vibe algorithm is the use of pixel areas to create a background model, background model and the current input pixel values to detect foreground by comparison.
Works model
Select background pixel samples (the point in the past field and pixels thereof): The points in the neighborhood using the eight-neighbor method of selecting randomly. X pixels in the image acquired at the values given Euclidean color space, each background pixel x a set of N sample values of the background is represented by V (x) is modeled
M(x)={v1,v2,……vN}
(X) the pixel value V (x) are classified according to the model M, to define a V (x) of radius R centered on the spherical surface S (v (x)), which is compared with the value closest to the sample . The predetermined threshold value is referred to as #min, if
#{SR(v(x))∩M(x)}>#min
The background pixel point. On the contrary, the former attractions.
Model initialization
Background initialization is the process of establishing model, a video sequence detection algorithm generally requires a certain length of study is completed, the impact of real-time detection, and when a sudden change in the video screen to re-learn the background model takes a long time.
vibe algorithm is the use of a single frame of the video sequence initialization background model, for a pixel, adjacent pixels have binding spatial distribution characteristics close to the pixel values, the pixel values of randomly selected points in the neighborhood of its sample value as its model. We assume that a first frame index t = 0, N G (x) is a spatial neighborhood of the pixel position x, position y is selected randomly in accordance with the law of unity.
M0 (x) = {v0 (y | y∈N G (x)}
Disadvantages: the emergence of a moving object in a first frame will be introduced Ghost.
Update Model
The conservative update strategy: the pixel samples only when the sample is classified as the background, in order to be included in the background model, and attractions never before been used to fill the background model. This method ensures accurate detection of moving objects, provided that they are not similar to the background color. The disadvantage is that the situation will lead to a deadlock and Ghost. The solution is introduced into the space or the information of interest before counting method.
Blind update strategy: regardless of whether the sample is classified as the background are added to the background model. The advantage is not sensitive to the deadlock, the difference disadvantage is slow moving target detection capability, gradually slow moving targets are included in the background model. One possible solution is to use a large pixel size model, which covers the time window. But this is increased memory usage and higher computational cost for the price.
vibe algorithm is a conservative policy update method of combining spatial information, you may also be added to the foreground counting method.
Prospects counting method: The statistics pixel, if a pixel N times in a row is detected as a prospect, it is updated to the background points.
This updated policy includes three features:
No memory update: ensure a probability sample value remains in the sample set of monotonic decay, this method is combined with the conservative update strategy, prospects for value should not be included in the model. The new replacement value randomly selected sample, rather than first replace the old value. After the time dt is the probability sample is retained
P(t,t+dt)=[(N-1)/N](t+dt)-t
It can also be written as
P(t,t+dt)=e-ln(N/(N-1))dt
This equation shows that independent of the sample in the time dt is reserved probability t, that is, the past has no effect on the future.
Time sub-sampling: do not update the background frame for each new model, when a pixel is divided into the background, it is some probability update the background model.
Spatial neighborhood update strategy: Since we are using conservative update strategy must address its shortcomings introduction of spatial information. When a pixel is selected sample values to update the background model, and some also to update the probability that the sample values of the neighborhood. The method utilizes pixel values of the spatial propagation characteristics, background model gradually spread out, the Ghost facilitate rapid identification.