WISP Localization

1
WISP Localization

• Jonathan Huang and Ali Rahimi

2
The Localization Problem

• Track the 3D position and rotation of a WISP
  based on harvested signal strengths

WISP Power Harvester
3
Chair
4
This is difficult

• The mapping from positions to voltages is
  environment specific
• Most environments are too complex to be
  analytically modeled
• Voltage measurements are noisy

5
(No Transcript)
6
(No Transcript)
7
(No Transcript)
8
Approach

• Ultimate goal learn the mapping from positions
  to voltages while tracking
• But for now learn the mapping from ground truth
  obtain by camera system

9
Current Approach

• Collect ground truth for training and evaluation
• Learn the function which maps positions to
  voltage readings
• Use dynamics to track tags

10
Processing the Voltage Readings
Raw Signal (3 Antennas)
11
Processing the Voltage Readings

• The Alien Reader cycles through each antenna.
  Each cycle begins with a spike and ends with a
  long rest
• The received signal is actually a somewhat
  corrupted version of this

(Zoomed In)
12
Processing the Voltage Readings
HMM signal model States long break, start
spike, ant1, ant2, ant3, short
break Observations gtthresh, ltthresh
Parse using Viterbi to separate signal
from spurious spikes, and to get antenna id.
13
Processing the Voltage Readings
Raw Signal (with errors shown)
The most-likely parse as determined by Viterbi
14
Processing the Voltage Readings

• For each antenna, we average all of the readings
  in one frame to obtain the final signal

20 seconds
4 seconds
15
Visual Localization

• Setup

Groundtruthing cameras
Testing and data collection setup
RFID antenna
Lights to simplify camera-based tracking
RFID antenna
16
Calibrating the Camera Setup
Input Images with corner detections
Recovered Camera Pose
17
Input
Thresholded Background Subtraction
Thresholded Brightness
Connected Components Labeling
Estimate the mean of the largest two components
Output
18
Movie
19
Vision Results

• Performance
• 15 fps, processing two time-synced PtGrey
  Dragonfly2 cameras
• Limitations
• Algorithm sensitive to lighting conditions,
  random movements
• Works well when lighting is constant and scene is
  static (except for the lights)
• The known distance between the lights provides a
  sanity check and adds robustness

20
Vision Results
21
Current Approach

• Collect ground truth for training and evaluation
• Learn the function which maps positions to
  voltage readings
• Use dynamics to track tags

22
One Dimensional Localization

• Data Collection
• We waved a WISP in (roughly) a straight line
  between two antennas
• Learning
• We regressed positions against two voltage
  readings using a Gaussian Process model

23
One-Dimensional Dataset
24
1-D GP Regression Results
25
Current Approach

• Collect ground truth for training and evaluation
• Learn the function which maps positions to
  voltage readings
• Use dynamics to track tags

26
Using Dynamics

• We model the pose evolution with stochastic
  linear dynamics
• Use the GP observation model
• Estimate state using Kalman filter

27
1-D Tracking Results
Accuracy 16 cm average error over a range of 3
meters
28
Thank You