Digital Signal Processing Reference
In-Depth Information
Fig. 16.8
MAO architecture for dynamic operation of graphs
This section contains a short overview of the most popular methods for feature
tracking, especially those which can be implemented in embedded hardware. These
methods are essentially correlation-based but they may use in addition qualitative
motion heuristics and Bayesian reasoning.
16.3.1 Correlation-Based Feature Tracking
In general, it is very difficult to estimate displacements from one image to another.
The difficulties are caused by changes in illumination, the complexity of the dis-
placements, and the effects of un-modeled noise. One solution is to track salient
points rather than an entire image region, on the grounds that salient points can be
matched even if the noise level is high.
One of the most popular methods for tracking is proposed by [
55
]. The method
involves the matching of relatively small squares of pixels from one image to the
next. The displacements are assumed to be small translations, usually of the order
of a pixel. This assumption is reasonable if the movement is slow compared to the
camera sampling speed. The windows most suited to matching are those in which
the eigenvalues of the matrix in Eq.
16.4
) are above a threshold [
83
] and have a
weak dynamics:
gg
T
w
=
G
(16.4)
F
where
F
is a given window,
g
is the luminosity gradient in
F
as a function of posi-
tion, and
w
is a weighting function. The displacement of each window is determined
by Eq. (
16.5
):
Gd
=
hgw
(16.5)
F
Search WWH ::
Custom Search