LOCALIZATION

1
LOCALIZATION

• Distributed Embedded Systems
• CS 213/Estrin/Winter 2003
• Speaker Lewis Girod

2
What is Localization

• A mechanism for discovering spatial relationships
  between objects

3
Why is Localization Important?

• Large scale embedded systems introduce many
  fascinating and difficult problems
• This makes them much more difficult to use
• BUT it couples them to the physical world
• Localization measures that coupling, giving raw
  sensor readings a physical context
• Temperature readings ? temperature map
• Asset tagging ? asset tracking
• Smart spaces ? context dependent behavior
• Sensor time series ? coherent beamforming

4
Variety of Applications

• Two applications

Passive habitat monitoring Where is the
bird? What kind of bird is it?
Asset tracking Where is the projector? Why is it
leaving the room?
5
Variety of Application Requirements

• Very different requirements!

• Outdoor operation
• Weather problems
• Bird is not tagged
• Birdcall is characteristic but not exactly known
• Accurate enough to photograph bird
• Infrastructure
• Several acoustic sensors, with known relative
  locations coordination with imaging systems

• Indoor operation
• Multipath problems
• Projector is tagged
• Signals from projector tag can be engineered
• Accurate enough to track through building
• Infrastructure
• Room-granularity tag identification and
  localization coordination with security
  infrastructure

6
Multidimensional Requirement Space

• Granularity Scale
• Accuracy Precision
• Relative vs. Absolute Positioning
• Dynamic vs. Static (Mobile vs. Fixed)
• Cost Form Factor
• Infrastructure Installation Cost
• Communications Requirements
• Environmental Sensitivity
• Cooperative or Passive Target

7
Axes of Application Requirements

• Granularity and scale of measurements
• What is the smallest and largest measurable
  distance?
• e.g. cm/50m (acoustics) vs. m/25000km (GPS)
• Accuracy and precision
• How close is the answer to ground truth
  (accuracy)?
• How consistent are the answers (precision)?
• Relation to established coordinate system
• GPS? Campus map? Building map?
• Dynamics
• Refresh rate? Motion estimation?

8
Axes of Application Requirements

• Cost
• Node cost Power? ? Time?
• Infrastructure cost? Installation cost?
• Form factor
• Baseline of sensor array
• Communications Requirements
• Network topology cluster head vs. local
  determination
• What kind of coordination among nodes?
• Environment
• Indoor? Outdoor? On Mars?
• Is the target known? Is it cooperating?

9
Variety of Mechanisms

• Weve seen a broad spectrum of application
  requirements
• There are also a broad spectrum of localization
  mechanisms appropriate for different applications

10
Returning to our two Applications

• Choice of mechanisms differs

Passive habitat monitoring Minimize environ.
interference No two birds are alike
Asset tracking Controlled environment We know
exactly what tag is like
11
Variety of Localization Mechanisms

• Very different mechanisms indicated!

• Bird is not tagged
• Passive detection of bird presence
• Birdcall is characteristic but not exactly known
• Bird does not have radio TDOA measurement
• Passive target localization
• Requires
• Sophisticated detection
• Coherent beamforming
• Large data transfers

• Projector is tagged
• Projector might know it had moved
• Signals from projector tag can be engineered
• Tag can use radio signal to enable TOF
  measurement
• Cooperative Localization
• Requires
• Basic correlator
• Simple triangulation
• Minimal data transfers

12
Taxonomy of Localization Mechanisms

• Active Localization
• System sends signals to localize target
• Cooperative Localization
• The target cooperates with the system
• Passive Localization
• System deduces location from observation of
  signals that are already present
• Blind Localization
• System deduces location of target without a
  priori knowledge of its characteristics

13
Active Mechanisms

• Non-cooperative
• System emits signal, deduces target location from
  distortions in signal returns
• e.g. radar and reflective sonar systems
• Cooperative Target
• Target emits a signal with known characteristics
  system deduces location by detecting signal
• e.g. ORL Active Bat, GALORE Panel, AHLoS
• Cooperative Infrastructure
• Elements of infrastructure emit signals target
  deduces location from detection of signals
• e.g. GPS, MIT Cricket

14
Passive Mechanisms

• Passive Target Localization
• Signals normally emitted by the target are
  detected (e.g. birdcall)
• Several nodes detect candidate events and
  cooperate to localize it by cross-correlation
• Passive Self-Localization
• A single node estimates distance to a set of
  beacons (e.g. 802.11 bases in RADAR Bahl et
  al., Ricochet in Bulusu et al.)
• Blind Localization
• Passive localization without a priori knowledge
  of target characteristics
• Acoustic blind beamforming (Yao et al.)

