Ferret: RFID Localization for Pervasive Multimedia

1
Ferret RFID Localization for Pervasive Multimedia

• Xiaotao Liu,
• Mark Corner, Prashant Shenoy

University of Massachusetts, Amherst
2
Scenario Ive Lost my Keys

• People frequently misplace common items
• books, keys, tools, clothing, etc.
• difficult due to the sheer scale we interact
  with gt1000s of items
• Need a system to find objects quickly and
  efficiently
• then tell the user where the object is

3
Problems

• Tracking objects can be broken into sub-problems
• Locate find position, perhaps not exact, but a
  general idea
• Store keep object locations in a convenient
  place
• Update when objects move, need to change store
• Display Present locations to user in a helpful
  way

4
Solution Ferret

• Provides a real-time augmented reality service
• locates, stores, updates, and displays object
  locations
• intended for nomadic objects not mobile ones
• Leverage passive RFID, multimedia, and location
  systems
• passive RFID inexpensive, scalable,
  maintenance-free
• multimedia systems provide convenient display
  and storage
• location systems bootstrap process of finding
  locations
• Goal is to pack all functions into a hand held
  device
• including RFID detection, storage, and display
• a combination of video camera and RFID reader

5
Outline

• Motivation and Applications
• Overview of Use
• Design of Ferret
• Sensor model
• Offline location algorithm
• Online location algorithm
• Display
• In paper Storage, Update for nomadic objects
• Prototype implementation
• Experiments
• Speed and accuracy
• Robustness to different movement patterns
• Related Work
• Conclusions

6
Overview of Operation

• User selects some object(s) that she is looking
  for
• She wanders around a room, or building, holding
  Ferret system
• During this process, the reader scans for nearby
  RFID tags
• Ferret detects the RFID tag of interest,
  localizes tag
• It then displays an outline of where the object
  is on the screen
• willing to settle for a probable region of where
  the object is
• depend on human skill to find the exact location
• refine region as system runs
• present improved results in real-time

7
RFID Localization
2. use RF energy to charge up

• Passive RFID tags are not self-locating
• Instead we depend on the handheld to locate tags
• Passive RFID tags have significant error rates
• false negatives are frequent
• false positives due to reflections
• Locate using probabilistic model
• inspired by Hähnel et. al

RFID reader
8
Bayesian Probability Model

• Goal p(xD1n) Probability of tag at x given
  readings
• Initially, without readings, p(xD0) is uniformly
  distributed
• Assume we have p(xD1n)
• Positive reading
• p(Dn1Truex)
• Bayes rule p(xD1n1) a p(xD1n) p(Dn1x)
• a normalization factor
• Similarly, for negative readings
• p(Dn1Falsex) 1 - p(Dn1Truex)

9
Tag Detection Probability
Manually measure probability of detecting tag
(positive reading) p(D Truex) x tags
position
10
Ferret Localization Algorithm ( reading)

• Multiple readings come from user mobility,
  previous, or shared readings

11
Ferret Localization Algorithm (- reading)
Repeated intersection of positive and negative
readings
12
Offline Algorithm Complexity

• We refer to the previous algorithm as the
  offline algorithm
• Each or - reading Ferret performs O(n3)
  operations
• n is the number of sample points
• it must rotate, translate the RFID sensor model
• multiply each sample point against every other
  sample point
• must do this for each object!
• Computational requirements at least 0.7s on a
  laptop
• reader is producing at least 4 readings per
  second
• some readings include multiple objects
• Algorithm most useful for back-annotating video

13
Online Algorithm

• To address real-time concerns use an online
  algorithm
• instead of intersecting all interior points, just
  find convex intersection
• only uses positive readings, not negative ones
  (keeps shape convex!)
• Complexity reduced to O(n2) or 6ms per reading

14
Display

• Each RFID location is a 3-D shape
• To display we simply project this 3-D shape onto
  a 2-D screen

