CN101777186A - Multimodality automatic updating and replacing background modeling method - Google Patents

Abstract

The invention discloses a multimodality automatic updating and replacing background modeling method which comprises the following steps: firstly, carrying out modeling of a main background and an auxiliary background, wherein the modeling of the main background is carried out by the following steps: initializing a main background model, modifying the main background model, and updating threshold; and the modeling of the auxiliary background is carried out by the following steps: establishing an alternate auxiliary background sequence, counting classified data, and updating the threshold of the auxiliary background model; and then computing the background to be updated, and determining the background to replace the auxiliary background according to the occurring frequency of the background to be updated. The method of the invention adopts a multimodality updating and replacing design concept, adopts various modalities to form a vector to carry out the modeling of the background, and finishes background adaptation to the change of ambient lighting through continuous updating and replacing among the modalities. The background modeling method of the invention is suitable for an intelligent monitoring system, and adopts a background differential method for detecting a moving target.

Description

The background modeling method that a kind of multimodality automatic updating is replaced

Technical field

The invention belongs to technical field of video monitoring, relate to a kind of background modeling method that video frequency motion target detects that is used for, be specifically related to the background modeling method that a kind of multimodality automatic updating is replaced.

Background technology

Moving object detection is to finish the link of the most critical that goal behavior is analyzed in the intelligent monitor system, considers to monitor the uncertainty of moving target on direction of motion, speed in the visual field, adopts the background subtraction point-score to finish motion target detection usually mostly.Yet adopt the background subtraction point-score to carry out the effect of moving object detection, depend on whether background model is effective.Because in background, though scenery does not change, but unforeseen variation takes place in the illumination meeting of environment, except that the variation of natural lighting slowly, also exist as cross bursts such as cloud, wind leaf shake to change, for this reason, the background model that just needs be set up has certain adaptability, though had several different methods such as mixed Gauss model to realize having certain adaptive background modeling at present, still existed many problems in actual applications.

Summary of the invention

The purpose of this invention is to provide the background modeling method that a kind of multimodality automatic updating is replaced, solved the burst variation adaptability not high problem of existing background modeling method environment.

The technical solution adopted in the present invention is, the background modeling method that a kind of multimodality automatic updating is replaced is specifically implemented according to following steps:

Step 1: the modeling of main background and auxiliary background:

Main background modeling is specifically implemented according to following steps:

A. the initialization of main background model;

B. main background model is revised;

C. upgrade main background model threshold value;

The auxiliary background modeling is specifically implemented according to following steps:

A. set up alternate auxiliary background sequence;

B. grouped data is added up;

C. upgrade the threshold value of auxiliary background model;

Step 2: the calculating of background to be updated:

According to following formula, carry out context update:

F _k=[f _k(i, j)] _{M * n}Be the present frame of monitoring video, B=[b (i, j)] _{M * n}Be background model, O _k=[o _k(i, j)] _{M * n}Be the target detection result of present frame, th is the background judgment threshold.

Characteristics of the present invention also are,

The initialization of the main background modeling a of step 1 wherein main background model in the step, specifically implement according to following steps:

Get continuous N frame video and carry out background study, the size of N satisfies in the sequence of frames of video, and each pixel not frame number of passive movement target occlusion accounts for 98% of totalframes, adopts single Gaussian processes, obtains the initial value B of main background _Main=[b _Main(i, j)] _{M * n}, m, n are respectively the line number and the columns of video frame images, b _Main(i, j)=μ (i, j), i=1,2 ..., m, j=1,2 ..., n, wherein, μ (i, j) be N frame video frame images point (i, the j) average on, promptly $\mathrm{\μ}(i,j)=\frac{1}{N}\underset{k=1}{\overset{N}{\mathrm{\Σ}}}{f}_k(i,j),$ f _k(i j) is the k frame video image, and add up this N frame video image standard deviation sigma (i, j), that is:

The main background modeling b of step 1 wherein revised main background model in the step, specifically implemented according to following steps:

To the sequence of video images of N+1 frame, calculate revised main background model B by following to the 2N frame _Main=[b _Main(i, j)] _{M * n}, that is:

Wherein, threshold value th1 (i, j)=2 σ (i, j), σ (i, j) standard deviation of the preceding N frame video image that calculates for step 1a step.

