Stanford CS223B Computer Vision, Winter 2006 Lecture 12 Filters / Motion Tracking 2

1
Stanford CS223B Computer Vision, Winter
2006Lecture 12 Filters / Motion Tracking 2

• Professor Sebastian Thrun
• CAs Dan Maynes-Aminzade, Mitul Saha, Greg
  Corrado

2
Team Names

1. noname
2. Skynet
3. AED
4. WeeDrivers
5. Fear Factor
6. Dangerous Driver
7. First Star on the Left
8. Team Lee
9. The Five-Seconds Rule
10. Colorado
11. Pien
12. Triangulated
13. Making Drivers Obsolete
14. LyricalAssasins
15. Gal
16. Brad Evan
17. Three Blind Mice
18. Optical Optima
19. The Visionaries

3

• Online data
• cs223b.stanford.edu/competition

4
USB Frash Drive

• team.txt -replace with you team name, number
• /data
• /data/clip1
• /data/clip2
• /data/clip3
• /results
• Will not be available
• /submit
• clip1.png
• clip2.png
• clip3.png

5
Competition

• 5 seconds of driving (150 frames)
• 3 seconds of blackout

6
Important Information

• Memorize your team number
• Pick up flash drive today (SCPD will post on
  Web)
• Add files in submit directory
• Change team.txt
• Return on Monday

7
Moving Objects
8
Kalman Filter Tracking
9
Particle Filter Tracking
10
Mixture of KF / PF (Unscented PF)
11
Review Kalman Filters
prior
12
Summary Kalman Filter

• Estimates state of a system
• Position
• Velocity
• Many other continuous state variables possible
• KF maintains
• Mean vector for the state
• Covariance matrix of state uncertainty
• Implements
• Time update prediction
• Measurement update
• Standard Kalman filter is linear-Gaussian
• Linear system dynamics, linear sensor model
• Additive Gaussian noise (independent)
• Nonlinear extensions extended KF, unscented KF
  linearize
• More info
• CS226
• Probabilistic Robotics (Thrun/Burgard/Fox, MIT
  Press)

13
Particle Filters

• An alternative technique for tracking
• Easier to implement
• Nonlinear
• Better for data association
• In CV, known as Condensation Algorithm

14
Particle Filter
15
Particle Filters Basic Idea
See e.g., Doucet 98, deFreitas 98
16
Basic Particle Filter Algorithm
Initialization X0 ? n particles x0 i
p(x0) particleFilters(Xt -1 ) for i1 to n
xt i p(xt xt -1i)
(prediction) wti p(zt
xti)
(importance weights) endfor for i1 to
n include xt i in Xt with probability ?
wti (resampling)
17
Particle Filters Illustration With Wolfram
Burgard, Dieter Fox, Frank Dellaert
18
Particle Filter Explained
19
Monte Carlo Localization (1)
20
Monte Carlo Localization (2)
21
Particles Robustness
22
Tracking People from Moving Platform

• ? robot location (particles)
• ? people location (particles)
• ? laser measurements (wall)

With Michael Montemerlo
23
Tracking People from Moving Platform

• ? robot location (particles)
• ? people location (particles)
• ? laser measurements (wall)

With Michael Montemerlo
24
Examples
Siu Chi Chan McGill University
25
Examples
D. Stavens. D. Lieb. A. Lookingbill
Particle filter
Optical flow
26
Another Example
Mike Isard and Andrew Blake
27
Tracking Fast moving Objects
K. Toyama, A.Blake
28
More Particle Filter Tracking
David Stavens, Andrew Lookingbill, David Lieb,
CS223b Winter 2004
29
Nonlinearity in the Particle Filter
30
Data Association in Particle Filters

• Suppose k features in image, k state variables.
  Which ones to marry?
• Particle filter Each particle selects its own
  data association
• Probabilistic interpretation Particles are
  posteriors over continuous state and discrete
  data associations. (KF can only to continuous
  state)

31
Summary Kalman Filter
Particle

• Estimates state of a system
• Position
• Velocity
• Many other continuous state variables possible
• KF maintains
• Mean vector for the state
• Covariance matrix of state uncertainty
• Implements
• Time update prediction
• Measurement update
• Standard Kalman filter is linear-Gaussian
• Linear system dynamics, linear sensor model
• Additive Gaussian noise (independent)
• Nonlinear extensions extended KF, unscented KF
  linearize
• More info
• CS226
• Probabilistic Robotics (Thrun/Burgard/Fox, MIT
  Press)

and discrete
set of particles (example states)
predictive sampling
resampling, importance weights
fully nonlinear
easy to implement