Tracking with Linear Dynamic

Models

Presented By:
Sajid Ali
CSE (Regular)-2010
NITTTR, Chandigarh

1
 Contents:
Tracking
Three main issues
Assumptions
Tracking as Induction
Linear Dynamic Models
Kalman Filter
References

2 Sajid Ali,NITTTR
 Tracking
 Tracking is the problem of generating an inference about
the motion of an object given a sequence of images.

 The key technical difficulty is maintaining an accurate

3 Sajid Ali,NITTTR
 Tracking(Cont..)
 Very general model:
We assume there are moving objects, which have an
underlying state X
There are measurements Y, some of which are functions
of this state
There is a clock
at each tick, the state changes
at each tick, we get a new observation
 Examples
object is ball, state is 3D position+velocity,
measurements are stereo pairs
object is person, state is body configuration,
measurements are frames, clock is in camera (30 fps)
4 Sajid Ali,NITTTR
 Three main Issues in Tracking

5 Sajid Ali,NITTTR
Simplifying Assumptions

6 Sajid Ali,NITTTR
 Tracking as induction
Assume data association is done
we’ll talk about this later; a dangerous assumption
Do correction for the 0’th frame
Assume we have corrected estimate for i’th frame
show we can do prediction for i+1, correction for i+1

7 Sajid Ali,NITTTR
 Base case

8 Sajid Ali,NITTTR
 Induction step
Given

9 Sajid Ali,NITTTR
 Induction step (Cont.)

10 Sajid Ali,NITTTR
 Linear dynamic models
 Use notation ~ to mean
“has the pdf of”, N(a, b) is
a normal distribution with 
y i  N Mi x i ; m i 
mean a and covariance b. xi  N D x ; d i
i1 i1 
 Then a linear dynamic
model has the form
 This is much, much more
general than it looks, and
extremely powerful
 We are on a boat at night
and lost our position
 We know: star position

11 Sajid Ali,NITTTR
 Examples
Drifting points
we assume that the new position of the point is the old
one, plus noise.
For the measurement model, we may not need to observe
the whole state of the object
 e.g. a point moving in 3D, at the 3k’th tick we see x, 3k+1’th tick
we see y, 3k+2’th tick we see z
 in this case, we can still make decent estimates of all three
coordinates at each tick.
This property, which does not apply to every model, is
called Observability

12 Sajid Ali,NITTTR
 Examples
Points moving with constant velocity
Periodic motion
Points moving with constant acceleration

13 Sajid Ali,NITTTR
 Points moving with constant velocity
We have

ui  ui1  tvi1   i
vi  vi1   i

(the Greek letters denote noise terms)

u  1 t u 

14 Sajid Ali,NITTTR
 Points moving with constant
acceleration
We have
ui  ui1  tvi1   i
vi  vi1  tai1   i
ai  ai1  i
(the Greek letters denote noise terms)
Stack (u, v) into a single state vector
u  1 t 0 u 
v  0 1 t v   noise
    
ai 0 0 1 a i1
which is the form we had above

15 Sajid Ali,NITTTR
16 Sajid Ali,NITTTR
17 Sajid Ali,NITTTR
 The Kalman Filter
Key ideas:
Linear models interact uniquely well with Gaussian
noise - make the prior Gaussian, everything else
Gaussian and the calculations are easy
Gaussians are really easy to represent --- once you know
the mean and covariance, you’re done

18 Sajid Ali,NITTTR
 What is it used for ?
Tracking missiles
Tracking heads/hands/drumsticks
Extracting lip motion from video
Fitting Bezier patches to point data
Lots of computer vision applications
Economics
Navigation

19 Sajid Ali,NITTTR
 The Kalman Filter in 1D
Dynamic Model

Notation

Predicted mean

Corrected mean

20 Sajid Ali,NITTTR
 Prediction for 1D Kalman filter
The new state is obtained by
multiplying old state by known constant
adding zero-mean noise
Therefore, predicted mean for new state is
constant times mean for old state
Predicted variance is
sum of constant^2 times old state variance and noise
variance
Because:
old state is normal random variable, multiplying normal rv by constant
implies mean is multiplied by a constant variance by square of constant, adding
zero mean noise adds zero to the mean, adding rv’s adds variance
21 Sajid Ali,NITTTR
22 Sajid Ali,NITTTR
 Correction for 1D Kalman filter
Pattern match to identities given in book
basically, guess the integrals, get:

Notice:
if measurement noise is small,
we rely mainly on the measurement,
if it’s large, mainly on the
prediction

23 Sajid Ali,NITTTR
In higher dimensions,
derivation follows the
same lines, but isn’t as
easy. Expressions
here.

24 Sajid Ali,NITTTR
25 Sajid Ali,NITTTR
26 Sajid Ali,NITTTR
 Smoothing
Idea
We don’t have the best estimate of state - what about the
future?
Run two filters, one moving forward, the other backward
in time.
Now combine state estimates
 The crucial point here is that we can obtain a smoothed estimate
by viewing the backward filter’s prediction as yet another
measurement for the forward filter
 so we’ve already done the equations

27 Sajid Ali,NITTTR
28 Sajid Ali,NITTTR
29 Sajid Ali,NITTTR
30 Sajid Ali,NITTTR
 Data Association
Nearest neighbours
choose the measurement with highest probability given
predicted state
popular, but can lead to catastrophe
Probabilistic Data Association
combine measurements, weighting by probability given
predicted state
gate using predicted state

31 Sajid Ali,NITTTR
32 Sajid Ali,NITTTR
33 Sajid Ali,NITTTR
34 Sajid Ali,NITTTR
35 Sajid Ali,NITTTR
36 Sajid Ali,NITTTR
 References:
http://www.cs.utexas.edu/~grauman/courses/fall2008/s
lides/lecture23_tracking.ppt
http://luthuli.cs.uiuc.edu/~daf/book/bookpages/Slides/
Tracking.ppt

37 Sajid Ali,NITTTR
 Thanks

38 Sajid Ali,NITTTR