Digital Signal Processing Reference
In-Depth Information
4
Improvement of Mixture Gaussian Model Algorithm
Mixture Gauss model effectively solves the pixel multi peak distribution problem, can
be accurately modeling in the complex background image with the light gradient and
branches shaking, but it still has the following disadvantages: 1. When the object in
the background suddenly began to exercise, the revealed background area will be
false detected as moving targets, to form the "ghost". 2. When the moving object
became static, in a very long period of time it will still be judged as foreground, to
form the "smear". 3. Light mutation. This paper uses an improved mixture Gauss
model to solve these defects effectively.
4.1
The Elimination of "ghosting" and "smear"
The Gauss mixture model determine the foreground and background according to the
Image pixel values. In the background, if the original stationary objects have a
suddenly movement (such as a car drive away), the background which is covered by
objects is revealed, these revealed part obviously can not match the long training
background distribution so will be determined as the foreground to produce
"ghosting". Similarly, when the moving object gradually be stationary into the
background, in a long period of the time, it will still be judged as the foreground to
produce "smear ". We can use the idea of the frame difference method to improve the
algorithm to achieve the elimination of the above phenomenon.[7-8]
Figure 1(a) is the schematic diagram for the phenomenon of "ghosting". When
object C1 moves from the frame k to C2 in frame k+1, the revealed background is A,
the detected foreground is A+C2. Then object moves from C2 to C3 in frame k+2, the
revealed background is A+B, the detected foreground is A+B+C3. It is easy to know
that region A + C2 and the region A + B + C3 in Part A is the same, so make the
foreground image of the frame k+2 and the hole image of the frame k+1 difference.
Using appropriate threshold, if the difference does not exceed the threshold, the area
A should be background. First correct error detection area in frame k+1 and establish
a new Gauss model for the pixels on area A. Then take the current pixel value as the
mean, take appropriate variance and weights to replace the lowest priority Gauss
distribution of the original background distribution, updated them into the background
distribution. So that we can eliminate the phenomenon of "ghost".
Figure 2(b) is the schematic diagram for the phenomenon of "smear". The object
was still moving in frame k, the detected foreground is C1. The object has stopped
since the frame k+1. Because of the update rule of the mixture Gaussian model, the
background update rate of the area detected as foreground object is slow. So after this
frame, object C will still be detected as the foreground in a long time. According to
the above method, to make the foreground image C3 of the frame k+2 and the hole
image of the frame k+1 difference, we will find the C3 area matches with the previous
frame. It means the object has been static. First correct the foreground image of frame
k +1 as the background and establish a new Gauss model for the pixels on area C2.
Then take the current pixel value as the mean, take appropriate variance and weights
to replace the lowest priority Gauss distribution of the original background
distribution, updated them into the background distribution. So that we can eliminate
the phenomenon of "smear".