The main background modeling c of step 1 wherein upgrades main background model threshold value in the step, specifically implements according to following steps:

Make σ ^Old(i, j)=th ₁(i, j)/2,

Wherein, i=1,2 ..., m, j=1,2 ..., n, k=N+1, N+2 ..., 2N,

Th ₁(i, j)=2 σ (i, j), i=1,2 ..., m, j=1,2 ..., n, α are turnover rate.

Step 1 auxiliary background modeling a wherein sets up alternate auxiliary background sequence in the step, specifically implements according to following steps:

In the sequence of video images of 3N frame, threshold series th is set at the 2N+1 frame ₂(i, j)=th ₁(i, j)+σ (i, j), th ₃(i, j)=th ₂(i, j)+σ (i, j) ..., th _K+1(i, j)=th _k(i, j)+σ (i, j), k is a positive integer,

If | f _k(i, j)-b _Main(i, j) |＞th ₁(i, j), and | f _k(i, j)-b _Main(i, j) |≤th ₂(i, j), then this frame (i, j) pixel value on is classified as C1 (i, j) class;

In like manner, | f _k(i, j)-b _Main(i, j) |＞th _m(i, j), and | f _k(i, j)-b _Main(i, j) |≤th _M+1(i, j), then this frame is in that (i, j) pixel value on is classified as C _m(i, j) class.

Step 1 auxiliary background modeling b wherein added up grouped data in the step, specifically implemented according to following steps:

Add up each classification C _k(i, j), k=1,2 ..., the number of pixels among the m, note is done

If the auxiliary background that is provided with is L, then from m classification, select the maximum L class of number of pixels as auxiliary background, not selected classification is then deleted, and calculates the average of this L classification respectively according to following formula:

Afterwards, add up the pixel distribution standard deviation in the selected classification, that is:

Step 1 auxiliary background modeling c wherein upgrades the threshold value of auxiliary background model in the step, specifically implement according to following steps:

According to the pixel distribution standard deviation that step 1 auxiliary background modeling b obtained in the step, upgrade the judgment threshold of auxiliary background according to following formula:

th _k(i，j)＝2σ _k(i，j)i＝1，2，...，m，j＝1，2，...，n，k＝1，2，...，L。

The calculating of step 2 wherein background to be updated, specifically implement according to following steps:

When background is when slowly changing, according to following formula:

Wherein, bs (i j) is a main background and L auxiliary background, ths (i j) is the judgment threshold of a main background and L auxiliary background,

Simultaneously, the pixel distribution number of background model is added up, that is:

Higher limit N is set _{Lim it}, work as N _s(i, j) 〉=N _{Lim it}The time, force N _s(i, j)=N _{Lim it},

At first, upgrade the standard deviation of each background model according to following formula:

Afterwards, upgrade the judgment threshold of background:

th _s(i，j)＝2σ _s(i，j)，

Wherein, σ _s(i j) is the standard deviation of a main background and L auxiliary background model;

When saltus step takes place in the part point in the background, if background to be updated and the main background b that is provided with before _Main(i, j), and the difference between L auxiliary background all surpassed setting threshold, is judged as the foreground point, its value is copied to background to be updated, that is:

b _renew(i，j)＝f _k(i，j)，

Simultaneously, threshold value th is set _New(i, j)=min{th ₁(i, j), th _s(i, j) (s=1,2 ..., L) }, th ₁(i j) is the judgment threshold of main background, th _s(i j) is the judgment threshold of s auxiliary background, to this background to be updated after the P frame in add up, if in the P frame Q frame is arranged, Q is no less than the lower limit that sets in advance, and satisfies

|b _renew(i，j)-f _j(i，j)|≤th _new，j＝k+1，...，P

Then be judged as and belong to background to be updated, be i.e. the number of pixels N of background to be updated _Renew(i j) is:

N _renew(i，j)＝Q，

Then the average with this Q frame deposits in the update background module, upgrades the judgment threshold of model to be updated, otherwise, this point is judged as prospect, from background to be updated, delete, wait for rejudging of next time, if N _Renew(i, j) greater than the pixel distribution number of some auxiliary background in L the auxiliary background, then background to be updated is replaced less one of statistical value in L the auxiliary background, as new auxiliary background, the auxiliary background deletion that replaces, also delete simultaneously background to be updated, wait for rejudging of next time.

The invention has the beneficial effects as follows, in background training study and renewal process, set up a plurality of mode and carry out the background description, and according to the current state of declaring, select suitable mode to carry out background and describe, and when detecting moving target, set a plurality of mode are upgraded and replacement, thus, improve when moving object detection adaptation to greatest extent to photoenvironment.