15
Active vs. Passive

• Active techniques tend to work best
• Signal is well characterized, can be engineered
  for noise and interference rejection
• Cooperative systems can synchronize with the
  target to enable accurate time-of-flight
  estimation
• Passive techniques
• Detection quality depends on characterization of
  signal
• Time difference of arrivals only must surround
  target with sensors or sensor clusters
• TDOA requires precise knowledge of sensor
  positions
• Blind techniques
• Cross-correlation only may increase
  communication cost
• Tends to detect loudest event.. May not be
  noise immune

16
Building Localization Systems

• Given a set of application requirements, how do
  we build a system that meets them?
• Outline
• Overview of a typical system design
• A quick example
• Ranging technologies
• Coordinate system synthesis techniques
• Spatial scalability
• Recent results the GALORE panel

17
Localization System Components

• Generally speaking, what is involved with a
  localization system?

Stitching and Refinement
This step applies to distributed construction of
large-scale coordinate systems
This step estimates target coordinates (and often
other parameters simultaneously)
Coordinate System Synthesis
Coordinate System Synthesis

• Parameters might include
• Range between nodes
• Angle between nodes
• Psuedorange to target (TDOA)
• Bearing to target (TDOA)
• Absolute orientation of node
• Absolute location of node (GPS)

18
Example of a Localization System

• SHM system, developed at Sensoria Corp.

Each node has 4 speaker/ microphone pairs,
arranged along the circumference of the
enclosure. The node also has a radio system and
an absolute orientation sensor that senses
magnetic north.
Microphone
Speaker
12 cm
19
System Architecture

• Ranging between nodes based on detection of coded
  acoustic signals, with radio synchronization to
  measure time of flight
• Angle of arrival is determined through TDOA and
  is used to estimate bearing, referenced from the
  absolute orientation sensor
• An onboard temperature sensor is used to
  compensate for the effect of environmental
  conditions on the speed of sound

20
System Architecture

• Nodes periodically emit acoustic pulses. Other
  nodes detect these pulses and compute a range and
  angle of arrival.
• Range data, angle data, and absolute orientation
  are broadcast N hops away.
• Based on this table of ranges, angles, and
  orientations, each node applies a multilateration
  algorithm with iterative outlier rejection to
  compute a consistent coordinate system.

Range, Angular Data
Multilat Engine
Range, Angular Data
Multilat Engine
Range, Angular Data
Multilat Engine
21
Localization System Components

• Sensing layer Ranging, AOA, etc.

This step applies to distributed construction of
large-scale coordinate systems
Stitching and Refinement
This step estimates target coordinates (and often
other parameters simultaneously)
Coordinate System Synthesis
Coordinate System Synthesis

• Parameters might include
• Range between nodes
• Angle between nodes
• Psuedorange to target (TDOA)
• Bearing to target (TDOA)
• Absolute orientation of node
• Absolute location of node (GPS)

22
Active and Cooperative Ranging

• Measurement of distance between two points
• Acoustic
• Point-to-point time-of-flight, using RF
  synchronization
• Narrowband (typ. ultrasound) vs. Wideband (typ.
  audible)
• RF
• RSSI from multiple beacons
• Transponder tags (rebroadcast on second
  frequency), measure round-trip time-of-flight.
• UWB ranging (averages many round trips)
• Psuedoranges from phase offsets (GPS)
• TDOA to find bearing, triangulation from multiple
  stations
• Visible light
• Stereo vision algorithms
• Need not be cooperative, but cooperation
  simplifies the problem

23
Passive and Non-cooperative Ranging

• Generally less accurate than active/cooperative
• Acoustic
• Reflective time-of-flight (SONAR)
• Coherent beamforming (Yao et al.)
• RF
• Reflective time-of-flight (RADAR systems)
• Database techniques
• RADAR (Bahl et al.) looks up RSSI values in
  database
• RadioCamera is a technique used in cellular
  infrastructure measures multipath signature
  observed at a base station
• Visible light
• Laser ranging systems
• Commonly used in robotics very accurate
• Disadvantages directionality, expense, no
  positive ID of target

24
Using RF for Ranging

• RF TOF techniques
• Accurate, deterministic transponders hard to
  build
• Temperature-dependence problems in timing of path
  from receiver to transmitter
• But, you can use RBS techniques (compare
  receptions)
• Measuring TOF requires fast, synchronized clocks
  to achieve high precision (c ? 1 ft/ns)
• Fast synchronized clocks generally at odds with
  low power
• Trade-off synchronized infrastructure vs. nodes
  (e.g. GPS)
• Ultra wide-band ranging for sensor nets?
• Current research focus in RF community
• Based on very short wideband pulses, measure RTT
  to fixed, surveyed base stations
• FCC licensing?

