WO2016165064A1

WO2016165064A1 - Robust foreground detection method based on multi-view learning

Info

Publication number: WO2016165064A1
Application number: PCT/CN2015/076533
Authority: WO
Inventors: 王坤峰; 王飞跃; 刘玉强; 苟超
Original assignee: 中国科学院自动化研究所
Priority date: 2015-04-14
Filing date: 2015-04-14
Publication date: 2016-10-20

Abstract

Provided is a robust foreground detection method based on multi-view learning, comprising: acquiring a reference background image from an input video by means of a time domain median filtering method, and performing iterative search and multi-scale integration on a current image and the reference background image, so as to acquire heterogeneous features; calculating the conditional probability density of a foreground type and the conditional probability density of a background type using the conditional independence of the heterogeneous features, and calculating the posterior probability of a foreground and the posterior probability of a background using a Bayesean rule according to a foreground likelihood, a background likelihood and a priori probability; and constructing an energy function of a Markov random field model by means of the posterior probability of the foreground, the posterior probability of the background and a time-space consistency constraint, and minimizing the energy function using a belief propagation algorithm, so as to obtain a segmentation result of the foreground and the background. The present invention can realize robust foreground detection in a complex challenging environment.

Description

Robust foreground detection method based on multi-view learning

Technical field

The invention relates to an intelligent video monitoring technology, in particular to a robust foreground detection method based on multi-view learning.

Background technique

Intelligent video surveillance is an important means of information collection, and foreground detection or background reduction is a challenging underlying problem in intelligent video surveillance research. Based on the foreground detection, other applications such as target tracking, recognition, and anomaly detection can be realized. The basic principle of foreground detection is to compare the current image of the video scene with the background model to detect areas with significant differences. Although seemingly simple, foreground detection often encounters three challenges in practical applications: motion shadows, illumination changes, and image noise. The motion shadow is caused by the occlusion of the light source by the foreground target, which is a hard shadow on a sunny day and a soft shadow on a cloudy day. Regardless of the form, the motion shadow is easily detected as foreground, interfering with the extraction of the size and shape information of the segmented foreground object. Light changes are common in traffic scenes. For example, as the sun moves through the sky, the light changes slowly; as the sun enters or moves out of the clouds, the light may change rapidly. In addition, noise is inevitably introduced during image acquisition, compression, and transmission. If the signal-to-noise ratio is too low, it will be difficult to distinguish the foreground target from the background scene.

We divide the foreground detection technology into sparse models, parametric models, nonparametric models, machine learning models, and so on. The sparse model mainly uses various variants of principal component analysis and matrix decomposition to model the background as a low rank representation and the foreground as a sparse outlier. However, the computational complexity of such methods is high and it is difficult to detect foregrounds that are similar in color to the background. The parametric model uses a certain probability distribution to model the background. Nonparametric models have greater flexibility in probability density estimation. The machine learning model uses machine learning methods such as support vector machines and neural networks to classify foreground and background.

The prior art has the following problems. First, only use the brightness feature, but the brightness feature is on the light Sensitive to changes and motion shadows. Second, only the background model is established, and the foreground pixels are identified as outliers, making it difficult to distinguish foregrounds that are similar to the background color. Third, the spatiotemporal consistency constraints in the video sequence are not utilized.

Summary of the invention

The robust foreground detection method based on multi-view learning provided by the invention can accurately realize the segmentation of the foreground and the background.

According to an aspect of the present invention, a robust foreground detection method based on multi-view learning is provided, including:

Obtaining a reference background image by using a time domain value filtering method, performing iterative search and multi-scale fusion on the current image and the reference background image to obtain a heterogeneous feature;

Calculating the conditional probability density of the foreground class and the conditional probability density of the background class by using the conditional independence of the heterogeneous feature, and calculating the posterior probability and background of the foreground using Bayesian rules according to foreground likelihood, background likelihood and prior probability Posterior probability

The energy function of the Markov random field model is constructed by the posterior probability of the foreground, the posterior probability of the background, and the spatiotemporal consistency constraint, and the energy function is minimized by the belief propagation algorithm to obtain the segmentation result of the foreground and the background. .

The multi-view learning based robust foreground detection method provided by the embodiment of the present invention can calculate the posterior probability of the foreground and the posterior probability of the background by using the Bayesian rule according to the foreground likelihood, the background likelihood and the prior probability, and The energy function of the Markov random field model is constructed by the posterior probability of the foreground, the posterior probability of the background and the spatio-temporal consistency constraint, so that the foreground and background segmentation can be accurately realized by the belief propagation algorithm.

