SonarBased RealWorld Mapping and Navigation

1
Sonar-Based Real-World Mapping and Navigation
Alberto Elfes, 1986

• Lewis Girod
• CS547 Presentation
• 13 October 1999

2
Background

• Previous mapping and navigation efforts based on
  cameras and machine vision
• In the mid-80s several groups began
  investigating the use of cheap ultrasound sonar
  transducers
• Paper has Traditional AI flavor
• But using low-res sonar encourages bottom-up
  design

3
Sensor Characteristics

• Sensors can detect in range 0.9-35 feet
• Main sensitivity in a 30? cone, 1/2 of response
  from middle 15?
• Accuracy ?0.1 feet
• Sensing based on aperture same size object less
  likely to be seen farther away

4
Sensor Array

• 24 transducers, configured in a ring, spaced 15?
  apart
• To avoid interference sensors must be fired
  sequentially (200ms/firing)
• Sensor array mounted on two mobile robots
  (Neptune and Terregator)
• Preprocessor rejects out of range data and
  averages several sequential readings

5
Occupancy Grid

• Interprets sensor readings made at a particular
  location and generates a map
• Three phases
• Collect readings in Probably Empty and Somewhere
  Occupied grids
• Compare grids to generate map
• Match maps to register with readings from other
  locations

6
Probability Profiles

• Each sensor reading is converted into a
  probability profile

Ranging error
Somewhere Occupied
Probably Empty
Range Measurement
Rmin
7
Superposition of Profiles

• Asymmetric composition rules
• P. Empty profiles are combined additively
• Additional empty readings strengthen probability
  that the region is empty

8
Occupied Regions

• Evidence that a region is probably empty is used
  to narrow occupied distribution

• Subtract empty probability values from somewhere
  occupied, and normalize
• Then additively superimpose resulting
  distributions

9
Matching

• Each set of readings is accumulated and combined
  into a single map
• Each new map must be integrated with existing
  maps from past readings
• This is done by finding a rotation and
  translation transform that produces the best
  correlation in overlapping areas

10
Speeding up Matching

• Concentrate on occupied cells
• Occupied cells are rare but important
• Initial matching on downsampled grids
• Identify likely places to search by matching on
  reduced resolution grids
• Found that grids needed to be blurred to filter
  out high frequency noise

11
Post processing

• Overall a traditional AI layered approach
• Readings are taken when the robot is stopped
• Matcher is used to accrete new view into local
  map
• Connected regions are located to produce
  geometric map
• Frame-based symbolic representation of map

12
Multi-axis Representation

• The system makes map data available to
  applications in a variety of representations
• Abstraction, Geographical, Resolution axes
• Raw sensor data, geometric, symbolic
• Single view, local map, global map
• Low, medium, high resolution

13
Path Planning

• Performed at three levels
• Symbolic level decide area to traverse
• Geometric select path through open area
• Sensor level verify path for safety
• Path is converted to line segments
• Maintains a running position estimate
• Reactive override by sonar bumper which can
  cause an emergency stop

14
Results

• Showed the results of two runs, one 30 foot run
  indoors and one 50 foot run outdoors
• Both runs topologically simple
• mostly demonstrated avoidance of an irregularly
  shaped/cluttered wall
• optimistic about effectiveness of mapping
  technique future work is mostly tweaks

15
Analysis

• As we will see, problem is far from solved
• Runs barely extend beyond sensor range
• Correlation technique will introduce cumulative
  error if used for longer runs
• Blurring probably fixed this for small scale
• Simple topology ? simple maps
• Interesting topologies have circuits
• Basic occupation grid idea is useful