# Caraoke: An E-Toll Transponder Network for Smart Cities
Abari et al., ACM SIGCOMM '15

## Motivation/Introduction
- **Background:** RFID is pervasive, everyone has an EZPass, e.g., transponder (mandatory in PA?)
- Deploy RFID readers on street lamps, intersections, everywhere in city (at some cost), as shown in **Figure 3**.
- Interact with e-toll (EZPass) transponders anywhere in an urban environment
- Enables some new **use cases:**
	1. Vehicle flow at traffic intersections
	2. Red light enforcement via localization
	3. Street parking management and billing

### Transponder Background
- *Reader* (high powered multi-Watt radio + directional antenna) sends message to tags, tag replies
- But **no MAC protocol** so common practice is to place readers in non-overlapping locations (spatial reuse)
- **Figure 2:** Shows the reader transmitting the 915 MHz sinewave *Query* signal and the transponder's *Response* data packet
	- Simple radio so carrier frequency offset up to 1.2 MHz (that's one part per thousand -- typical Wi-Fi radio CFOs are 10-30 parts per million)
	- Transponder response is **On-Off Keyed (OOK):** Each symbol period, transmit carrier for a one, or be silent for a zero.
	- Receive picture -- peaks at the CFO value Delta-f (on board)

### Contribution/Key Idea
- Let tags collide when they are read simultaenously but view collision in the **frequency domain** -- use **carrier frequency offset (CFO)** to separate the tags

## Design
- LTE backhaul, solar-powered lamppost-mounted reader

### §5: CFO-Based Sharing
- **Goal:** Estimate the number of vehicles crossing a busy intersection
- Take the FFT of collision (sum of many received tags replying at the same time)
	- **Figure 4:** Most likely, # peaks = # transponders responding

1. Need to look at the **distribution** of the transponder CFOs
2. Need to consider **frequency resolution** of the FFT
	- Inversely proportional to FFT **time window T*
	- Tag response 512 us so frequency resolution = 1.95 KHz
	- That's 615 "bins" over a range of 1.2 MHz
	- Eq. 7 gives prob of an accurate transponder count if we just count peaks in the frequency domain

- **Insight:** Time shifting property of DFT -- time shift in in the time domain transforms with a phase shift linear with frequency, in the frequency domain.

- Receive two signals R(f) and R(f'), then looking at that with a time shift you get R(f)exp{j2pi\*f\*tau} + R(f')exp{j2pi\*f'\*tau}
	- Paper claim: Then magnitude of that bin changes b/c the two frequencies rotate by different amounts and then sum...
- **Algorithm:** Test each non-empty bin using the above time-shifting property, if detect multiple responses in the same bin, count that as **two** responses
- Now a miscount happens when **three or more** transponders end up in same FFT bin
	- Very unlikely (math in Eq. 9)

### §6: Localizing Transponders
- **Goal:** Localize the transponders, to enable smart parking, red-light violations
- Calculate the AoA at a two-antenna reader via the same math as spatial signatures (**Figure 5** and Eq. 10)
- Do this on a per-frequency basis to separate the AoA estimates for each tag (unnumbered equation on p. 302, lower right column)
- Issue: Sensitivity to small perturbations of Delta-phi at the ends of the two antenna array
	- Fix: Use three antennas in a triangle, switch to the pair that doesn't end-face the tag
	- But need a story for either adding a radio OR switching and then re-querying to get more readings (only partially discussed in paper)
- Result (**Figure 7**): A *cone* locus, which intersects the road at a *hyperbola*
	- Two readings from two readers results in two hyperbolas, intersect them for location fix.

### §7: Measuring Speed
- Localize car at two different locations, NTP timesync, compute average speed

### §8: Decoding IDs
- Spectrum is sinc-squared power versus frequency, not white noise
- Nice trick to separate closely-spaced replies in frequency:
	- Estimate channel to one tag
	- Receive another batch of collisions, estimate channel to the same tag and equalize
	- Add the results: the one tag adds coherently, the others are random and average to zero
	- Repeat with even more readings for better results (**Figure 8**)

### §9: Coordinating Multiple Readers
- Preceding was interference between tag replies; now consider two readers initiating queries at the same time
	- This is usually OK: sum of two sinuoids is a sinusoid
- Collision of query with tag reply:
	- Listen for duration of query + reply (this is the slot time), then use CSMA/CA like Ethernet/ALOHA

### §11: Evaluation
- Implemented in hardware, experimented on quiet urban environment streets
- EZPass transponders in real cars, on real street

#### Counting Accuracy
- **Figure 11:** Vary the number of transponders responding, accuracy maintains high
	- Collect traces by measuring each transponder in isolation, using highly directional antenna
	- Then sum up varying numbers of transponders (so know ground truth) to obtain another trace instance

#### Localization Accuracy
- Two readers, 175 localization instances
- Measure **bearing estimation accuracy**, not localization accuracy
	- Accurate to within 4 degrees on average -- low intensity multipath reflections outside, line-of-sight path from car to elevated reader.
	- Measured the MUSIC AoA profile (**Figure 14**) to support this.