Robust Range Estimation Using Acoustic and Multimodal Sensing

1
Robust Range Estimation Using Acoustic and
Multimodal Sensing

• Lewis Girod and Deborah Estrin
• 24 April 2001

2
Goal A Fine-grained, Ad-hoc Deployed
Localization System

• Sub-cm scale localization that is
• Independent of environment
• (foliage, clutter, obstructed sky)
• Self-configuring/self-calibrating
• (active calibration)
• Minimal requirements on deployment
• (at most a few rules of thumb)
• Implemented by low-power wireless nodes
• (small, inexpensive, distributed)

3
Why?

• Fine grained localization enables many
  applications for sensor networks
• Routing
• Energy expenditure is a function of path
• Location is a natural namespace for physically
  motivated applications
• Collaborative Sensing
• Location is a natural context for defining
  relations among sensors
• e.g. relative position of two microphones and a
  camera
• UI
• Define task in context of location
• Physical UI, e.g. pointing, moving objects

4
Motivating Example Habitat Monitoring

• Wildlife habitat
• Instrumented with cameras and microphones
• Task is to detect presence of bird and photograph
  it
• One approach
• Use microphones to detect birdcall and estimate
  location
• Then, select a camera that has the bird in field
  of view

5
Habitat Monitoring contd

• Problem Starting with no knowledge, acoustically
  localizing the bird is very difficult
• Simultaneously estimate sensor positions, bird
  position and source signal
• Assuming its possible at all, convergence will
  be slow
• More tractable if we start with knowledge of the
  positions and orientations of the sensors BUT
• We dont want to measure positions
• Want to be resilient to sensors moving
  occasionally

6
Habitat Monitoring contd

• Fine-grained, ad-hoc deployable, cooperative
  localization
• Sensors determine relative positions during
  initialization phase
• Periodic verification accounts for environmental
  changes
• Enables sensor collaboration
• Acoustic sensors (synchronized by radio)
  collaborate to estimate the birds position
• Position identified as being in FOV of a camera,
  which can then capture an image.

7
Our focus

• Thesis
• We are validating this idea through
  implementation
• Building a testbed for multimodal localization
  indoors
• Experimenting with acoustics and cameras

Any individual mode of sensing used for
localization (e.g. acoustic) will suffer from
unrecoverable ambiguities and undetectable
errors. In many cases, these errors are
persistent and are not readily eliminated
statistically. A more robust solution to this
problem is to cross-validate sensor data with
data gathered from alternate perspectives and
from other modes of sensing.
8
Acoustic Ranging

• Our initial experiments have focused on
  developing an active acoustic ranging system
  based on measurement of time of flight of sound.

Radio
Radio
CPU
CPU
Speaker (Kingstate KDS-27008)
Microphone
9
Signaling and Detection

• Wideband ranging signal
• 511 bit M-sequence
• Modulated using BPSK, 12 kHz
• Detected by matched filter
• Earliest peak in output of sliding correlator
• M-sequences autocorrelation properties result in
  good process gain

10
Detection Algorithms

• Earliest peak is most correct Secondary peaks
  represent echoes
• Need to estimate noise level in order to
  correctly identify early peak
• Noise level changes dynamically over the duration
  of the measurement

11
Initial Experiments

• We performed some initial experiments
• Indoors, in our lab
• No special effort to prevent environmental noise
  or multipath interference
• For each experiment, statistical analysis of 20
  trials
• Experimented with
• Measurement of distances from 0-8m
• Temperature / Humidity dependence
• Dependence on orientation or speaker/microphone
• Interference from various obstructions

12
Initial Results

• When the system works, it works well

• Sub-cm ranging accuracy
• (Averaged data after removing outliers)
• No significant error as a function of distance up
  to 10m

13
Fixing minor problems

• In this process, we added numerous algorithms and
  fixes to improve the performance of the estimator
• Reasonable results can be achieved by optimizing
  the performance of the acoustic system
• However, no amount of optimization solves all the
  problems

14
Problems Temp/Humidity

• Variations in speed of sound
• Problem Dependence on temperature and humidity
• Solution model speed of sound based on average
  sensor readings from temp/humidity sensors
• Local variations (lt sensor granularity)
• Problem variations over short distances (e.g.
  sunlight heating surface)
• Solution long-term averaging

15
Problems Orientation

• Orientation dependence
• Problem When speaker or microphone point away
  from line-of-sight, sound diffracts around edges,
  resulting in several cm of error.
• Solution Two speakers back to back and two
  microphones back to back. The microphones are
  input through the stereo line-in port and some
  orientation information can be recovered by
  comparing phase.

16
Problems Obstructions

• Obstructions to the line-of-sight
• Problem Obstructions can reliably prevent
  detection of the line-of-sight path
• Solution Under some circumstances, obstructed
  conditions can be detected
• Multimodal distributions in successive estimates
• Failure of the triangle inequality

Transient detection failure leads to multimodal
distribution
Failure of the triangle inequality
17
Obstructed LOS not always locally detectable

• Obstructed LOS can introduce ambiguities that are
  not always locally resolvable based on acoustic
  range data

• For example, reflections can cause ambiguity.
  The simplest explanation (blue nodes), is not the
  true explanation (green nodes).

• To solve this kind of problem, alternative
  hypotheses must be formulated.
• These hypotheses can have non-local effects (BAD
  for scaling)

18
Cross-validation

• Cross-validation across sensor modes
• Simpler, local algorithms
• Less communication overhead
• Resolves ambiguities faster, more certainty
• Example Cameras and IR LEDs

• Camera sees LEDs (orange dots)
• LEDs emit ID coded pulses
• Acoustic ranges estimated between all components

19
Rich set of cross-checks

• Provides rich set of cross-validation
  opportunities
• Any LED seen by the camera is known to be LOS,
  therefore acoustic range between camera and that
  node is true
• LED images have known range (from acoustics)
  therefore depth map from only one camera
• With a depth map, the distance between two
  visible LEDs can be approximately determined,
  enabling further validation
• Stereopsis is simplified, because the coded LED
  pulses solves the correspondence problem

20
Conclusions and Future Work

• Conclusion Lots of future work
• Complete implementation of dual-microphone
  orientation estimation upgrade testbed
• Multilateration experiments with upgraded testbed
• Camera LED system characterization integration