15
Ferret Prototype

• ThingMagic Mercury4 RFID reader
• 30dBm (1 Watt), monostatic circular antenna
• Alien Technology M RFID Tag
• EPC Class 1, 915 MHz
• Sony Motion Eye web-camera
• 320x240 at 12fps
• Cricket Ultrasound 3-D locationing system
• global location not necessary, but need relative
  locations at least
• Sparton SP3003 Digital Compass
• Pan, tilt, and roll
• Software
• translate between coordinate systems, rotate, and
  display

16
Ferret Prototype
Built-in Camera
Cricket locationing sensor
Compass
ThingMagic RFID reader
RFID antenna
17
Evaluation

• Evaluation metrics
• Size of location region for many objects
• Speed of localization for a particular object
• Robustness of localization to mobility patterns
• Evaluation setup for many objects
• Place 30 objects with passive tags around the
  room
• Move Ferret system around the room by human for
  20 minutes
• CDF of localization over 30 objects
• Evaluation setup for single object
• Place single object in room with passive tag
• Move Ferret system in and out of view randomly
  and using a specific pattern
• Size of localization after some amount of time

18
Online Vs Offline (CDF-30 Objects)
Offline algorithm outperforms online, but most
objects localized to 0.2 m3
19
Refinement Relative Volume (1 Object)

• Volume size drops down 100 times to 0.02m3 in 2
  mins
• When starting with previous readings,
  localization is faster

20
Refinement Relative Projection Area
Final projection area decreases 33 times in 2
mins to a 54 pixel diameter circle
21
Different Movement Patterns
Circular motion pattern performs the worst no
diversity in views Offline algorithms advantage
comes from negative readings so head-on and
circular perform similarly
22
Related Work

• Grown out of our work on Sensor Enhanced Video
  Annotation
• SEVA ACM Multimedia 2005 (Best Paper Award)
• Used active sensors for location
• RFID Localization inspired by techniques from
  Hähnel et. al
• 2-D sensor model, application of Bayes rule
  positive readings
• we add 3-D model, negative readings, and online
  technique
• focuses on SLAM/localizing reader, we focus on
  reverse
• LANDMARC and SpotON RFID locationing
• active RFID and signal strength

23
Conclusions

• Ferret a scalable, RFID-based, augmented reality
  system
• localize objects augmented with passive RFID tags
• display probable location regions to a user in
  real-time
• Uses two algorithms online and offline
• both are accurate and efficient (localizes
  objects to 0.2m3 in minutes)
• robust to a variety of user mobility patterns
• Ferret lays the ground work for other augmented
  reality applications

24
Ferret RFID Localization for Pervasive Multimedia

• Xiaotao Liu,
• Mark Corner, Prashant Shenoy

University of Massachusetts, Amherst
25
(No Transcript)
26
Location Storage

• Locations (3-Dimensional probability maps)
• Storage on reader
• simple to implement, but must acquire readings as
  it goes
• Database
• any Ferret readers can take advantage of prior
  knowledge
• also permits offline searching, but
  privacy/authorization concerns
• Storage on writable tags
• tags self-locating and provide locations to
  non-Ferret systems

27
What if objects move?

• Nomadic objects may have moved since previous
  readings
• when online algorithm detects empty intersection,
  reset
• offline algorithm more complex, uses a
  probability threshold

28
Ferret Software Architecture
Ferret System
Visualization Module (modified from FFmpeg)
Intercept original display function
Display projection boundary
Use optics model
Compute projection of location estimates
Fuse video, tags location together
Deal with large amount of data, Optimized for
real-time usage
Bayesian Locationing Module
Video Recording
via TCP, Use SQL-like language
RFID Module (operate RFID reader)
Device Drivers for Cricket and Compass
29
Hähnel et. al

• To each of the randomly chosen potential
  positions we
• assign a numerical value storing the posterior
  probability
• p(x z1t) that this position corresponds to the
  true pose of
• the tag. Whenever the robot detects a tag, the
  posterior is
• updated according to Equation (1) and using the
  sensor model
• described in the previous section.
• In this paper we analyze whether recent Radio
  Frequency Identification (RFID) technology can be
  used to improve the localization of mobile robots
  and persons in their environment.