DRAWINGS

FIG. 1 is a flowchart of a method for detecting a robust foreground based on multi-view learning according to an embodiment of the present invention;

2 is a schematic diagram of an input video image and a reference background image according to an embodiment of the present invention;

3 is a schematic diagram of a pyramid search template according to an embodiment of the present invention;

4 is a schematic diagram of texture variation features based on iterative search and multi-scale fusion according to an embodiment of the present invention;

FIG. 5 is a schematic diagram of an RGB color model according to an embodiment of the present invention; FIG.

FIG. 6 is a schematic diagram of brightness variation characteristics and chromaticity change characteristics according to an embodiment of the present invention; FIG.

FIG. 7 is a flowchart of a method for acquiring a candidate background according to an embodiment of the present invention;

FIG. 8 is a heterogeneous characteristic frequency histogram according to an embodiment of the present invention; FIG.

FIG. 9 is a schematic diagram of image marking results according to an embodiment of the present invention; FIG.

FIG. 10 is a schematic diagram of segmentation results of foreground and background according to an embodiment of the present invention.

detailed description

The multi-view learning based robust foreground detection method provided by the embodiment of the present invention is described in detail below with reference to the accompanying drawings.

FIG. 1 is a flowchart of a method for detecting a robust foreground based on multi-view learning according to an embodiment of the present invention.

Referring to FIG. 1, in step S101, an input video is acquired by a time domain value filtering method to obtain a reference background image, and an iterative search and multi-scale fusion of the current image and the reference background image are performed to obtain a heterogeneous feature.

In step S102, the conditional probability density of the foreground class and the conditional probability density of the background class are calculated using the conditional independence of the heterogeneous feature, and the foreground is calculated using the Bayesian rule according to the foreground likelihood, the background likelihood, and the prior probability. Probability and background posterior probability.

In step S103, an energy function of a Markov random field model is constructed by a posterior probability of the foreground, a posterior probability of the background, and a spatiotemporal consistency constraint, and the energy function is minimized by a belief propagation algorithm to obtain a foreground and The segmentation result of the background.

Further, the obtaining, by the input domain video, the reference background image by using a time domain value filtering method includes:

Reading each frame of the input video;

Obtaining each pixel in the threshold time window by using the time domain value filtering method Median value;

The reference background image is obtained according to a median value of each of the pixels.

Here, the threshold time window is the duration of the image of 500 frames, with specific reference to the input video image and the reference background image provided by the embodiment of the present invention as shown in FIG. 2, (a) is an input video image, and (b) is a reference background. image.

Further, the heterogeneous feature is a texture change feature, and the iterative search and multi-scale fusion of the current image and the reference background image to obtain the heterogeneous feature includes:

Calculating the texture change feature according to formula (1):

Wherein, TV _i is the texture change feature, i is a current pixel, and [I _R (i), I _G (i), I _B (i)] is a color value of a color model RGB of the current pixel, and j is a background pixel corresponding to the current pixel, [E _R (j), E _G (j), E _B (j)] is a color value of RGB of the background pixel, and m∈N(i) is the current pixel The spatial neighborhood, n ∈ N (j) is the spatial neighborhood of the background pixel.

Here, for any pixel i of the current image, it is assumed that its spatial neighborhood N(i) is 8 neighborhoods.

Texture variation features are robust to motion shadows and illumination changes, but are sensitive to dynamic backgrounds. If not properly processed, a swaying textured background area can result in large texture changes. Therefore, in order to solve the above problem, the pixel i in the current image is matched with the pixel j in the reference background image by an iterative search and a multi-scale fusion strategy.

FIG. 3 is a schematic diagram of a pyramid search template according to an embodiment of the present invention. As shown in FIG. 3, (a) is a large pyramid search template, and (b) is a small pyramid search template. The specific search process is as follows: First, the large pyramid template is used for rough search. Before the first round of iteration, the pixel point i is initialized to the center point of the search template, and each iteration needs to examine 9 positions, and the optimal position (ie, Minimizing the position of TV _i ) is set to the center point of the next iteration, repeating this iterative process until the optimal position happens to be the center point of the search template; secondly, using the small pyramid template for fine search, in the fine search process It is only necessary to examine 5 positions, and determine which pixel of the TV _i is minimized as the optimal position; finally, the pixel j in the reference background image corresponding to the pixel i of the current image is acquired.

