Parameters of
three-dimensional motion (rotation, translation) play an important role in
image motion modeling. If these parameters are known precisely, the object
motion can be accurately predicted. The 3D relative movement of objects and
camera induces 2D motion on the image plane via a suitable projection system
called as apparent motion or optical flow. Motion is perceived when a change in
the image intensity on the screen is seen.
Motion estimation
methods
• feature/region
matching methods
• pel-recursive
methods
• deterministic
model based methods
• stochastic model
based methods
• optical flow based
methods
1. feature/region matching methods
Estimate motion by
correlating/matching features or particular regions of the sequence from one
frame to another.
Matching process –
Assumed
that pixels belonging to the block are displaced with the same amount, which
means a requirement of smoothness and decreasing side information on motion
vectors.
The
metrics used to perform motion estimation by block matching can be either
maximizing the cross-correlation function or minimizing an error criterion.
There are two type of search
strategies
1. full-search (exhaustive)
The former is an
exhaustive search within a predetermined maximum displacement range.
2. suboptimal (non-exhaustive)
The latter is a
non-exhaustive search procedure specified to decrease the search time,
developed in several algorithms.
a.
2D
logarithmic search
·
Stepwise implementation
·
evaluating
the matching criterion for five points (the center of the search window and the
four midpoints between the center and the four borders)
·
winner-one showing minimum dissimilarity
among them
·
other four points around the winner with
unchanged distance are taken, unless the current central point or a boundary
point give a minimum matching criterion value. In this case, the distance among
the five points has to be reduced. The minimum is reached when the last set of
points is on a 3 × 3 2D grid.
b.
Conjugate
direction search
·
two step approach,
·
matching
criterion is evaluated separately by rows and then by columns, or vice versa
·
The
row coordinate is fixed at an initial position, and the minimum of the chosen
criterion is found along the horizontal direction. Then, the same kind of
evaluation is done for the vertical direction starting from the column coordinate
obtained by the previous step.
c.
Multiresolution
block matching
There are pyramid with set
of images, taken at different resolutions
lowest level is the
original image itself,
at each step, the
upper level is obtained low-pass filtering the lower one using particular
weights, and then subsampling by 2×2
search algorithm
finds the initial motion vectors estimation starting from the top,
and then refine
them on, making them propagate to the next level. Each level gives an
improvement in the
evaluation of the vectors.
d.
Hierarchical
block matching
Method first
applied to the two frames at their lowest resolution level. (From low to high)
the result of each
level is passed onto the next higher resolution level and used as initial estimates
of the displacement vector field. These initial values are then updated by the
values
lower resolution
levels purpose - estimate the major part
of the displacement;
higher resolution
levels purpose - fine-tune the estimates
obtained at lower resolution levels.
e.
Overlapped
block matching
·
to reduce or totally eliminate blocking
artefacts.
·
applying
a windowing function to the error and then evaluated the “windowed MAD” to find
the best match and consequently the displacement vector.
·
This
method leads to both a good trade-off between higher computational cost and
also less prediction error.
2.Pel-recursive methods
Aims to estimate the displacement
vectors for each pixel of a frame in a recursive manner. This is done using a
recursive method (called descent method) that moves pixel by pixel both
spatially (horizontally and vertically) and temporally (same position in two
consecutive frames).
Several algorithms have been
developed in this field, which mainly differ on the
kind of approach
they implement.
·
Netravali-Robbins
It is based on the steepest descent
algorithm and it aims to estimate the motion
vectors minimizing the squared DFD (Displacement
Frame Difference). The estimated displacement is defined in a recursive
way
·
Bergmann
He proposed a modified version of
the Netravalli-Robbins by using the Newton-
Raphson’s algorithm instead of the
steepest descent one. The advantage is that the proposed
method converges
to the minimum faster than using the steepest descent one
3.Optical flow based methods
Several techniques
have been developed to approach the computation of optical flow
- ·
Gradient-based
approach
The first method
being proposed is the Horn and Schunck
one. The method initially makes the assumption of moving object and moving
sensor. The second method is the Lucas
and Kanade one. The iteration is not done in a strict pixel by pixel way,
but considering a small neighborhood of the pixel.
- ·
Correlation-based
approach
The frame is here
divided into fixed-size non overlapping blocks and a matching by a searching
window defined by some a priori rule is performed through the image. The displacement
to the block that shows the highest likelihood or the lowest dissimilarity is
taken as measure of the translation a motion vector to our reference block.
- ·
Spatio-temporal
energy based approach
Gabor filter is
able to select some frequency components from the signal it is filtering. The
main idea is to detect motion in the frequency domain, usingparticular
spatiotemporal filters, Gabor energy filters
