Kanade–Lucas–Tomasi Tracking (KLT tracker)

advertisement
Kanade–Lucas–Tomasi Tracking (KLT tracker)
Tomáš Svoboda, svoboda@cmp.felk.cvut.cz
Czech Technical University in Prague, Center for Machine Perception
http://cmp.felk.cvut.cz
Last update: September 17, 2008
importance for Computer Vision
gradient based optimization
good features to track
Talk Outline
experiments
Importance in Computer Vision
Improved many times, most importantly by Carlo Tomasi [5, 4]
Free implementation(s) available1.
Firstly published in 1981 as an image registration method [3].
2/22
1
2
After more than two decades, a project2 at CMU dedicated to this
single algorithm and results published in a premium journal [1].
Part of plethora computer vision algorithms.
http://www.ces.clemson.edu/~stb/klt/
http://www.ri.cmu.edu/projects/project_515.html
Tracking of dense sequences — camera motion
3/22
I
J
Tracking of dense sequences — object motion
4/22
I
J
Alignment of an image (patch)
5/22
Goal is to align a template image T (x) to an input image I(x). x column
vector containing image coordinates [x, y]>. The I(x) could be also a small
subwindow withing an image.
Original Lucas-Kanade algorithm I
6/22
Goal is to align a template image T (x) to an input image I(x). x column
vector containing image coordinates [x, y]>. The I(x) could be also a small
subwindow withing an image.
Set of allowable warps W(x; p), where p is a vector of parameters. For
translations
x + p1
W(x; p) =
y + p2
W(x; p) can be arbitrarily complex
The best alignment minimizes image dissimilarity
X
x
[I(W(x; p)) − T (x)]2
Original Lucas-Kanade algorithm II
7/22
X
[I(W(x; p)) − T (x)]2
x
is a nonlinear optimization! The warp W(x; p) may be linear but the pixels
value are, in general, non-linear. In fact, they are essentially unrelated to x.
It is assumed that some p is known and best increment ∆p is sought. The
the modified problem
X
[I(W(x; p + ∆p)) − T (x)]2
x
is solved with respect to ∆p. When found then p gets updated
p ← p + ∆p
Original Lucas-Kanade algorithm III
8/22
X
[I(W(x; p + ∆p)) − T (x)]2
x
linearized by performing first order Taylor expansion
X
x
∂W
[I(W(x; p)) + ∇I
∆p − T (x)]2
∂p
∂I ∂I
, ∂y ] is the gradient image3 computed at W(x; p). The term
∇I = [ ∂x
the Jacobian of the warp.
3
∂W
∂p
is
As a vector it should have been a column wise oriented. However, for sake of clarity of equations row
vector is exceptionally considered here.
Original Lucas-Kanade algorithm IV
9/22
Derive
X
x
∂W
[I(W(x; p)) + ∇I
∆p − T (x)]2
∂p
with respect to ∆p
> X
∂W
∂W
I(W(x; p)) + ∇I
∆p − T (x)
2
∇I
∂p
∂p
x
setting equality to zero yields
>
X
∂W
−1
[T (x) − I(W(x; p))]
∇I
∆p = H
∂p
x
where H is the Hessian matrix
H=
X
x
>
∂W
∂W
∇I
∇I
∂p
∂p
The Lucas-Kanade algorithm—Summary
10/22
Iterate:
1. Warp I with W(x; p)
2. Warp the gradient ∇I with W(x; p)
3. Evaluate the Jacobian
image ∇I ∂∂W
p
∂W
∂p
4. Compute the Hessian H =
5. Compute ∆p = H
h
P
−1
x
at (x; p) and compute the steepest descent
P h
∇I
x
∇I
∂W
∂p
i>
∂W
∂p
∇I
∂W
∂p
i
[T (x) − I(W(x; p))]
6. Update the parameters p ← p + ∆p
until k∆pk ≤ i> h
Example of convergence
11/22
video
Example of convergence
12/22
Convergence video: Initial state is within the basin of attraction
Example of divergence
13/22
Divergence video: Initial state is outside the basin of attraction
What are good features (windows) to track?
14/22
How to select good templates T (x) for image registration, object tracking.
−1
∆p = H
X
x
∂W
∇I
∂p
>
[T (x) − I(W(x; p))]
where H is the Hessian matrix
H=
X
x
>
∂W
∂W
∇I
∇I
∂p
∂p
The stability of the iteration is mainly influenced by the inverse of Hessian.
We can study its eigenvalues. Consequently, the criterion of a good feature
window is min(λ1, λ2) > λmin (texturedness).
What are good features (windows) to track?
15/22
Consider translation W(x; p) =
1 0
∂W
∂p =
0 1
x + p1
. The Jacobian is then
y + p2
i> h
P h
i
∂W
∂W
∇I
∇I
x
∂p "
∂p
#
∂I
P
1 0
1 0
∂I ∂I
∂x
[ ∂x , ∂x ]
=
∂I
x
0 1
0 1
∂y


2
2
∂I
∂I
P
∂x
∂x∂y
2 

=
2
x
∂I
∂I
H =
∂x∂y
∂y
The image windows with varying derivatives in both directions.
Homeogeneous areas are clearly not suitable. Texture oriented mostly in one
direction only would cause instability for this translation.
What are the good points for translations?
16/22
The Hessian matrix

H=
∂x
X

x
∂I 2
2
∂I
∂x∂y
2
∂I
∂x∂y

2 
∂I
∂y
Should have large eigenvalues. We have seen the matrix already, where?
What are the good points for translations?
16/22
The Hessian matrix

H=
∂x
X

x
∂I 2
2
∂I
∂x∂y
2
∂I
∂x∂y

2 
∂I
∂y
Should have large eigenvalues. We have seen the matrix already, where?
Harris corner detector [2]!
Experiments - no occlusions
17/22
video
Experiments - occlusions
18/22
video
Experiments - occlusions with dissimilarity
19/22
video
Experiments - object motion
20/22
video
References
21/22
[1] Simon Baker and Iain Matthews. Lucas-Kanade 20 years on: A unifying framework. International Journal
of Computer Vision, 56(3):221–255, 2004.
[2] C. Harris and M. Stephen. A combined corner and edge detection. In M. M. Matthews, editor,
Proceedings of the 4th ALVEY vision conference, pages 147–151, University of Manchaster, England,
September 1988. on-line copies available on the web.
[3] Bruce D. Lucas and Takeo Kanade. An iterative image registration technique with an application to
stereo vision. In Proceedings of the 7th International Conference on Artificial Intelligence, pages 674–679,
August 1981.
[4] Jianbo Shi and Carlo Tomasi. Good features to track. In IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), pages 593–600, 1994.
[5] Carlo Tomasi and Takeo Kanade. Detection and tracking of point features. Technical Report
CMU-CS-91-132, Carnegie Mellon University, April 1991.
End
22/22
Download