In order to deal with the problem that the iterative search is trapped in the local minimum value, the present invention utilizes the complementary information contained in the multi-scale image for feature extraction, and specifically refers to the texture change based on iterative search and multi-scale fusion provided by the embodiment of the present invention as shown in FIG. Feature map.

First, the size of the current image and the reference background image are sequentially scaled to 1/2 times and 1/4 times of the original size, and feature extraction is performed on both the original size image and the scaled image; secondly, the features of the three scales are original Fusion is performed on the scale, and the fusion operator is a median operator.

Further, the heterogeneous feature is a brightness change feature, and the iterative search and multi-scale fusion of the current image and the reference background image to obtain the heterogeneous feature includes:

Calculating the brightness variation characteristic according to formula (2):

BV _i =(α _i -1)||OE _j || (2)

Wherein, BV _i is the brightness variation characteristic, α _i is a ratio of a brightness of the current pixel to a brightness of the background pixel, and E _j is a color value of RGB of the background pixel, and ||OE _j || The linear distance between the origin O and E _j .

Here, the difference between the current pixel and the reference background pixel in the RGB space is decomposed into the luminance variation feature BV and the chrominance variation feature CV, with reference to the RGB color model diagram provided by the embodiment of the present invention as shown in FIG. 5 . As shown in FIG. 5, for the pixel i ∈ I in the current image I, the change in luminance of I _i with respect to the reference background pixel value E _j is calculated. Let [I _R (i), I _G (i), I _B (i)] represent the RGB color value of the current pixel i, and [E _R (j), E _G (j), E _B (j)] Background RGB color value of pixel j. The specific process is: first calculating the ratio of the current pixel brightness to the background brightness α _i , α _i is known by the formula (3); secondly, the brightness variation characteristic BV _i of the pixel i is the signed distance of α _i E _j with respect to E _j , Specifically, it can be known from formula (2):

It can be seen from the formula (2) that ||OE _j || represents the linear distance between the origin O and E _j , and if the current pixel brightness is equal to the background luminance, BV _i =0. If the current pixel brightness is less than the background brightness, then BV _i <0. If the current pixel brightness is greater than the background brightness, then BV _i >0. Therefore, the brightness change BV _i reflects the difference in brightness between the current pixel and the corresponding background pixel.

Further, the heterogeneous feature is a chroma change feature, and the iterative search and multi-scale fusion of the current image and the reference background image to obtain the heterogeneous feature includes:

Calculating the chromaticity change characteristic according to formula (4):

Wherein CV _i is the chromaticity change characteristic, and α _i is a ratio of the brightness of the current pixel to the brightness of the background pixel, [I _R (i), I _G (i), I _B (i)] For the color value of RGB of the current pixel, [E _R (j), E _G (j), E _B (j)] is the color value of RGB of the background pixel.

Here, the specific process of the luminance variation feature and the chrominance variation feature based on the iterative search and the multi-scale fusion is as follows: First, the current image and the reference background image are sequentially scaled to 1/2 times and 1/4 times of the original size, in the original Feature extraction is performed on both the size image and the scaled image. Secondly, the features of the three scales are fused on the original scale to obtain the final brightness variation feature and chromaticity variation feature. For details, refer to the luminance variation feature and the chrominance variation feature provided by the embodiment of the present invention as shown in FIG. 6 .

Both BV _i and CV _i are distances in the RGB color space and have the same unit of measure. The invention directly quantizes the values of these two features into integers, and can achieve high efficiency kernel density estimation.

Since the luminance variation characteristic, the chrominance variation feature and the texture variation feature reflect the characteristics of different sides of the image, in the case of a given pixel class marker C, the probability distribution conditions of the three features are independent, and the formula (5) shows that:

p(BV, CV, TV|C)=p(BV|C)p(CV|C)p(TV|C) (5)

Wherein, the category tag C may be a foreground class or a background class.

