Array signal processing: localization

ArrayTrack, NSDI '13

Array signal processing: indoor radar

(Left) Wi-Vi, SIGCOMM '13
(Right) Wi-See, MobiCom '13

Array signal processing: WiFi security

SecureArray, MobiCom '13

Phase measurements are key

Can we do this on every WiFi AP?

WiFi APs are ubiquitous; huge potential for large-scale deployments!

Need phase data usually only available on specialized hardware (e.g. Rice WARP)

Can we do this on every WiFi AP?

Consumer NICs sometimes leak this data, but have other challenges.

Unless accounted for, these APs are useless for signal processing.

Array signal processing: Our goal

• Get phase readings as if they were taken at the antennas
• Have an array with many antennas, and the ability to scale up
• Operate in real-time

Array signal processing: Our goal

• Get phase readings as if they were taken at the antennas
• Have an array with many antennas, and the ability to scale up
• Operate in real-time

Have no fear, Phaser is here!

• Get phase readings as if they were taken at the antennas
• Have an array with many antennas, and the ability to scale up
• Operate in real-time

A single 802.11 NIC is not enough

• A single consumer NIC typically has ≤3 antennas
• But array processing fidelity increases with # of antennas
• 802.11ac will help, although only marginally

Can we combine data from multiple NICs to make One Large Array™?

Complications abound

• NICs measure the channel at different times

  

• Cannot directly compare phase of same sample index!

Complications abound

• NICs measure the channel at different times
• Oscillators are not frequency locked between NICs
  
  

  

  

Complications abound

• NICs measure the channel at different times
• Oscillators are not frequency locked between NICs
• Sampling rate varies between NICs

  

Two observations about DFT

Time-domain shift is subcarrier (\(k\)) dependent rotation!

\(h\left[n - \color{green}{\tau_0}\right] \overset{DFT}{\longleftrightarrow} H[k]e^{ 2\pi j \color{green}{\tau_0} k/N}\)

Time-domain rotation is equal to frequency-domain rotation!

\(h[n]e^{j\color{blue}\rho} \overset{DFT}{\longleftrightarrow} H[k] e^{j\color{blue}\rho}\)

Opportunity!

All inter-card differences are frequency-domain rotations!

Find set of rotations to apply to one card to synchronize with other card

How do we find the right rotation?

Shared information is key

Idea: Introduce a shared antenna using a splitter.

\(\psi\)

• \(\color{blue}{\psi}\) is the same for both cards for shared antenna
• Can simply measure the rotation for each sample!

Oscillator alignment woes

On one NIC, oscillators are frequency-locked, but have phase offsets.

\(\psi_0\) \(=\psi_0\)
\(\psi_1\) \(=\psi_1\color{red}{+\phi_1}\)
\(\psi_2\) \(=\psi_2\color{red}{+\phi_2}\)

$$(\psi_1-\psi_0)\color{red}{+\phi_1} \neq \color{green}{\pi\cos\theta}$$

$$(\psi_2-\psi_0)\color{red}{+\phi_2} \neq \color{brown}{2\pi\cos\theta}$$

Need to find \(\color{red}{\{\phi_i\}}\).

Phase offsets skew AoA estimate

• With estimate of bearing, can search for {φ_i}
• Assuming free-space, correct {φ_i} maximizes energy toward estimated bearing and minimizes energy elsewhere.

Are we done yet?

No, not quite...
Exhaustive search becomes too expensive for larger arrays.

For five antennas, need to find φ₁ through φ₄.
With granularity of 4°, ~65M combinations!

Skip bad combinations early

Some combinations of φ₁ and φ₂ are obviously incorrect.

Can skip these when searching φ_3..n!

Probabilistic insertion

1. Create an empty, finite-size population
2. Draw initial candidates from all combinations of φ₁,φ₂.
3. Probabilistically insert each candidate into population
   • Better AoA match = higher probability
   • Increased diversity (variance) = higher probability
4. New candidates are each surviving φ₁...φ_n-1 + every φ_n
5. Repeat 3-5 until all {φ_i} have been searched

Are we done now?

• No, multipath throws off AoA estimate!
• True bearing may not be known accurately
• {φ_i} with best AoA correspondence is not necessarily the right one!

Beating multipath

Beating multipath

Beating multipath

Beating multipath

• Cluster all final populations
• Distance: Euclidian distance of {φ_i}
• Identify best^* cluster
• Centroids are best estimate of true {φ_i}

^*Highest product of average estimate quality and number of samples in cluster.

Multi-NIC operation

in an Anechoic chamber

AoA with 3 antennas across 800 packets in a controlled environment.

Multi-NIC operation

in an Anechoic chamber

AoA with 5 antennas across 800 packets in a controlled environment.

Accuracy improves by ~80%!

Autocalibration

Four APs in an indoor environment

• Manual cable-based calibration for ground truth
• All APs at same height
• Calibration with five antennas (two NICs)
• Measure phase error for each \(\phi_i\) across APs

Autocalibration

Four APs in an indoor environment

AoA estimate error 2.5° median, 6.5° 95th percentile.

Autocalibration

Four APs in an indoor environment

AoA estimate error 4.5° median, 9.5° 95th percentile.

Contributions

• Automatic phase calibration down to 20° phase error
• A technique for building large arrays with small NICs
• A method for suppressing multipath (in the paper)
• And all this on off-the-shelf hardware!
  • Our goal is not to replace WARPs

Where do we go from here?

• Tune probabilistic insertion and clustering
• Determine when multicard operation breaks
• Switch design to avoid sacrificing an antenna
• Run other array processing-based applications on top of Phaser
  • Wi-Vi
  • SecureArray
  • Object recognition

Phaser in practice