Lucas-Kanade in A Nutshell: 1 Motivation

Lucas-Kanade in a Nutshell
Prof. Dr. Raúl Rojas∗
1 Motivation
The Lucas-Kanade optical flow algorithm is a simple technique which can

provide an estimate of the movement of interesting features in successive
images of a scene. We would like to associate a movement vector (u, v) to
every such ”interesting” pixel in the scene, obtained by comparing the two
consecutive images.
The Lucas-Kanade algorithm makes some implicit assumptions:
– The two images are separated by a small time increment ∆t, in such a
way that objects have not displaced significantly (that is, the algorithm
works best with slow moving objects).
– The images depict a natural scene containing textured objects exhibit-

ing shades of gray (different intensity levels) which change smoothly.
The algorithm does not use color information in an explicit way. It does not
scan the second image looking for a match for a given pixel. It works by
trying to guess in which direction an object has moved so that local changes
in intensity can be explained.
∗
Freie Universität Berlin, Dept. of Computer Science, Arnimallee 7, 14195 Berlin,
Germany
1
2 Technique
Assume that we watch a scene through a square hole. The intensity a visible
through the hole is variable.
mask
a - increasing brightness
?
increasing brightness
In the next frame the intensity of the pixel has increased to b. It would be
sensible to assume that a displacement of the underlying object to the left
and up has occurred so that the new intensity b is now visible under the
square hole.
v
6
movement u
b
mask
If we know that the increase in brightness per pixel at pixel (x, y) is Ix (x, y)
is the x-direction, and the increase in brightness per pixel in the y direction
is Iy (x, y), we have a total increase in brightness, after a movement by u
pixels in the x direction and v pixels in the y direction of:
Ix (x, y) · u + Iy (x, y) · v
This matches the local difference in intensity (b − a) which we call It (x, y),
so that
Ix (x, y) · u + Iy (x, y) · v = −It (x, y)
The negative sign is necessary because for positive Ix , Iy , and It we have a
movement to the left and down (think about this for a minute).
2
3 Neighborhoods
Of course, a simple pixel does not usually contain enough ”structure” useful
for matching with another pixel. It is better to use a neighborhood of pixels,
for example the 3 × 3 neighborhood around the pixel (x, y). In that case we
set 9 linear equations:
Ix (x + ∆x, y + ∆y) · u + Iy (x + ∆x, y + ∆y) · v = −It (x + ∆x, y + ∆y)

for ∆x = −1, 0, 1 and ∆y = −1, 0, 1
The linear equations can be summarized as the matrix equality

!
u →
S =t
v

where S is a 9 × 2 matrix containing the rows Ix (x + ∆x, y + ∆y), Iy (x +

→
∆x, y + ∆y) and t is a vector containing the 9 terms −It (x + ∆x, y + ∆y).
The above equation cannot be solved exactly (in the general case). The Least
Squares solution is found by multiplying the equation by S T
!
T u →
S S = ST t
v
and inverting S T S, so that

!
u →
= (S T S)−1 S T t
v
4 Invertibility
The solution given above is the best possible, whenever S T S is invertible.

This might not be the case, if the pixel (x, y) is located in a region with
no structure (for example, if Ix , Iy and It are all zero for all pixels in the
3
neighborhood). Even if the matrix is invertible it can be ill conditioned, if
its elements are very small and close to zero.
One way of testing how good the inverse of S T S for our purposes is, is to
look at the eigenvalues of this matrix. S T S is a symmetrical matrix, and as
such can be diagonalized and written in the form
!
T λ1 0
S S=U UT
0 λ2
where U is a unitary 2 × 2 matrix. If S T S is not invertible, λ1 or λ2 , or both,

are zero. If one of them, or both, are very small, then the inverse matrix is
ill-conditioned.
Testing the size of the eigenvectors can be done by solving the characteristic
equation
det(S T S − λI) = 0
which reduces to
 X X 
Ix2 (x + ∆x, y + ∆y) − λ Ix (x + ∆x, y + ∆y)Iy (x + ∆x, y + ∆y)
 N N

det  X X =0
 Ix (x + ∆x, y + ∆y) · Iy (x + ∆x, y + ∆y) Iy2 (x + ∆x, ` + ∆y) − λ 
N N
This is a quadratic equation for λ, which can be readily solved. The sums
are computed over the neighborhood of pixels N .
5 Conclusions
The Lucas-Kanade algorithm makes a ”best guess” of the displacement of a

neighborhood by looking at changes in pixel intensity which can be explained
from the known intensity gradients of the image in that neighborhood. For a
simple pixel we have two unknowns (u and v) and one equation (that is, the
system is underdetermined). We need a neighborhood in order to get more
equations. Doing so makes the system overdetermined and we have to find
a least squares solution. The LSQ solution averages the optical flow guesses
over a neighborhood.
4
We assume that all intensity changes can be explained by intensity gradients.
The method breaks down when the gradients are random (think of an image
of random points) or when the gradients are negligible (no structure, as in flat
surfaces). The Lucas-Kanade algorithm eliminates regions without structure
by looking at the invertibilily of the matrix S T S in an indirect way, that is,
through the eigenvalues of this matrix.
The result of the algorithm is a set of optical flow vectors distributed over
the image which give an estimation idea of the movement of objects in the
scene. Of course, some optical flow vectors will be erroneous.
The main advantage of the algorithm is, that for a neighborhood of fixed size,
→
the number of operations needed to compute (S T S)−1 S T t are constant,
and therefore the complexity of the algorithm is linear in the number of
pixels examined in the image. Alternative algorithms that match similar
regions using a neighborhood, and scanning the second image, have quadratic
complexity.
Summarizing: The Lucas-Kanade algorithm is an efficient method for ob-

taining optical flow information at interesting points in an image (i.e. those
exhibiting enough intensity gradient information). It works for moderate
object speeds.
References:
B.D. Lucas, T. Kanade, ”An Image Registration Technique with an
Application to Stereo Vision”, in Proceedings of Image Understanding
Workshop, 1981, pp. 121-130.

Lucas-Kanade in A Nutshell: 1 Motivation

Uploaded by

Copyright:

Available Formats

Lucas-Kanade in A Nutshell: 1 Motivation

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Lucas-Kanade in A Nutshell: 1 Motivation

Uploaded by

Copyright:

Available Formats

Lucas-Kanade in a Nutshell

Prof. Dr. Raúl Rojas∗

The Lucas-Kanade optical flow algorithm is a simple technique which can

The Lucas-Kanade algorithm makes some implicit assumptions:

– The images depict a natural scene containing textured objects exhibit-

Ix (x + ∆x, y + ∆y) · u + Iy (x + ∆x, y + ∆y) · v = −It (x + ∆x, y + ∆y)

The linear equations can be summarized as the matrix equality

and inverting S T S, so that

The solution given above is the best possible, whenever S T S is invertible.

where U is a unitary 2 × 2 matrix. If S T S is not invertible, λ1 or λ2 , or both,

The Lucas-Kanade algorithm makes a ”best guess” of the displacement of a

Summarizing: The Lucas-Kanade algorithm is an efficient method for ob-

You might also like