Further, calculating the conditional probability density of the foreground class and the conditional probability density of the background class by using the conditional independence of the heterogeneous feature includes:

Calculating the conditional probability density of the foreground class according to formula (6):

p(BV|FG)=p(BV|CV>τ _CV or TV>τ _TV ),

p(CV|FG)=p(CV|BV>τ _BV or TV>τ _TV ), (6)

p(TV|FG)=p(TV|BV>τ _BV or CV>τ _CV ),

Wherein FG is the foreground class, p(BV|FG) is a probability density of the brightness change characteristic under the condition of the foreground class, and p(CV|FG) is the condition of the foreground class The probability density of the chrominance variation feature, p(TV|FG) is the probability density of the texture variation feature under the condition of the foreground class, τ _BV is the threshold of the luminance variation feature, and τ _CV is the chromaticity The threshold of the variation feature, τ _TV is the threshold of the texture variation feature.

Here, the luminance variation feature, the chrominance variation feature, and the texture variation feature select a trusted foreground pixel in the current image, accumulate and continuously update the frequency histogram of the three features, and use the multi-view learning method to estimate the conditional probability density of the foreground class. .

It can be known from formula (6) that if the value of one of the brightness change characteristic, the chromaticity change feature and the texture change feature is sufficiently large, indicating that the pixel is a trusted foreground pixel, a frequency histogram can be added to estimate other The foreground class conditional probability density of the feature. In the embodiment of the present invention, setting

τ _CV =20, τ _TV =3.6, here

Represents the median of the BV over the entire image to compensate for the global brightness change of the image.

Further, the calculating the conditional probability density of the foreground class and the conditional probability density of the background class by using the conditional independence of the heterogeneous feature further includes:

Obtaining a trusted foreground region from the current image;

Extending the trusted foreground area to expand the credible foreground area;

From the current image, an area outside the expanded trusted foreground area is used as a candidate background area, and a conditional probability of the background class is calculated according to the candidate background area. degree.

The candidate background acquisition method is specifically referred to the flowchart of the candidate background acquisition method provided by the embodiment of the present invention as shown in FIG. 7 . If the features of certain pixels in the current image satisfy BV > τ _BV or CV > τ _CV or TV > τ _TV , then these pixels belong to the trusted foreground region.

FIG. 8 is a heterogeneous characteristic frequency histogram according to an embodiment of the present invention. As shown in FIG. 8, figures a and d are luminance change characteristics, figures b and e are chromaticity change characteristics, and figures c and f are texture change characteristics. Figures a, b and c are feature frequency histograms based on ground-truth, and figures d, e and f are feature frequency histograms based on multi-view learning.

Here, using the kernel density estimation, modeling the foreground-like conditional probability density and the background-like conditional probability density, quantizing the values of the luminance change and the chrominance change into integers, and quantifying the value of the texture change feature to 0.1 interval, using a Gaussian kernel The function sets the kernel widths of the three features to σ _BV = 2.0, σ _CV = 2.0, and σ _TV = 0.2, respectively.

Further, the calculating the posterior probability of the foreground and the posterior probability of the background by using the Bayes rule according to the foreground likelihood, the background likelihood, and the prior probability include:

Calculate the posterior probability of the foreground according to formula (7):

Where P _i (FG|x) is the posterior probability of the foreground, p(x|C) is the foreground likelihood or background likelihood, and P _i (C) is the prior probability or the foreground of the foreground The prior probability of the background.

Calculate the posterior probability of the background according to formula (8):

P _i (BG|x)=1-P _i (FG|x) (8)

Where P _i (FG|x) is the posterior probability of the foreground, and P _i (BG|x) is the posterior probability of the background.

Here, the prior probabilities can be spatially distinct, with trees, buildings, and sky in the scene. Compared to the area, the road area should have a greater prospective prior probability. The prior probability can also be changed with time. If a pixel is marked as foreground more frequently than the previous period in the recent period, its foreground prior probability increases, otherwise its foreground prior probability decreases. . Therefore, the present invention constructs a dynamic prior model based on the marking result of the previous image, which is specifically known by the formula (9):

P _i,t+1 (FG)=(1-ρ)P _i,t (FG)+ρL _i,t (9)

Where P _i,t+1 (FG) is the foreground prior probability of pixel i at time t+1, P _i,t (FG) is the foreground prior probability of pixel i at time t, and L _i,t represents pixel i is the mark at time t, and ρ is the learning rate parameter.

If pixel i is marked as foreground at time t, then L _i,t =1; if pixel i is marked as background at time t, then L _i,t =0. ρ is the learning rate parameter, and ρ is set to 0.001. At system startup, P _i,t (FG) is set to 0.2.

