Page 137 - Classification Parameter Estimation & State Estimation An Engg Approach Using MATLAB

126 STATE ESTIMATION

The computational structure of the Viterbi algorithm is comparable to
that of the forward algorithm. The computational complexity is also on
2
the order of (i þ 1)K .

Example 4.9 Offline license plate detection in videos
Figure 4.19 shows the results of the two offline state estimators
applied to the video line shown in Figure 4.15. Figure 4.20 provides
the results of the whole image shown in Figure 4.14.
Both methods are able to prevent the delay that is inherent in online
estimation. Nevertheless, both methods show some falsely detected
license plate pixels on the right side of the plate. These errors are
caused by a sticker containing some text. Apparently, the statistical
properties of the image of this sticker are similar to the one of a license
plate.
A comparison between the individually estimated states and the
jointly estimated states shows that the latter are more coherent, and
that the former are more fragmented. Clearly, such a fragmentation
increases the probability of having erroneous transitions of estimated
states. However, usual the resulting erroneous regions are small. The
jointly estimated states do not show many of these unwanted transi-
tions, but if they occur, then they are more serious because they result
in a larger erroneous region.

true license
plate pixels:

offline individually
estimated states:

detected pixels:

offline, jointly
estimated states:

detected pixels:

Figure 4.19 Offline state estimation