25
Practical Difficulties with RSSI

• RSSI is extremely problematic for fine-grained,
  ad-hoc applications
• Path loss characteristics depend on environment
  (1/rn)
• Shadowing depends on environment
• Short-scale fading due to multipath adds random
  high frequency component with huge amplitude
  (30-60dB) very bad indoors
• Mobile nodes might average out fading.. But
  static nodes can be stuck in a deep fade forever
• Potential applications
• Approximate localization of mobile nodes,
  proximity determination
• Database techniques (RADAR)

Path loss Shadowing Fading
Distance
Ref. Rappaport, T, Wireless Communications
Principle and Practice, Prentice Hall, 1996.
26
Using Acoustics for Ranging

• Key observation Sound travels slowly!
• Tight synchronization easily achieved using RF
  signaling
• Slow clocks are sufficient (v 1 ft/ms)
• With LOS, high accuracy can be achieved cheaply
• Coherent beamforming can be achieved with low
  sample rates
• Disadvantages
• Acoustic emitters are power-hungry (must move
  air)
• Solid obstructions block sound completely ?
  detector picks up reflections
• Audible sound has good channel properties but
  isnt always appropriate

27
Typical Time-of-Flight AR System

• Radio channel is used to synchronize the sender
  and receiver (or use a service like RBS!)
• Coded acoustic signal is emitted at the sender
  and detected at the emitter. TOF determined by
  comparing arrival of RF and acoustic signals

28
Narrowband vs. Wideband

• Narrowband technique pulse train at f0
• Works with tuned resonant ultrasound transducers
• COTS parts implement detection (SONAR modules)
• Crosstalk between nodes is a problem, introduces
  significant coordination overhead to system
  design
• Used in ORL Active Bat, MIT Cricket, UCLA AHLoS
• Wideband technique pseudonoise burst
• Detection requires 100M FLOPs, 128K RAM
• High accuracy, excellent interference rejection
• 100m range achieved over grass in outdoor
  environment
• Excellent crosstalk rejection each xmitter uses
  diff. code
• Used in GALORE Panel, IPAQs, SHM

29
Wideband Acoustic Detection
Autocorrelation
Value of Correlation Function
First arrival and echoes
Offset in samples (20.8?s, or 0.71 cm)
30
Ringing
Ringing Introduced by Speaker
Probable Echo
Can be corrected at the source (deconvolve source
waveform) or at the receiver (correlation)
31
An Acoustic Ranging Error Model

• A useful model for error in acoustic ranges is
• Rij Xi Xj2 nij
  Nij,
• where
• nij is a gaussian error term (?0,?1.3)
• Nij is a fixed bias present only when LOS blocked
• Error reduction
• nij can be reduced by repeated observations
• Nij cannot because it is caused by persistent
  features of the environment, such as detection of
  a reflection.
• The Nij errors must be filtered at higher layers
• Cross-validation of multiple sensor modalities
• Geometric consistency, error terms during
  multilateration

32
Typical Angle-of-Arrival AR System

• TOF AR system with multiple receiver channels
• Time difference of arrivals at receiver used to
  estimate angle of arrival

Radio
Radio
CPU
CPU
Speaker
33
Sources of error in bearing position est.

• Quantization error in detection
• Sample rate of detector lower bounds phase
  accuracy
• Synchronization error
• Detection error
• Noise, interference, ringing, blurring in channel
• Excess path length due to clutter
• Non-LOS detection of reflected paths
• Usually easy to filter if sensor positions are
  known
• Angular dependence in sensor
• Error in measurement or estimation of baseline
• Other... various relatively unlikely
  occurrences
• Collisions between codes, system failures, etc
• Its hard (impossible?) to root out the last 0.1
  of the problems!