FIG. 9 is a schematic diagram of image marking results according to an embodiment of the present invention. As shown in FIG. 9, graph a is the foreground prior probability of the pixel, and graph b is the foreground posterior probability of the pixel. As can be seen from Figure a, the road area has a greater foreground prior probability than the tree area. As can be seen from Figure b, the real foreground target area has a greater foreground a posteriori probability than other areas.

Further, the constructing the energy function of the Markov random field model by using the posterior probability of the foreground, the posterior probability of the background, and the space-time consistency constraint comprises:

Calculate the energy function according to formula (10):

Where f is the labeling process, E(f) is the energy function, D _i (f _i ) is the data item, and W(f _i , f _u ) is the smoothing term.

Here, let I be the set of pixels in the current image, and L be the set of marks. Marked as an estimate for each pixel, the estimate of the foreground is labeled 1 and the estimate of the background is labeled 0. The marking process f is to assign a flag f _i ∈L to each pixel i ∈ I. Under the Markov random field framework, the markers can change slowly in the image space, but at some locations, such as target boundaries, the markers can change rapidly, and the quality of the markers depends on the energy function E(f).

As can be seen from equation (10), N represents the edge set in the graph model structure, D _i (f _i ) is the data item, which measures the cost of assigning the label f _i to the pixel i, and W(f _i , f _u ) is smoothed. Item, which measures the cost of assigning the flags f _i and f _u to two pixels i and u that are spatially adjacent. The marker that minimizes the energy function corresponds to the maximum a posteriori estimate of the Markov random field.

The data item D _i (f _i ) consists of two parts. first part

A posteriori probability that each pixel belongs to the foreground and a posterior probability that belongs to the background, namely:

Wherein, the data item D _i (f _i ) imposes a constraint on each pixel, and the encouragement flag is consistent with the pixel observation value.

the second part

Apply a time domain consistency constraint to the tag. It is assumed that a pair of associated pixels should have the same mark in a continuous image. When calculating the optical flow, the current image (ie, the image at time t) is inversely mapped to the previous frame image (ie, the image at time t-1), so that each current pixel i∈I and the pixel in the previous frame image v association. Since the mark f _{v is} known,

It can be known from formula (12):

Where γ>0 is a weight parameter. γ is set to γ=0.5 due to the influence of noise, large motion, boundary effect, and the like.

Combine the two parts and the data item becomes

However, it should be noted that if the frame rate of the video is very low, the time domain consistency constraint will not be available, so

The smoothing term W(f _i , f _u ) encourages the spatial consistency of the markers. If two spatially adjacent pixels have different markings, there is a price to pay. Specifically, it can be known from formula (13):

Where φ=5.0 is a weight parameter, and Z(I _i , I _u ) is a decreasing function controlled by the luminance difference of the pixels i and u. The function Z is known by the formula (14):

Where σ _I is the variance parameter and σ _{I is} set to 400.

As shown in FIG. 10, the first column is the embodiment number, the second column is the original image, the third column is the foreground detection result, and the fourth column is the ground-truth. According to the quantitative analysis, the average recall rate (recall) of the present invention was 0.8271, the average precision was 0.8316, and the average F-measure was 0.8252.

The figure includes motion shadow, illumination variation, image noise and other interference. The robust foreground detection method based on multi-view learning proposed by the present invention has strong robustness, can overcome these interferences, and accurately obtain foreground detection results.

The above is only a specific embodiment of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art can easily think of changes or substitutions within the technical scope of the present invention. It should be covered by the scope of the present invention. Therefore, the scope of the invention should be determined by the scope of the appended claims.

Claims

A robust foreground detection method based on multi-view learning, characterized in that the method comprises:

Obtaining a reference background image by using a time domain value filtering method, performing iterative search and multi-scale fusion on the current image and the reference background image to obtain a heterogeneous feature;

Calculating the conditional probability density of the foreground class and the conditional probability density of the background class by using the conditional independence of the heterogeneous feature, and calculating the posterior probability and background of the foreground using Bayesian rules according to foreground likelihood, background likelihood and prior probability Posterior probability

The energy function of the Markov random field model is constructed by the posterior probability of the foreground, the posterior probability of the background, and the spatiotemporal consistency constraint, and the energy function is minimized by the belief propagation algorithm to obtain the segmentation result of the foreground and the background. .
The method according to claim 1, wherein the obtaining the reference background image by the input domain video by the time domain value filtering method comprises:

