figure 1(b): High-power LEDs have a rise time of about 1-2 $\mu$s, low power LEDs about 500-800 ns
- Receiver side (photodiode)
	- Needs to respond in time to not miss the short pulse
	- Figure 2(b): Shows two alternatives, they choose the **SD5421** for data, which responds in about 200 ns but takes > 1.4 $\mu$s to reach peak response, and has a low gain (so short distance for communication)
		- Missing out on full response to their short data pulses, perhaps?  (Their pulse width, as we'll see later, is 500 ns)
- The issue of ambient light **interference**
	- Switching off/on other lights, sun, reflections etc. are the sources of interference
	- Figure 3: Their short pulsed data appears as the short pulses atop the slower interference
	- Don't have phase information from photodiodes, so no spatial separation techniques

## DarkLight Design (§3)
- Figure 4: They reduce the rise time (marginally) with amplified imput voltage
- They increase the photodiode receiver gain (significantly) using an amplifier (no numbers, though)
- **Modulation (§3.2)**
	- Time divided into _symbols_, symbol time divided into $2^M$ _slots_, each of length $L$ seconds
		- Overlapped Pulse Position Modulation (OPPM): encode bits as which slot the pulse starts in.  The slots' durations overlap (so only rising edge position matters).
- **Demodulation:** Look for the rising edge, so take derivative
	- Issue: Derivative amplifies noise
		- Solution: Smooth the received signal first (Gaussian filter), then take derivative
	- Issue: Need to acquire packet timing
		- Solution: Preamble with three pulses in the first slot of each symbol

### Ambient Light Adaptation (§3.3)
- Transmitter has another photodiode for sensing ambient light level, alongside the LED
- When ambient light is brighter, they use fewer slots per symbol
- Data rate is $M/(2^M\cdot L)$ bits per second
- Duty cycle $d=t_{ON}/(2^M\cdot L)$ and ambient light levels determine visibility
	- So data rate is also equal to $M\cdot d/t_{ON}$
- In brighter ambient light, they decrease $M$, increasing $d$ and data rate (look at the first data rate equation above, and look ahead at **Table 1**)

## DarkLight Networks (§4)
- Light is directional, but overlap will occur.  _Many DarkLight links operating at the same time and interfering within overlap?_
- **Goal:** Receive multiple simultaenous streams
- Receiver-side ADC sampling drift causes variation in rise time offset around 100-200 ns (Figure 6)
- They remember the slot alignments for each data _stream_ (light source) and allocate new streams when they see unaligned slots
	- Misses "colliding" streams (_slot collisions_)

## Prototype Implementation (§5)
- Parameters: $t_{ON}=500$ ns, $L=3.2$ $\mu$s, $t_{symbol}=6.55$ ms, which implies $M=11$, but that's not shown in Table 11 (?)
- Hardware:
	- Transmitter-side LED _for data transmission_: **Cree CXA 2520**
	- Transmitter-side photodiode _for ambient light adaptation_: **OPT101**
	- Receiver-side photodiode _for data reception_: **SD5421**

## Experiments (§6)

### User Perception Study (§6.1)
- Setup DarkLight LED on ceiling with a lampshade, measure human perception: ask people to look directly at the LED and indirectly at objects in the room (evidently people know LED is present)
	- Indirect viewing: Indistinguishable in bright ambient light, distinguishable in low ambient light
	- Directly looking at LED: More distinguishable, down to 65% in bright light

### Single-link Performance (§6.2)
- Test link over a distance of **1.3 meters**, varying $L$ (timeslot length), results in up to 1.8 Kbit/second throughput (Figure 10a)
- Increasing pulse width allows the LED to hit max brightness (Figure 10b) -- is this optimized fully, then?
- Viewing angle is about 15 degrees (Figure 10c)
- The adaptation loop takes 5 seconds to converge (Figure 11b)
- Power consumption
	- Table 5: Transmitter side, FPGA dominates (can move to ASIC)
	- Receiver side, amplifier dominates
	- Appears that transmitter can be made quite low power, receiver perhaps not

### Multi-link Performance (§6.3)
- Figure 13: To maximize throughput, need to increase the slot width from $3.2$ $\mu$s to $32$ $\mu$s -- slowing down by an order of magnitude!
	- When LEDs are synchronized, can space out the slots, otherwise they'll collide
	- Good match with their analytical model for collision frequency