34
Localization System Components

• Coordinate system synthesis refinement layer

Stitching and Refinement
Stitching and Refinement
This step applies to distributed construction of
large-scale coordinate systems
This step estimates target coordinates (and often
other parameters simultaneously)
This step estimates target coordinates (and often
other parameters simultaneously)
Coordinate System Synthesis
Coordinate System Synthesis
Coordinate System Synthesis
Coordinate System Synthesis

• Parameters might include
• Range between nodes
• Angle between nodes
• Psuedorange to target (TDOA)
• Bearing to target (TDOA)
• Absolute orientation of node
• Absolute location of node (GPS)

35
Position Est. and Coord. Systems

• Position Estimation, Triangulation
• Some of the nodes have known positions
• Targets position inferred relative to known
  nodes
• e.g. Active Bat, single GALORE Panel
• Forming a coordinate system, Multilateration
• Most nodes have unknown positions
• Consistent coordinate system constructed based on
  measured relationships between nodes
• Multilateration is a commonly used term
• e.g. N-hop ML, SHM, GALORE IPAQ system

36
Optimization Problems

• Often implemented as an overconstrained
  optimization problem
• Input is set of measurements
• Ranges, angles, other relationships
• Output is estimated node position map
• Environmental parameters often estimated
  concurrently
• Gaussian error ? least-squares minimization
• Careful filtering required to ensure this
  property
• Centralized vs. distributed

37
Simple Example GALORE Panel

• Pythagorean Theorem yields constraints
• Object is to find position estimate that
  minimizes squared sum of error terms

GALORE panel system 4 mics at measured locations
Mote emitter at unknown location
Where is measurement error in the range
measurement
38
GALORE Panel Position Estimator

• Rewrite to get error as function of position and
  meas. range
• Problem error function is not linear must use
  NLLS
• Approximate the constraint function by a Taylor
  series
• Neglecting higher-order terms, and choosing an
  initial guess X0,
• Linear approximation of the error function in
  that neighborhood
• Solve for ?X that minimizes error
• Iteratively improve X0 until it converges
• Good initial guess ? likely convergence if well
  constrained

Ref. Strang, and G, Borre, K, Linear Algebra,
Geodesy, and GPS, Wellesley-Cambridge Press, 1997
39
Importance of Outlier Rejection

• Outliers in source data cause problems
• Acoustic localization data
• 480 ranges, 30 nodes
• Outliers caused by code collisions and multipath
• 4 outliers rejected

40

• Even if most of the data is good, a few outliers
  can cause significant position error for this
  algorithm
• Finding the outliers is easier with more data
• Studentized residual analysis weights error
  terms by their impact on solution

41
Scaling NLLS Position Estimation

• NLLS algorithm with outlier rejection scaling
  issues
• Centralized all ranges are brought together
• O(N4), thus does not scale well for large nets
• Solution partial distribution
• Collect ranges from a few hops away
• Accrete small set of well-constrained nodes, and
  process them into coordinate system
• Pass along that system for neighbors to extend
• Current data shows 1 degradation using set size
  of 14 nodes in a 30 node network with high
  connectivity

42
N-Hop Multilateration1

• Fully distributed algorithm
• Very low overhead avoids mutliplies
• Designed to be robust to long ranges
• Initial implementation
• Running on MK-2 platform (ARM-thumb Atmel 103
  radio controller, RFM transceiver)
• Initial application Smart Kindergarten project
• Measured and deployed beacons on ceiling
• Small receiver tags on toys and children

1A. Savvides, H. Park, M. Srivastava, The bits
and flops of the n-hop multilateration primitive
for node localization problems, First ACM
Workshop on Wireless Sensor Networks and
Application (WSNA), pp112-121, Atlanta GA,
September 2002.
43
N-hop System Layers

• Ultrasound ranging
• Uses COTS ultrasound detection chipset, RF sync
• N-hop lateration
• Nodes compute range to beacons
• Pass their range to beacons to RF neighbors
• If you have no direct range to a beacon
• Compute your range to neighbors who see that
  beacon
• Add your range to that neighbor to their range to
  beacon
• Keep minimum computed range to each beacon

44
Key Points
? Although the summed ranges will be larger than
the true range, with sufficient and uniform node
density the excess path range is bounded and
fairly small
? And, long ranges will automatically be dropped!
? Caveat Important to filter out short ranges
45
Low-cost Multilateration Step

• Once each node has ranges to N beacons
• Min-Max algorithm
• Approximates beacon ranges to squares
• Computes centroid of bounding box

? Approximation is often quite close (blue vs red)
? Redundant long estimates are implicitly
dropped
? Caveat Approximation as well as error
tolerance is best if the node is inside a convex
hull of beacons
46
N-hop multilateration refinement

