CN105741277A - ViBe (Visual Background Extractor) algorithm and SLIC (Simple Linear Iterative Cluster) superpixel based background difference method - Google Patents

Abstract

The invention discloses a ViBe (Visual Background Extractor) algorithm and SLIC (Simple Linear Iterative Cluster) superpixel based background difference method and belongs to the technical field of image processing. The method comprises the following steps of superpixel segmentation based background modeling, foreground detection and background updating. The invention proposes a background modeling and foreground detection algorithm in combination with SLIC superpixels under a ViBe algorithm framework. While the advantages of the ViBe algorithm are effectively utilized, the neighbourhood range of the pixels is extended in combination with the SLIC superpixels, and the spatial relativity of background pixels is utilized more fully, so that the phenomena of any disturbance, camera shaking and the like can be better coped with. A background updating mode for randomly replacing a background library by the ViBe algorithm is changed to an updating mode for Gaussian distribution, so that the instability of the algorithm caused by a random replacement policy is avoided. Three groups of frame sequences in an I2R data set are used for performing a background modeling and foreground detection experiment in an experimental part, and a GMM (Gaussian Mixture Model) algorithm and the ViBe algorithm are compared with existing algorithms.

Description

Background subtraction based on ViBe algorithm Yu SLIC super-pixel

Technical field

The present invention relates to the background subtraction based on ViBe algorithm Yu SLIC super-pixel, belong to technical field of image processing.

Background technology

In the last few years, along with the demand of monitor in real time is grown with each passing day by people, background modeling foreground detection techniques was widely used in field of video monitoring.Meanwhile, the kind of monitoring scene also emerges in an endless stream, and arrives outdoor indoor, from static scene to the complex scene containing strong disturbance.But, complex scene also brings some challenges for background modeling foreground detection.

Illuminance abrupt variation: illuminance abrupt variation makes background pixel point produce strong variations, is therefore easily mistaken for prospect when foreground detection.This situation is very common at the outdoor scene of cloudy weather.

Background object shifts: when there is background object displacement, two places will be had in background to be judged as prospect.One place is the object of displacement, and this should be judged as prospect.Another place is the position covered before object displacement, and this should be absorbed in background as soon as possible.

Model for dynamic scene: most of background modeling methods wish that training sample is static.But, this is impossible under some scenes, such as to the highway monitor in real time that wagon flow is busy.This requires that background modeling algorithm can be modeled for dynamic scene.

In existing popular algorithm, the method based on mixed Gauss model is usable in complex scene, but is operated because each pixel is distributed multiple Gauss distribution by it, and time complexity is higher.Nonconservative more New Policy is adopted to cause him that the foreground detection effect of slowly movement is bad.Illuminance abrupt variation is sensitive.ViBe (VisualBackgroundExtractor) algorithm make use of the spatial coherence of background pixel, by allowing each pixel that its neighborhood is carried out context vault renewal so that it can tackle disturbance slight in background.But no matter being that 4 neighborhoods or 8 neighborhoods are processed, the contiguous range of single pixel is still very limited, thus the complex scene containing strong disturbance in background can not well be tackled.And in order to make the speed ViBe of algorithm adopt random manner that context vault is updated, thus bring certain unstability to algorithm.

Summary of the invention

The present invention is directed to problem above, develop the background subtraction based on ViBe algorithm Yu SLIC super-pixel.This algorithm inherits the superperformance of ViBe algorithm, make use of the spatial coherence of background pixel more fully simultaneously, reaches to successfully manage powerful disturbance in complex scene the purpose that foreground object is accurately detected.

The present invention comprises the steps:

The first step: based on the background modeling of super-pixel segmentation,

Second step: foreground detection,

3rd step: context update.

The principle of the invention and beneficial effect: 1. belong to the most pixels in the same area in natural image and obey same Gauss distribution.2. the strong disturbance overwhelming majority in complex scene is caused by the displacement (the rustle of leaves in the wind, camera shake) of neighborhood territory pixel.We propose background subtraction based on ViBe algorithm Yu SLIC super-pixel in conjunction with SLIC super-pixel in conjunction with both features under original ViBe algorithm frame.Compared with existing background subtraction, the method make use of the spatial coherence of background pixel more fully, it is possible to successfully managing powerful disturbance in complex scene, false alarm rate is lower, and degree of accuracy is higher.

Accompanying drawing explanation

Fig. 1 flow chart of the present invention.

Fig. 2 SLIC super-pixel segmentation schematic diagram.A width artwork in (a), (b), (c) respectively Fountain, Campus and WaterSurface frame sequence in figure, (d), (e), (f) are corresponding SLIC super-pixel segmentation result figure.

