SLIDE 1
Introduction to Collaborative Signal Processing Feng Zhao Xerox - - PowerPoint PPT Presentation
Introduction to Collaborative Signal Processing Feng Zhao Xerox - - PowerPoint PPT Presentation
Introduction to Collaborative Signal Processing Feng Zhao Xerox Palo Alto Research Center January 15-16, 2001 29 Palms, August 2000 Model-Directed Sensing and Estimation Insight: Model provides spatio- temporal priors to focus sensing and
SLIDE 2
SLIDE 3
Signals and sensing Models Sensor fusion
Model-Directed Sensing and Estimation
Insight: Model provides spatio- temporal priors to focus sensing and signal processing! Aggregation Features Parameter estimation Event components Source Separation
Event Event Event
SLIDE 4
Tracking job flow in a printshop
- The problem: Identify workflow of printshops consisting of multiple machines
– Determine job initiation, duration, and component conditions – Optimize the shop operations using the information
- Requirement: Minimal disruption to existing printshop operation and intrusion to
customer networks (security)
- Approach: Use distributed microphones and RF tags to locate/track job flows
Media Storage Printing Binding Plate Printing
Shipping and Receiving
Collating
Shipment Preparation (boxing)
Microphones
Printshop
Wireless link PC Internet connection RF tag reader Printing
SLIDE 5
Idea: Event model as a temporal prior in distributed monitoring
- As in speech processing, identify
grammar and inter-word constraints
- Use a Petri net event model to
constrain the search for next “word”
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
- 0.06
- 0.04
- 0.02
0.02 0.04 0.06
time [s] acoustic signature stack elevation AS pull-in FM ramp-up Unknown AS off FM off Unknown
[ ]
) ( ) | ( ) | ( exp ) , | ( ) , , ( ) (
2 1 ) ( 2 1 1 1 ) ( ) 1 ( 1 ) 1 ( 1 ) 11 ( 11 1 2 1 t t t t t t T t t t m s t s nm t nm n t n m t m t t n z t z t t t
e P e P k e P k P e e M e P ⋅ ⋅ = ⋅ = =
− Σ − ∗ − − − − = − −
z z z
r r
: Update (3) : function Likelihood : separation Source (2) : Prediction (1) τ α
τ δ α τ δ α τ δ α τ δ α M L M L M
L
- 5
- 4
- 3
- 2
- 1
1 2 3 4 5 0.04 0.05 0.06 0.07 0.08 0.09 0.1 0.11
time deviation [ms] posterior predicted time Estimated time
Multi-channel data streams
SLIDE 6
Dynamics as a prior in multiple target tracking
Particle Filter (Condensation) as an implementation for Bayesian Filtering
- Multi-modal, non-Gaussian
distribution
- Uniform representation
Probability distribution at k-1 Prediction Probability distribution at k Observation
) | (
k k Z
S P
) | (
1 1 − − k k
Z S P
) | (
1 − k k S
S P ) | (
k k S
Z P
Two vehicle targets cross over
SLIDE 7
A few issues in distributed implementation
- Distribute hypotheses and models
- Transmit results v.s. raw data
- Exchange and combine data with great
variability in latency and quality
- Dense small nodes vs. sparse large nodes
- Track pixels v.s. predicates
- Sensor wakeup and cueing
- Taxonomy of application scenarios
SLIDE 8