• Iterative refinement step can improve upon the
  min-max algorithm
• Locally collects inter-node ranges and
  iteratively improves the estimates
• Must include scheme to reject outlier data
• Overall assessment of N-Hop Multilateration
• Simple and fully distributed
• Implicitly eliminates many forms of outlier data
• Position accuracy at least order of node density

47
Scaling Positioning Across the Network

• Idea from RBS transform to local time at every
  hop
• Improves scalability by avoiding need for global
  time
• Similar technique may be useful for localization
• Transform to local coordinate system at each hop
• However,
• Error propagation characteristics are more
  difficult to estimate and model than timing
  errors highly dependent on geometry and
  environment.
• Likely not zero-mean Gaussian but drift can be
  mitigated by including beacons or survey points
  with absolute positions.

48
System Design Case Study IPAQ-based Localization

• Followon to GALORE Panel System
• Intended to be an ad-hoc deployable system of
  small units (IPAQs) that can self-organize and
  then localize motes.
• Computational cost of sender is low IPAQs do
  detection

IPAQs are unmodified, running Familiar Linux
uses internal speaker and microphones. Acoustic
Mote adds spkr, amp on daughterboard (N. Busek)
49
Acoustic Mote
50
Experimental Setup
51
System Architecture

• Multi-hop Timesync (J. Elson)
• Audio Server Daemon (L. Girod)
• Acoustic Range Estimation (L. Girod)
• Multilateration (V. Bychkovskiy)
• Self-configuration phase (IPAQ-IPAQ positioning)
• Operation phase (Mote-IPAQ positioning)

52
Multihop Timesync

• Multihop Timesync System (J. Elson)
• Provides conversion parameters among the many
  components of the system
• IPAQ system clocks
• IPAQ codec clocks
• MoteNIC system clocks
• Acoustic Mote system clocks
• Timesync service transparent to applications
• Conversions can span many RF hops and independent
  networks

53
Timesync Connectivity Graph
54
Multihop Timesync in Action
Acoustic Mote
CODEC 1
CODEC 2
Convert to Mote1 (RMS 15uS)
Compare Acoustic Arrival Times, Mote Send Time
Convert to CPU2 (RMS 0uS)
Convert to CPU1 (RMS 0uS)
Mote 1
Mote Broadcaster
CPU 1
CPU 2
Convert to CPU1 (RMS 1uS)
Convert to CPU1 (RMS 2uS)
IPAQ 1
802.11
802.11
IPAQ 2
IPAQ 802.11Broadcaster
55
Audio Server

• The audio server performs two functions
• Buffers last N seconds of audio data
• Logs DSP transfer interrupts and supplies data to
  Timesync system to enable time conversions.
• Vastly simplifies the system
• Enables multiple user applications to process
  data
• Decouples detection from message timing concerns
  pull data out of the past
• Timesync eliminates triggered recording

56
Multilateration

• Based on mass-spring model
• Larger region of convergence, reduces
  requirements on initial guess
• Customized force function makes shorter ranges
  stronger so they reel in long ranges
• Outliers identified by excess tension after
  convergence
• Weird IPAQ geometry
• Mic and Speaker in different locations
• Orientation is also a variable
• Initial pass averages considers ranges from
  center
• Second pass allows them to re-orient to relax
  network

57

• Thank you!

58
Next Steps Applications

• Tracking in a mote field
• Acoustic threshold detection in mote field
  triggers responses (N. Busek)
• Using RBS and the GALORE Localization system,
  motes will be able to correlate their
  observations in time and space
• Coherent Signal Processing
• One or more panels can collaborate to do passive
  localization and beamforming (H. Wang)
• RBS provides accurate synchronization
• GALORE Localization system determines precise
  relative positions of the receivers.

59
Multi-channel Correlation
Value of Correlation Function
Arrival times at the four channels
Offset in samples (20.8?s, or 0.71 cm)
60
Parallel Rays Approximation

• Parallel rays approximation
• True for distant point source
• ? acos(?/B)
• At shallow angles, small variations in relative
  range yield large angular errors
• When is it valid?
• Red is error in degrees between parallel rays and
  true value
• Blue region shows cases where error is less than
  0.5 degrees

61
Intuition from parallel rays

• Error in bearing estimate from quantization error
• Quantization error is error caused by sample
  timing
• Derivative of acos(x) represents sensitivity to
  small errors in relative ranges

X (???)/B X?1 as ??0