Collaborative Signal Processing For Sensor Networks

1
Collaborative Signal Processing For Sensor
Networks

• Sensit CSP Workshop, Jan 15, 2001
• Akbar Sayeed
• Yu Hen Hu
• Kerry Wong and Dan Li
• University of Wisconsin-Madison

2
What Tasks Can We Perform?

• Detect multiple targets
• Count the number of targets
• Classify targets
• Locate targets
• Track the trajectory of targets
• Predict the location/trajectory of targets
• Put confidences on the accuracy of tasks

3
What Properties Should We Sense?

• FEATURES to distinguish targets (bandwidth)
• Energy detection for CPA
• Subspace energy detector outputs (classification)
• Subspace matched to targets (training data)
• Progressive (multi-resolution) accuracy by
  increasing subspace dimension to trade
  performance for latency/computation
• Background noise power in between detected events
  for dynamic threshold adjustment and confidence
  assessment

4
Why Multiresolution Processing?

• Facilitate low-complexity, low-latency processing
• Processing resolution ?? complexity
• Tailor the resolution to the task at hand
• E.g., detection may be accomplished at lower
  resolution followed by classification at higher
  resolution
• Wavelet-based processing
• a natural multiresolution framework
• accomplishing each task with minimum
  resolution/complexity

5
Detect, Track, Classify Scenario

• Detect (CPA)
• Track and predict trajectory from past CPAs
• Progressively accurate classification along the
  trajectory
• Required subspace dimension passed between
  successive sensors

6
Identifying Subspaces (Features)

• Principal component analysis on identified
  features
• Identify features via time-frequency
  representations/wavelets
• Dominant harmonics
• Instantaneous frequency
• Further dimension reduction via representation at
  coarse wavelet scales

7
Short-Time Fourier Transform of SITEX Data
(08030800 DATA, NODE A01)
8
AAV Event 9
9
HMMWV Event 1
10

11
Classification of AAVs versus HMMWVs
Subspace energy detector Outputs (rank 9 512
length vectors)