Highrate sensing wants smart, interactive sensing

1
High-rate sensing wantssmart, interactive sensing

• Lewis Girod
• MIT/CSAIL

This material discusses work performed at the
Center for Embedded Networked Sensing (NSF Grant
No. CCF-0120778) and the Networks and Mobile
Systems group at CSAIL.
2
High Rate Continuous Sensing Systems and
Applications
Systems and applications developed at UCLA/CENS
Acoustic ENSBox (PXA255) 4 x 48KHz x 16bit
(385KB/s) Study of Dusky Antbird, Marmot
Cyclops Imager (Mica2) 16KB images x 8fps
(128KB/s) Hummingbird nest boxes
Vango/Neuromote (Telos) 115KHz x 12bit
(172KB/s) Neural signals from living mice other
acoustic sensing applications
(Raw data rate) (deployment size) gtgt Available
BW Storage is not infinite ? Huge value in moving
processing toward the sensor
3
One common solution Filtering at the sensor

• If our interest is Antbird calls, develop a
  filter that records and analyzes only that data
• Real time prefilter triggers slower analysis,
  storage, network
• Hints from other nodes

4
Developing and debuggingthis is not so easy

• Filtering algorithms
• Pre-deployment testing occurs under different
  conditions (noise, etc.)
• Source signals may not be understood
• Discovery of new phenomena!
• Need to debug in the field.. But how?
• Throwing away raw data
• At some point this is necessary
• How can this be done with confidence?

5
Two proposals

• Develop filter interactively in the field
• Start with raw data sequences
• Iteratively refine and test filters
• Researcher-in-the-loop applies domain knowledge
  and multimodal sensing
• Data collection with redundancy
• Periodically record raw data to validate filters
• Save lineage of filtering decisions so they can
  be checked, verified, debugged
• Upcoming deployments (July, Sept)
• Experiment with this general methodology
• Leverage this experience to develop a more
  convenient solution

6
Interactive filter development with Wavescope
MIT Networks and Mobile Systems Group

• Prior work TinyDB and similar CQ databases
• Express query, receive stream of responses
• Provides interactive interface
• But.. limited expressiveness, range of operators
• Wavescope query language
• Optimized for high-rate isochronous data
• Query defines dataflow with DSP operators
• UI Abstraction/composition query optimizer
• Goal natural, interactive, efficient operation
• Queries can be expressed easily and installed
  rapidly
• Concurrent queries, e.g. periodic raw data
• Refine system, feature set based on several use
  cases

7

• Thanks!

8
But we want all the data

• OK, but with computation at the sensor
• Longer deployments
• More sensor locations
• More useful data (with proper validation)
• Usually there are implicit limitations that
  youre used to video frame rate
• What the scientists are used to matters
• Seismic/geophysics vs. animal behavior