15.01.2013 Views

U. Glaeser

U. Glaeser

U. Glaeser

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

FIGURE 28.22 The (stylized) power spectra of a set of 3-D Gabor filters.<br />

FIGURE 28.23 Velocity estimation in the frequency domain via estimation of the slope of the spectrum.<br />

Optical Flow Using 3-D Gabor Filters<br />

Heeger [29] proposed the use of 3-D Gabor filters to determine this slope. Following the definition<br />

discussed for 2-D, a 3-D Gabor filter has the impulse response<br />

where<br />

© 2002 by CRC Press LLC<br />

ξt<br />

g( x, y, t)<br />

= gˆ x, y, t<br />

gˆ ( x, y, t)<br />

ξy<br />

ξt<br />

(28.53)<br />

(28.54)<br />

To detect motion in different directions, a family of these filters is defined, as shown in Fig. 28.22.<br />

In order to capture velocities at different scales (high velocities can be thought of as occurring over large<br />

scales, because a large distance is covered per unit time), these filters are applied to a Gaussian pyramidal<br />

decomposition of the sequence. Given the energies of the outputs of these filters, which can be thought<br />

of as sampling spatiotemporal/spatiotemporal-frequency space, the problem is analogous to that shown<br />

in Fig. 28.23. The slope of the line (corresponding to the slope of the plane which characterizes motion)<br />

ξx<br />

( )e j2π[ξ x (x − x<br />

0 0 ) + ξ (y − y y0 0 ) + ξ (t − t t0 0 )]<br />

1<br />

( 2π)<br />

3/2 =<br />

--------------------------------e<br />

σxσ yσt 1 x − x<br />

– -- ⎛ 0<br />

------------- ⎞<br />

2 ⎝ σ ⎠<br />

x<br />

2 y − y ⎛ 0<br />

------------- ⎞<br />

⎝ σ ⎠<br />

y<br />

2 t − t ⎛ 0<br />

----------- ⎞<br />

⎝ σ ⎠<br />

t<br />

2<br />

+ +<br />

ξx

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!