The usual approach to computing optical flow is to embed the estimator within a multi-resolution framework [ 1 , 9 , 4 ] and allow motion estimates from higher coarse-resolution levels in the hierarchy act as initial solutions at lower high-resolution levels. The primary motivation for such a complex architecture is to enable the method to detect and estimate large visual motions reliably.
In the previous section we demonstrated that the estimator derived in
section
2
can handle arbitrary large motions given a reasonably accurate initial
estimate. A
feed forward
architecture based on the above affine motion model has been developed[
7
] that takes advantage of the fact that the changes in motion from one
frame to the next (
i.e.
accelerations
) are often small. Motion estimates
derived for the current frame are fed forward to act as the initial
displacement field
of the next frame. From this data the estimator of equation
9
is used to generate the motion parameters
of the next frame. This approach is therefore
spatio-temporal
in character and eliminates the need for multi-resolution. The proposed
architecture is illustrated in figure
4
.
The warping procedure (labelled
in figure
4
) is required to create the future displacement field
from the affine motion parameters computed for the current frame based
on the assumption that the velocities remain constant
i.e.
As the motion model is affine, each rectangular neighbourhood in the
current image may be projected to a quadrilateral in the next. An
initial velocity field for the next image may then be generated for
pixels within these projected quadrilateral regions using the predicted
affine parameters computed as follows. For the motion model of equation
10
,
can be redefined as
Thus the predicted affine parameters
may be defined as
A warped image
is now recovered using bilinear interpolation.
Graeme Jones