Академический Документы
Профессиональный Документы
Культура Документы
Definition of Tracking
Tracking:
Generate some conclusions about the motion of
the scene, objects, or the camera, given a
sequence of images.
Knowing this motion, predict where things are
going to project in the next image, so that we
dont have so much work looking for them.
Why Track?
Detection and
C
recognition are expensive
If we get an idea of
Scene
where an object is in
A B
the image because we
Image Sequence
have an idea of the
motion from previous
C
images, we need less Detection
+ recognition
work detecting or
Tracking
recognizing the object.
(less detection,
no recognition)
New
person
Tracking a Silhouette by
Measuring Edge Positions
Observations are positions of edges along normals to tracked contour
Implicit Assumptions of
Tracking
Physical cameras do not move instantly
from a viewpoint to another
Object do not teleport between places
around the scene
Relative position between camera and scene
changes incrementally
We can model motion
Related Fields
Signal Detection and Estimation
Radar technology
We have measurements
Components of stable feature points in the
images.
Observations, projections of the state.
Necessary Models
We use models to describe a priori knowledge about
the world (including external parameters of camera)
the imaging projection process
Previous State
System Dynamics
Model
Next State
Measurement
Model (projection)
State
(u, v)
Measurement
C=
(x
2
i =1
a)
C
= 0 = ( xi a ) = xi n a
a
i =1
i =1
a
n
1
a = xi
n i =1
n
x
xi-a
xi
ti
Measurement
x
State variable
a
x
xi
ai 1 ai
i 1
xi
1
1
1
ai = xk = xk + xi
i 1
1
i k =1
i k =1
i
a
=
x
1
k
i 1
1
i 1 k =1
ai =
ai 1 + xi
i
i
1
ai = ai 1 + ( xi ai 1 )
i
Actual
measure
Predicted
measure
1
ai = ai 1 + ( xi ai 1 )
i
Estimate at step i
Innovation
Gain specifies how much
do we pay attention
to the difference
between what we expected
and what we actually get
xi = H i a
=
a
H1
measurement equation ... ...
xn H n
x1
x = Ha
that minimizes
Find estimate a
1
T
C = (X H a) (X H a)
2
T
1
T
We find a
= (H H) H X
Hi
x2
xi
H2
H1
Hi
xi
2. Recursive method
x2
H2
1
We had: ai = ai 1 + ( xi ai 1 )
i
i
Now we find: a
with
Predicted measure
= a i -1 + K i (X i H i a i -1 )
Gain matrix
K i = Pi H i
T
T 1
Pi = (H i H i )
Innovation
Xi
Dynamic System
X i = X i 1 + Vi 1 t
Vi = Vi 1 + Ai 1 t
Ai = Ai 1 + w
X i 1 t
V = 0 1
i
Ai 0 0
State of rocket
a i Vi
Ai
Measurement
xi
0 X i 1 0
Tweak factor
t Vi 1 + 0
ai = a i-1 + w
1 Ai 1 w
State equation for rocket
Xi
xi = [1 0 0] Vi + V
Ai
Noise
xi = H a i + V
Measurement equation
Measurement equation
xi = Hi ai + v i
w i ~ N (0, Q i )
Measurement noise
v i ~ N (0, R i )
Prediction for ai
a i = i a i -1 + K i (x i H i ia i -1 ) Prediction for xi
T
T
-1
K i = P'i H i ( H i P'i H i + R i ) Gain
P'i = i Pi -1 + Q i -1
T
i
Pi -1 = (I K i-1 H i -1 )P'i -1
Measurement equation
a i = f ( a i -1 ) + w i -1
xi = Hi ai + v i
a i = f ( a i -1 ) + K i (x i H i f(a i -1 ))
K i = P'i H i ( H i P'i H i + R i )
T
f
f
P'i =
Pi -1
+ Q i -1
a i
a i
T
Jacobian Matrix
Pi -1 = (I K i -1H i -1 )P'i-1
-1
Differences compared to
regular Kalman filter are
circled in red
Tracking Steps
Predict next state as i a i-1 using
previous step and dynamic model
Predict regions N ( H i i a i -1 , P'i )
of next measurements using
measurement model and
uncertainties
Prediction
Make new measurements xi in
region
predicted regions
Measurement (u, v)
Compute best estimate of next state
a i = i a i -1 + K i (x i H i ia i -1 )
Correction of predicted state
Measurement xi
State vector ai
i
Estimation a
Tracking as a Probabilistic
Inference Problem
Find distributions for state vector ai and for
measurement vector xi. Then we are able to
compute the expectations a i and x i
Simplifying assumptions (same as for HMM)
P (a i | a1 , a 2 ,L , ai -1 ) = P (a i | a i -1 )
(Only immediate past matters)
Tracking as Inference
Prediction
P (a i | x 1 , L , x i -1 ) = P(a i | a i-1 ) P (a i-1 | x1 , L , x i-1 )da i -1
Correction
P (a i | x 1 , L , x i ) =
P ( x i | a i ) P (a i | x1 , L , x i -1 )
P(x
| a i ) P (a i | x1 , L , x i -1 ) da i
ai = f ai-1 + wi-1
Measurement equation
xi = h ai + vi
wi ~ N (0 ,q )
Measurement noise
vi ~ N (0,r )
ai = f ai-1 + K i ( xi h f ai-1 )
2
-1
K i = p' i h (h p'i + r ) Gain
p'i = f pi 1 + q
2
Prediction for xi
Batch methods
Recursive methods
(Kalman filter)
References
Kalman, R.E., A New Approach to Linear Prediction
Problems, Transactions of the ASME--Journal of Basic
Engineering, pp. 35-45, March 1960.
Sorenson, H.W., Least Squares Estimation: from Gauss to
Kalman, IEEE Spectrum, vol. 7, pp. 63-68, July 1970.
http://www.cs.unc.edu/~welch/kalmanLinks.html
D. Forsyth and J. Ponce. Computer Vision: A Modern
Approach, Chapter 19.
http://www.cs.berkeley.edu/~daf/book3chaps.html
O. Faugeras.. Three-Dimensional Computer Vision. MIT
Press. Ch. 8, Tracking Tokens over Time.