Strategies for Sampling the Environment

1
Strategies for Sampling the Environment
Center for Embedded Networked Sensing
Xiangming Kong, Mohammed Rahimi, Richard Pon,
Bill Kaiser, Mark Hansen, Greg Pottie, Gaurav
Sukhatme, and Deborah Estrin
Explore different sampling strategies for
achieving high fidelity with limited resources
Motivation
Major Strategies

• Multiscale sampling
• Applied when different types of resources are
  available
• Examples
• Sun light intensity. Resources PAR sensor,
  camera
• Environmental temperature. Resources thermal
  sensor, infra camera
• Adaptive sampling
• Applied when only one type of resource is
  available
• Examples
• CO2 Respiration
• Water polution

• Fundamental challenges
• Phenomena are fast changing both spatially and
  temporally
• High fidelity reconstruction is required
• Extensive resource allocation is expensive
• Solution
• Probe ability of different resources
• Examine features in the phenomena
• Exploit mobility of sensing nodes

Strategy design

• Space-Filling Design
• Fill up the space
• Sample points with maximum distance
• Bounded convergence on expected value
• Modulate vs. cost of collecting sample
• Cost is the lost time to collect the sample
• Sampling cost (time)
• Navigation cost

• Adaptive Design
• Track interesting features
• Adaptivety can be misleading
• Balance between the desire to follow features and
  the desire to fill up space evenly
• Gradually relax toward adaptivity

• Multiscale sampling
• A combination of sparse multiscale/multimode
  sensing can yield accuracy of exhaustive sampling
• Pros and cons of each mode
• Camera has high speed and low accuracy
• PAR sensor has low speed and high accuracy
• Establish object model
• One level of sampling can direct another level of
  sampling based on the model

Experimental Results
 
Space Filling Desing

• Field Partition
• Whole field is non-homogeneous
• Can be partitioned into smaller homogeneous
  areas bright area, dark area, penumbra

Adaptive Design

• Track Features
• In the case of light the error was dominated by
  Bias or curvatures
• Find the amount of error in the model, compensate
  model error by more observation in high error
  areas.

• We digitize the world
• Limit the computation cost
• Occupancy map overlay for non standard geometric
  shapes
• Algorithm
• ? Sample Space
• S sampled point
• Static, mobile sensor
• x is robot position and d(x,y) is distance to a
  candidate sample point in ? then we can pick next
  sample point such that this criteria is maximum

• Bright and dark areas are approximately uniform
• Penumbra effect
• Due to non-point light source
• Light intensity change continuously
• Intensity curve can be approximated by linear
  function

• Regeneration of the scene
• More points alongside edges
• Silhouette is the location of the maximum error

• Voronoi Tessellation
• Maximum distance
• Continuously refinable
• Nice empirical convergence of the estimation error

• Calibration
• Outliers in the image due to different
  reflectivity
• Transformation between PAR sensor coordinate
  system and image coordinate system

• Effect of sensor delay
• From continuous design to a disconnected design

• Collaboration between different nodes
• Camera image provide global information
• In-situ static PAR sensor detect local intensity
• Mobile PAR sensor complement static sensor
• Camera directs mobile sensor to obtain new
  measurement

UCLA UCR Caltech USC CSU JPL UC
Merced