Embodiment

The present invention is described in detail below in conjunction with embodiment.

The background modeling method that multimodality automatic updating of the present invention is replaced, specifically implement according to following steps:

Step 1: the modeling of main background and auxiliary background

Main background modeling is specifically implemented according to following steps:

A. the initialization of main background model

During system start-up, at first enter the background learning phase, get continuous N frame video and carry out background study, the size of N depends on the movement velocity of moving target in this N frame, requires in the sequence of frames of video, and each pixel not frame number of passive movement target occlusion accounts for 98% of totalframes, at this moment, just need not to reject specially moving target, only need to adopt single Gaussian processes, can obtain the initial value B of main background _Main=[b _Main(i, j)] _{M * n}(m, n are respectively the line number and the columns of video frame images), that is:

b _main(i，j)＝μ(i，j)，i＝1，2，...，m，j＝1，2，...，n (1)

Wherein, μ (i, j) be N frame video frame images point (i, the j) average on, promptly $\mathrm{\μ}(i,j)=\frac{1}{N}\underset{k=1}{\overset{N}{\mathrm{\Σ}}}{f}_k(i,j),$ f _k(i j) is the k frame video image.

And add up this N frame video image standard deviation sigma (i, j), that is:

B. main background model is revised

To the sequence of video images of N+1 frame, by the following revised main background model B that calculates to the 2N frame _Main=[b _Main(i, j)] _{M * n}, that is:

Wherein, threshold value th ₀(i, j)=2 (i, j), (i j) is the standard deviation of the preceding N frame video image that calculated by formula (2) to σ to σ.

C. upgrade threshold value th ₀(i, j)

When utilizing formula (3) that background model is revised, also to threshold value th ₁(i j) upgrades, and can adapt to the variation of current ambient lighting to guarantee it.

That is: make σ ^Old(i, j)=th ₀(i, j)/2,

i＝1，2，...，m，j＝1，2，...，n，k＝N+1，N+2，...，2N。(4)

th ₀(i，j)＝2σ(i，j) i＝1，2，...，m，j＝1，2，...，n (5)

Wherein, α is a turnover rate, α ∈ [0,1].

The auxiliary background modeling is specifically implemented according to following steps:

The 2N+1 frame is provided with threshold series th in the sequence of video images of 3N frame ₁(i, j)=th ₀(i, j)+σ (i, j), th ₂(i, j)=th1 (i, j)+σ (i, j) ..., th _K+1(i, j)=th _k(i, j)+σ (i, j), (k is a positive integer).

A. set up alternate auxiliary background sequence

If | f _k(i, j)-b _Main(i, j) |＞th ₀(i, j), and | f _k(i, j)-b _Main(i, j) |≤th ₁(i, j), then this frame is in that (i, j) pixel value on is classified as C ₁(i, j) class.

In like manner, | f _k(i, j)-b _Main(i, j) |＞th _M-1(i, j), and | f _k(i, j)-b _Main(i, j) |≤th _m(i, j), then this frame is in that (i, j) pixel value on is classified as C _m(i, j) class.

B. grouped data is added up

Add up each classification C _k(i, j), k=1,2 ..., the number of pixels among the m, note is done

If the auxiliary background that is provided with is L, then from a top m classification, selecting the maximum L class of number of pixels (annotates: if classification number m≤L that top statistics obtains, then need not to select, all be auxiliary background, at this moment, make L=m), as auxiliary background, not selected classification is then deleted.Simultaneously, calculate the average of this L classification respectively according to formula (6):

Afterwards, add up the pixel distribution standard deviation in the selected classification, that is:

C. upgrade the threshold value of auxiliary background model

According to the standard deviation that formula (7) calculates, upgrade the judgment threshold of two auxiliary background respectively, that is:

th _k(i，j)＝2σ _k(i，j)i＝1，2，...，m，j＝1，2，...，n，k＝1，2，...，L?(8)

Step 2: the calculating of background to be updated

Finished after the modeling of main background and auxiliary background, just the threshold parameter substitution formula that obtains can have been carried out target detection

Wherein, F _k=[f _k(i, j)] _{M * n}Be the present frame of monitoring video, B=[b (i, j)] _{M * n}Be background model, O _k=[o _k(i, j)] _{M * n}Be the target detection result of present frame, th is the background judgment threshold, th ∈ { th ₀, th ₁..., th _L.

Detect moving target by the background subtraction point-score.In the process that detects, background is constantly upgraded to adapt to the variation of photoenvironment.When the situation difference of change of background, then the update method of this patent employing is also different.

