Development and Characterization of an Acoustic Rangefinder

1
Development and Characterization of an Acoustic
Rangefinder

• Lewis Girod
• 8 December 1999

2
Objective

• Tags can benefit from having accurate range data
  to other tags
• Many context-aware apps need spatial awareness
• Small scale of elements requires accurate info
• Indoor embedded systems can't use GPS
• Scaling issues.
• Pinpoint centrally manages location information
• Active Badge/Bat centrally managed, but base
  stations only handle contents of single room
• But both require significant investment in
  infrastructure

3
Requirements

• Conducive to decentralized algorithms
• High accuracy relative to tag size (cm)
• Good interference rejection
• Must work outdoors and in any environment
• Resistant to jamming
• Multiple access
• Tags should not interfere with each other

4
Active Bats

• Similar in design and requirements
• Infrastructure far from ad-hoc
• Centralized data collection
• Positions of receivers must be carefully measured
• System must coordinate readings globally
• Only one bat triggered at a time to avoid
  interference
• Limits the measurement rate as system scales
• Highly accurate (cm) but works only at low range
  
• Receivers installed every 5 ft
• Susceptible to jamming, and long paths when
  obstructed

5
Theory

• Measure time of flight of sound
• Sender emits a particular sound at T0
• Receiver detects that sound at T1
• Distance ??T Speed of Sound

6
Synchronization

• What if the clocks aren't synchronized?
• Speed of sound is 331.5 m/s
• 1 msec of error is 33 cm!
• Solution Out-of-band synchronization
• Use a radio signal.. It's 6 orders of magnitude
  faster than sound
• Synchronization is still not trivial
• each cm is only 30 microseconds
• O/S gets in the way

7
System Design Overview
Sender
Receiver
Chirp Request Message
Receive Chirp Request
T1
Record Timestamp
Begin Listening
Begin Radio Sync
T2
Radio packet transmission
Begin Chirp
Chirp
Listen
T3
Time
Critical synchronization Radio transmission Acoust
ic transmission
Detect T3 Result is (T3 - T2) VS
8
Experimental Testbed

• Testbed needed to explore the space
• Chose to use PCs and soundcards
• Drawback lower frequency
• sampling at 44.1KHz, max resolution is 7.5mm
• Benefits
• DSP freedom to select and detect almost
  arbitrary waveforms
• Standard audio hardware has fairly wide band

9
Design Issues

• Modulation strategy
• What sort of waveform is best?
• Which waveforms are most resilient to
  interference, jamming, crosstalk?
• Synchronization strategy
• How can we eliminate sources of sync error?
• Detection strategy
• What detection algorithms are most effective?
• Balance multipath rejection vs. noise rejection

10
Modulation

• Design principles
• Measure the channel characteristics
• What does the system do to the signal?
• DAC ? Amp ? Speaker ? Noisy, multipath
  environment ? Microphone ? ADC
• The only controllable aspect is the environment
  (e.g. quiet foam box vs. noisy machine room)

ADC
Mic/ Preamp
DAC
Amp/ Speaker
Environment
22KHz
20Hz-16KHz
22KHz
20Hz-20KHz
11
Channel Measurement

• Methodology
• Use DFT to calculate power spectrum
• Graphs magnitude of frequency components
• Perform experiments in anechoic chamber
• Measure spectrum of input generated at sender
• Measure spectrum of output observed at receiver
• Then compare the graphs
• Estimate the system pass band, then select a
  waveform that passes through adequately

12
White Noise Response
Drop-outs
Log plots of DFT of input waveform and measured
waveform
13
Spread Spectrum

• Problem Drop outs and interference
• Any single frequency is a bad choice in some
  environment
• Solution signals with wide spectrums
• Even if some frequency components are lost, the
  signal can still be detected
• Multiple access for free if correlation
  between signals is low, they will look like noise
  to each other