Fig. 3 foreground detection schematic diagram.A frame to be detected in (a), (b), (c) respectively Fountain, Campus and WaterSurface frame sequence in figure, d (), (e), (f) are road-scene, g testing result that (), (h), (i) are GMM (GaussianMixtureModel) algorithm, j testing result that (), (k), (l) they are ViBe algorithm, the testing result that (m), (n), (o) they are the method.

Fig. 4 is recall (recall rate), precision (accuracy rate) and fmeasure correlation curve,

$fmeasure=2\frac{recall\×precision}{recall+precision}---\left(3\right)$

In figure, abscissa is frame number, and vertical coordinate is the numerical value of corresponding evaluation metrics.Recall, precision, fmeasure correlation curve in (a), (b), (c) respectively Fountain experiment, recall, precision, fmeasure correlation curve in (d), (e), (f) respectively Campus experiment, recall, precision, fmeasure correlation curve in (g), (h), (i) respectively WaterSurface experiment.

Detailed description of the invention

The present invention includes four steps: based on the background modeling of super-pixel segmentation, foreground detection, context update.

The first step: based on the background modeling of super-pixel segmentation.

First with SLIC superpixel segmentation method, the first frame of video is carried out super-pixel segmentation, obtain size consistent super-pixel block substantially uniform with content.Calculate the brightness Brightness and mean flow rate aveBrightness of each pixel in super-pixel.

Brightness=0.3 × r+0.6 × g+0.1 × b (4)

$aveBrightness=\frac{1}{n}{\mathrm{\Σ}}_{i=1}^n{\mathrm{Brightness}}_i---\left(5\right)$

Wherein, n is pixel sum, Brightness in super-pixel_iFor the brightness of ith pixel point in super-pixel, r, g, b be the three-channel numerical value of pixel rgb respectively.

Utilize mean flow rate that pixel in super-pixel is divided into pixel value more than mean flow rate with less than or equal to two classes of mean flow rate.Distribute two Gauss distribution for each super-pixel and calculate the mean μ of two Gauss distribution, variance var, standard deviation std by two class pixels.Namely

$\mathrm{\μ}=\frac{1}{n}{\mathrm{\Σ}}_{i=1}^n{x}_i---\left(6\right)$

$var=\frac{1}{n}{\mathrm{\Σ}}_{i=1}^n{(x_i-\mathrm{\μ})}^2---\left(7\right)$

$std=\sqrt{var}---\left(8\right)$

Wherein, n is the pixel sum belonging to this Gauss distribution, x_iRepresent the ith pixel point belonging to this Gauss distribution.

If the pixel in super-pixel is identical, the standard deviation being computed drawing will be very little.The noise small when carrying out foreground detection all can be judged as prospect.In order to avoid this situation occurs, when calculating standard deviation, specify minima for it.If result of calculation is less than this minima, minima is used to be replaced.Experiment sets minima as 17.

Second step: foreground detection.

Its inside pixel to be detected is carried out foreground detection by two Gauss models first with super-pixel.If pixel to be detected at least meets the distance of pixel value and average with one of them Gauss model is not more than the condition of λ times of standard deviation, namely

|x-μ|≤λ×std(9)

Then it is judged to background dot；Wherein x is band detection pixel, average in μ and std respectively formula (6), (8) and standard deviation, and λ is the numerical value that we set ourselves, and in experiment, the span of λ is 0.8 to 2.

Without satisfying condition, then centered by super-pixel central point, super-pixel average side length is the interior neighborhood super-pixel of finding of eight neighborhood of radius, and the background model of use neighborhood super-pixel detects.If not still being judged as background dot, then it it is foreground point.Finally, in order to tackle illuminance abrupt variation, when the pixel number being judged as prospect exceedes total pixel number 50% time, again model.

3rd step: context update.

In the context update stage, adopt conservative update mode, be namely only not less than when pixel to be detected is judged as the distance of background dot and pixel and Gauss model averageTimes standard deviation time, namely

$|x-\mathrm{\μ}|\≥\frac{1}{2}\×\mathrm{\λ}\×std---\left(10\right)$

Just it is used for updating background model.Namely

μ=(1-α) × μ+α × x (11)

Var=(1-α) × var+ α × (x-μ)²(12)

Wherein, x is band detection pixel, average in μ, var, std respectively formula (6), (7), (8), variance, standard deviation, λ and α is the numerical value that we set ourselves, in experiment, the span of λ is 0.8 to 2, and the span of α is 0.01 to 0.5.Such more New Policy avoids original ViBe and adopts randomized policy when updating context vault and the unstability brought, simultaneously also substantial amounts of skimble-skamble when decreasing pixel to be detected with background model hypotelorism update operation, thus improve the speed of context update significantly.

