Page 129 - Classification Parameter Estimation & State Estimation An Engg Approach Using MATLAB

118 STATE ESTIMATION

The probability P(Z(i), x(i)) follows from the forward algorithm.

Example 4.8 Online license plate detection in videos
This example demonstrates the ability of HMMs to find the license plate
of a vehicle in a video. Figure 4.14 is a typical example of one frame of
such a video. The task is to find all the pixels that correspond to the license
plate. Such a task is the first step in a license plate recognition system.
A major characteristic of video is that a frame is scanned line-
by-line, and that each video line is acquired from left to right. The
real-time processing of each line individually is preferable because
the throughput requirement of the application is demanding. Therefore,
each line is individually modelled as an HMM. The hidden state of a
pixel is determined by whether the pixel corresponds to a license plate
or not.
The measurements are embedded in the video line. See Figure 4.15.
However, the video signal needs to be processed in order to map it
onto a finite measurement space. Simply quantizing the signal to a
finite number of levels does not suffice because the amplitudes of the
signal alone are not very informative. The main characteristic of a
license plate in a video line is a typical pattern of dark-bright and
bright-dark transitions due to the dark characters against a bright
background, or vice versa. The image acquisition is such that the
camera–object distance is about constant for all vehicles. Therefore,
the statistical properties of the succession of transitions are typical for
the imaged license plate regardless of the type of vehicle.

Figure 4.14 License plate detection