14
Spreading Techniques

• Frequency sweep
• Analog hardware can do this
• Good interference rejection, but no coding
• Direct sequence
• Long, high rate bit sequence
• Frequency hopping
• Shorter sequence of pulses selected from a large
  set of frequencies

15
Power Spectra (a) 4 KHz frequency sweep (b) 10
Kb/s sequence (128b) (c) 768 freq, 64 hop sequence
16
(No Transcript)
17
Cross Correlation

• Cross correlation occurs when the signal is
  similar to offset copies of itself
• Ex. A train of identical pulses
• Cross-correlation causes poor detection
• Leads to detection of ghost images
• Coding and Frequency Sweep methods exhibit low
  cross-correlation

18
Design Issues

• Modulation strategy
• Synchronization strategy
• How can we eliminate sources of sync error?
• Detection strategy
• What detection algorithms are most effective?
• Balance multipath rejection vs. noise rejection

19
Synchronization
Sender
Receiver
Chirp begins immediately after radio controller
accepts packet. Hopefully delay inside
controller is fixed
Sync timestamp with start of listen
T1
Kernel module catches radio interrupt and saves a
timestamp radio interrupt assigned maximum
priority
T2
Radio packet transmission
Chirp
Listen
T3
To achieve 1 cm resolution, maximum timing
variance is 30 ?s
20
Design Issues

• Modulation strategy
• Synchronization strategy
• Detection strategy
• What detection algorithms are most effective?
• Balance multipath rejection vs. noise rejection

21
Detection

• A matched filter, or autocorrelation function, is
  used to detect the signal
• Receiver constructs an identical copy
• Compares it to the observed signal at each
  possible offset
• The best matchis considered the true offset,
  and the time of flight is calculated

22
Matched Filter
Pairwise multiply and sum (Dot product)
23
Making it work

• Two components
• Selecting a signal that is easy to detect
• Lower cross-correlation ? sharper detection
• Random coding ? low correlation with
  environmental noise ? avoid false positives
• Detection heuristic
• In multipath and obstructed environments, the
  best match numerically is not always the right
  one must consider good matches that come earlier

24
Good Correlation
(Note log scale) Y axis is a measure of the
degree of correlation at each offset. There is a
prominent, narrow peak at sample 250. Sample
taken in a quiet, echo-free environment.
Peaks
Cross-correlation
25
Obstructed/Multipath
(Note linear scale) A peak detection function
selects the first good candidate. The noise
threshold must adapt to environmental noise and
to variations in signal power.
Cluster of Echoes
Looks like the peak
Shortest path was attenuated by an obstruction
26
Estimating Noise

• The magnitude of noise depends on environment
• Want to set threshold just above
  cross-correlation

Noise Threshold

• Solution Run a test correlation
• Generate a separate random sequence, use same
  modulation algorithms
• Correlation will exhibit cross-correlation
  without peaks

27
Detection Algorithm

• Estimate noise threshold
• Locate first above threshold correlate
• Based on knowledge of modulation, select the
  center of peak

• Shape of peak depends on modulation some are
  wider than others
• In some cases, curve fitting may help

Peak from frequency sweep correlation
28
Results
This graph shows the results of a series of tests
in a noisy machine room. Each point represents
about 10 trials. The tests were conducted at 1 m
intervals. The data in each point ranges about
?1.5 cm. The variance is about 0.01 cm
29
Results (detail)
30
Sources of inaccuracy

• For individual readings, synchronization is the
  biggest problem
• Averaging fixes this
• In good environments, low frequency and sample
  rate second largest problem
• In bad environments, environmental
  characteristics are the largest problem
• Little can be done about long paths when LOS is
  completely obstructed

31
Future Work

• Future testing should include
• Longer ranges, outdoor environments
• Environments with many beacons
• Environments with deliberate noise sources
• Use two detectors to extract orientation
• Improvements to the detector
• Use of specialized hardware to do better
  synchronization