The principle of the invention and beneficial effect: 1. belong to the most pixels in the same area in natural image and obey same Gauss distribution.2. the strong disturbance overwhelming majority in complex scene is caused by the displacement (the rustle of leaves in the wind, camera shake) of neighborhood territory pixel.We propose background subtraction based on ViBe algorithm Yu SLIC super-pixel in conjunction with SLIC super-pixel according to both features under original ViBe algorithm frame.Compared with existing background subtraction, the method make use of the spatial coherence of background pixel more fully, it is possible to successfully managing powerful disturbance in complex scene, false alarm rate is lower, and degree of accuracy is higher.

Experiment adopts the frame sequence (Fountain, Campus and WaterSurface) that in I2R data set, three groups is background with outdoor complex scene.Picture exists the background of motion, the such as current of fountain ejection, the branch waved, the sea that gleams of light are reflecting on waves in the river.Test recall, precision and fmeasure curve that result generates by contrasting this algorithm and original ViBe algorithm, GMM in three groups of videos proposed algorithm performance is made an appraisal.Testing result is as shown in Figure 3.Recall, precision and fmeasure correlation curve is as shown in Figure 4.Owing to the method extends the scope of neighborhood territory pixel, make use of the spatial coherence of background pixel more fully and more regular background model be updated, the flase drop brought by the strong disturbance of complex scene and camera shake in detection process is substantially reduced, and detection degree of accuracy is higher.

Claims (3)

1. the background subtraction based on ViBe algorithm Yu SLIC super-pixel, it is characterised in that:

The first step, based on the background modeling of super-pixel segmentation

First with SLIC superpixel segmentation method, the first frame of video is carried out super-pixel segmentation, obtain size consistent super-pixel block substantially uniform with content；Calculate the brightness Brightness and mean flow rate aveBrightness of each pixel in super-pixel；

Brightness=0.3 × r+0.6 × g+0.1 × b (1)

$aveBrightntess=\frac{1}{n}{\mathrm{\Σ}}_{i=1}^n{\mathrm{Brightntess}}_i---\left(2\right)$

Wherein, n is pixel sum, Brightness in super-pixel_iFor the brightness of ith pixel point in super-pixel, r, g, b be the three-channel numerical value of pixel rgb respectively；

Utilize mean flow rate that pixel in super-pixel is divided into pixel value more than mean flow rate with less than or equal to two classes of mean flow rate；Distribute two Gauss distribution for each super-pixel and calculate the mean μ of two Gauss distribution, variance var, standard deviation std by two class pixels；Namely

$\mathrm{\μ}=\frac{1}{n}{\mathrm{\Σ}}_{i=1}^n{x}_i---\left(3\right)$

$var=\frac{1}{n}{\mathrm{\Σ}}_{i=1}^n{(x_i-\mathrm{\μ})}^2---\left(4\right)$

$std=\sqrt{v\mathrm{\α}r}---\left(5\right)$

Wherein, n is the pixel sum belonging to this Gauss distribution, x_iRepresent the ith pixel point belonging to this Gauss distribution；

For specifying minima when calculating standard deviation, if result of calculation is less than this minima, minima is used to be replaced；

Second step, foreground detection；3rd step, context update.

2. a kind of background subtraction based on ViBe algorithm Yu SLIC super-pixel according to claim 1, it is characterised in that described foreground detection step is as follows,

Its inside pixel to be detected is carried out foreground detection by two Gauss models first with super-pixel；If pixel to be detected at least meets the distance of pixel value and average with one of them Gauss model is not more than the condition of λ times of standard deviation, namely

|x-μ|≤λ×std(6)

Then it is judged to background dot；Wherein x is band detection pixel, average in μ and std respectively formula (3), (5) and standard deviation, and λ is the numerical value that we set ourselves；

Without satisfying condition, then centered by super-pixel central point, super-pixel average side length is the interior neighborhood super-pixel of finding of eight neighborhood of radius, and the background model of use neighborhood super-pixel detects；If not still being judged as background dot, then it it is foreground point；Finally, when the pixel number being judged as prospect exceedes total pixel number 50% time, again model；

3rd step, context update.

3. a kind of background subtraction based on ViBe algorithm Yu SLIC super-pixel according to claim 1 and 2, it is characterised in that described context update step is as follows:

It is not less than when pixel to be detected is judged as the distance of background dot and pixel and Gauss model averageTimes standard deviation time, namely

$|x-\mathrm{\μ}|\≥\frac{1}{2}\×\mathrm{\λ}\×std---\left(7\right)$

Just it is used for updating background model；Namely

μ=(1-α+× μ+α × x (8)

Var=(1-α) × var+ α × (x-μ)²(9)

Wherein, x is band detection pixel, and the average in μ, var, std respectively formula (3), (4), (5), variance, standard deviation, λ and α is the numerical value set.