Title: Detection, Classification and Tracking of Targets in Distributed Sensor Networks
1Detection, Classification and Tracking of Targets
in Distributed Sensor Networks
Dan Li, Kerry Wong, Yu. H. Hu, and Akbar M. Sayeed
Presented by Prabal Dutta prabal_at_eecs
2Outline of the Talk
- Introduction
- Signal Processing Primitives
- Tracking
- Target Classification
- Issues and Challenges
- Future Research
- Conclusions
- Remarks
- Discussion
3Introduction
- This paper
- Outlines a framework for Collaborative Signal
Processing (CSP) in WSN - Proposes detection and tracking algorithms
- Implements and validates classification
algorithms - Argues that CSP can address challenges with
classification and tracking - Suggests CSP algorithms can benefit from
- Distributive processing compute and transmit
summary statistics - Goal-oriented, on-demand processing Only perform
signal processing when a query is present - Information fusion The farther I am, the fewer
details I need to know - Multi-resolution processing Different tasks
require different rates of sampling in space-time
4Signal Processing Primitives
- Detection
- Computes running average of signal power over
some window - Assumes noise is Gaussian
- Calculates a CFAR threshold based on mean and
variance - Event occurs when signal gt CFAR threshold
5Signal Processing Primitives (2)
- Target Localization
- Assumes isotropic, constant exponent signal
attenuation model - Uses energy-based source localization techniques
- Given 4 or more energy readings, uses non-linear
least squares to find best fit (target location
that minimizes error) - Observation Implicitly assumes calibrated and
localized sensors
6Tracking of a Single Target
- Assumes a target enters through one of the
corners - Active cells A, B, C, D
- Uses energy to detect
- Algorithm
- Nodes in cell detect target and report to manager
- Manager estimates current target location
- Manager predicts future position of target
- Manager creates and initializes new cells
- Manager hands off once the target is detected in
a new cell
7Tracking of Multiple Targets
- In the simple case
- Targets occupy distinct space-time cells
- Multiple instances of algorithm can be used in
parallel - In general case
- Multiple tracks may cross (simultaneously occupy
the same space-time cell) - Data association (which track to associate data
with?) - Classification is required to disentangle tracks
- Observation Depending on what the tracks are
used for, and whether it is permissible to
discard old state, classification may not be
required at all.
8Target Classification
- Focuses on classification at a single node
- Uses acoustic and seismic spectra of wheeled and
tracked targets as feature vectors - Extracts feature vectors from time series data
using FFT - Elements of the feature vectors are the Fourier
coefficients (corresponding to the signal power
at that frequency) - Acoustic Down-sampled to fs 5kHz, 1000 point
FFT, only used 0-1kHz BW, then compressed by 4x
and 10x to obtain 50 and 20 element feature
vectors - Seismic fs 256Hz, 256 point FFT using 64
samples and zero padded data segments
9Target Classification (2) Acoustic PSD
- Power Spectral Density plots of different targets
by the same sensor instances - Note the obvious differences in the prototype
signatures, allowing clean separations
10Target Classification (3) Seismic PSD
- Power Spectral Density plots of the same target
by different sensor instances - Note the signature differences in 5a and 5c
- What explains these differences?
11Target Classification (4) Algorithms and
Validation
- Three classification algorithms were tested
- k-Nearest Neighbor
- Maximum Likelihood Classifier
- Support Vector Machine
- Details of the classifiers not discussed here
- To cross-validate the performance of the
classifiers - Available data divided into three sets F1, F2,
F3 - Take two sets at a time for training and one for
testing - Experiment A Training F1F2 training Testing
F3 - Experiment B Training F2F3 training Testing
F1 - Experiment C Training F1F3 training Testing
F2
12Target Classification (5) Acoustic Performance
- SVM demonstrates best performance
- K-NN demonstrates next best performance
- ML demonstrates poorest performance
13Target Classification (6) Seismic Performance
- SVM demonstrates best performance
- K-NN demonstrates next best performance
- ML demonstrates particularly poor performance for
Wheeled Targets (77.6 correct classification
rate)
14Issues and Challenges
- Collaborative Signal Processing faces many
real-world hurdles - Uncertainty in temporal and spatial measurements
- Depends on accuracy of time synchronization
- Depends on accuracy of network node localization
- Variability in experimental conditions
- Classifications assumes that target signatures
are relatively invariant - Node locations and orientations may results in
signature variations - Environmental factors may alter signals
- These nuisance parameters and be included in a
higher dimension feature vectors at cost of
increased processing
15Issues and Challenges (2) - Doppler Effects
- Perceived frequency is a function of radial
velocity from source to sensor - Radial velocity changes as a target passes by
- Observation higher frequencies show greater
absolute changes in frequency
16Future Research
- Key directions
- Move toward more collaborative algorithms
- Extend feature space to higher dimensions
- Intra-sensor collaboration modal fusion
- Combine information from multiple sensors in
single node - Inter-sensor collaboration centralized
processing - Report raw time series data or statistics to a
central node - Doppler-based composite hypothesis testing
- Incorporate target velocity, CPA distance, and
angle between secant and radius (vertex is
targets position)
17Conclusions
- Outlined a framework for Collaborative Signal
Processing in Wireless Sensor Networks - Proposed detection and tracking algorithms
- Implemented and validated classification
algorithms - Discovered that signal or sensor variation can
cause problems with classification and tracking - Suggested that CSP can address some of these
challenges
18Remarks
- No simulations or empirical evidence supporting
single or multiple target tracking - Target models not provided and cell shape and
creation strategy unclear - Target tracking algorithm is purely conceptual
- Target tracking is simply the motivating scenario
for studying classification - Since multi-target tracking with crossing tracks
is the motivating scenario, classifier
performance for superimposed signatures would be
a good idea - Only tracking uses CSP
- Max signal does not always occur at CPA
- Interesting mix of position and results paper
19Discussion