Reading each frame of the input video;

And obtaining, by using the time domain value filtering method, the median value of each pixel in the threshold time window;

The reference background image is obtained according to a median value of each of the pixels.
The method according to claim 1, wherein the heterogeneous feature is a texture change feature, and the iterative search and multi-scale fusion of the current image and the reference background image to obtain the heterogeneous feature comprises:

The texture variation feature is calculated according to the following formula:

Wherein, TV i is the texture change feature, i is a current pixel, and [I R (i), I G (i), I B (i)] is a color value of a color model RGB of the current pixel, and j is a background pixel corresponding to the current pixel, [E R (j), E G (j), E B (j)] is a color value of RGB of the background pixel, and m∈N(i) is the current pixel The spatial neighborhood, n ∈ N (j) is the spatial neighborhood of the background pixel.
The method according to claim 3, wherein the heterogeneous feature is a brightness change feature, and the iterative search and multi-scale fusion of the current image and the reference background image to obtain a heterogeneous feature comprises:

The brightness change characteristic is calculated according to the following formula:

BV i =(α i -1)||OE j ||

Wherein, BV i is the brightness variation characteristic, α i is a ratio of a brightness of the current pixel to a brightness of the background pixel, and E j is a color value of RGB of the background pixel, and ||OE j || The linear distance between the origin O and E j .
The method according to claim 4, wherein the heterogeneous feature is a chroma change feature, and the iterative search and multi-scale fusion of the current image and the reference background image to obtain the heterogeneous feature comprises:

The chromaticity change characteristic is calculated according to the following formula:

Wherein CV i is the chromaticity change characteristic, and α i is a ratio of the brightness of the current pixel to the brightness of the background pixel, [I R (i), I G (i), I B (i)] For the color value of RGB of the current pixel, [E R (j), E G (j), E B (j)] is the color value of RGB of the background pixel.
The method according to claim 1, wherein the calculating the conditional probability density of the foreground class and the conditional probability density of the background class by using the conditional independence of the heterogeneous feature comprises:

Calculate the conditional probability density of the foreground class according to the following formula:

p(BV|FG)=p(BV|CV>τ CV or TV>τ TV ),

p(CV|FG)=p(CV|BV>τ BV or TV>τ TV ),

p(TV|FG)=p(TV|BV>τ BV or CV>τ CV ),

Wherein FG is the foreground class, p(BV|FG) is a probability density of the brightness change characteristic under the condition of the foreground class, and p(CV|FG) is the condition of the foreground class The probability density of the chrominance variation feature, p(TV|FG) is the probability density of the texture variation feature under the condition of the foreground class, τ BV is the threshold of the luminance variation feature, and τ CV is the chromaticity The threshold of the variation feature, τ TV is the threshold of the texture variation feature.
The method according to claim 6, wherein the calculating the conditional probability density of the foreground class and the conditional probability density of the background class by using the conditional independence of the heterogeneous feature further comprises:

Obtaining a trusted foreground region from the current image;

Extending the trusted foreground area to expand the credible foreground area;

From the current image, an area outside the expanded trusted foreground area is used as a candidate background area, and a conditional probability density of the background class is calculated according to the candidate background area.
The method according to claim 1, wherein the calculating the posterior probability of the foreground and the posterior probability of the background using the Bayes rule according to the foreground likelihood, the background likelihood, and the prior probability include:

Calculate the posterior probability of the foreground according to the following formula:

Where P i (FG|x) is the posterior probability of the foreground, p(x|C) is the foreground likelihood or background likelihood, and P i (C) is the prior probability or the foreground of the foreground The prior probability of the background.
The method according to claim 8, wherein the calculating the posterior probability of the foreground and the posterior probability of the background using the Bayes rule according to the foreground likelihood, the background likelihood, and the prior probability include:

Calculate the posterior probability of the background according to the following formula:

P i (BG|x)=1-P i (FG|x)

Where P i (FG|x) is the posterior probability of the foreground, and P i (BG|x) is the posterior probability of the background.
The method of claim 1 wherein said utilizing said foreground The posterior probability, the posterior probability of the background, and the space-time consistency constraint construct the energy function of the Markov random field model including:

The energy function is calculated according to the following formula:

Where f is the labeling process, E(f) is the energy function, D i (f i ) is the data item, and W(f i , f u ) is the smoothing term.