(1) background is a situation about slowly changing

This situation is meant that the illumination condition of environment does not have the situation of acute variation, and then context update calculates according to following formula.

Wherein, b _s(i j) is a main background and L auxiliary background, th _s(i j) is the judgment threshold of a main background and L auxiliary background.

Simultaneously, the pixel distribution number of the background model under it is added up, that is:

In order to prevent to monitor for a long time, accumulative frequency N _s(safeguard measure of Cai Yonging is here, and higher limit N is set for i, j) excessive overflowing of causing _{Lim it}, work as N _s(i, j) 〉=N _{Lim it}The time, force N _s(i, j)=N _{Lim it}

Afterwards, carry out next context update threshold value constantly.

At first, upgrade the standard deviation of each background model according to following formula:

Afterwards, upgrade the judgment threshold of background:

th _s(i，j)＝2σ _s(i，j) (13)

Wherein, σ _s(i j) is the standard deviation of a main background and L auxiliary background model.

(2) situation of saltus step takes place in the part point in the background

In the background because the point of permanent saltus step takes place in a variety of causes such as wind leaf generally is fewer and is discrete relatively, this variation begins that it(?) may not can have influence on the extraction of moving target, if but do not upgrade for a long time, may will take care of the pence even can have influence on detection to target, cause a lot of flase drops.The real-time renewal that this just requires the point to saltus step in this background also will carry out.

The way of this patent is, if differentiate itself and the main background b that is provided with before by criterion _Main(i, j), and the difference between L auxiliary background all surpassed setting threshold, is judged as the foreground point, simultaneously, might be the background of saltus step in order to prevent it, its value is copied to background to be updated, that is:

b _renew(i，j)＝f _k(i，j) (14)

Simultaneously, threshold value th is set _New(i, j)=min{th ₀(i, j), th _s(i, j) (s=1,2 ..., L) } (th ₀(i j) is the judgment threshold of main background, th _s(i j) is the judgment threshold of s auxiliary background), to this background to be updated after the P frame in add up, if having Q frame (Q is no less than the lower limit that sets in advance, and lower limit is an empirical value) to satisfy in the P frame

|b _renew(i，j)-f _j(i，j)|≤th _new，j＝k+1，...，P (15)

Then be judged as and belong to background to be updated, be i.e. the number of pixels N of background to be updated _Renew(i j) is:

N _renew(i，j)＝Q (16)

Then the average (calculating according to formula (6)) with this Q frame deposits in the update background module, and according to formula (7), (8) upgrade the judgment threshold of model to be updated.Otherwise, this point is judged as prospect, from background to be updated, delete, wait for rejudging of next time.

If N _Renew(i, j) greater than the pixel distribution number of some auxiliary background in L the auxiliary background (result who calculates by formula (11)), then background to be updated is replaced less one of statistical value in L the auxiliary background, as new auxiliary background, the auxiliary background deletion that replaces, also delete simultaneously background to be updated, wait for rejudging of next time.

Claims (8)

1. the background modeling method that multimodality automatic updating is replaced is characterized in that, specifically implements according to following steps:

Step 1: the modeling of main background and auxiliary background:

Main background modeling is specifically implemented according to following steps:

A. the initialization of main background model;

B. main background model is revised;

C. upgrade main background model threshold value;

The auxiliary background modeling is specifically implemented according to following steps:

A. set up alternate auxiliary background sequence;

B. grouped data is added up;

C. upgrade the threshold value of auxiliary background model;

Step 2: the calculating of background to be updated:

According to following formula, carry out context update:

F _k=[f _k(i, j)] _{M * n}Be the present frame of monitoring video, B=[b (i, j)] _{M * n}Be background model, O _k=[o _k(i, j)] _{M * n}Be the target detection result of present frame, th is the background judgment threshold.

2. the background modeling method that multimodality automatic updating according to claim 1 is replaced is characterized in that, the initialization of the main background modeling a of described step 1 main background model in the step is specifically implemented according to following steps:

Get continuous N frame video and carry out background study, the size of N satisfies in the sequence of frames of video, and each pixel not frame number of passive movement target occlusion accounts for 98% of totalframes, adopts single Gaussian processes, obtains the initial value B of main background _Main=[b _Main(i, j)] _{M * n}, m, n are respectively the line number and the columns of video frame images, b _Main(i, j)=μ (i, j), i=1,2 ..., m, j=1,2 ..., n, wherein, μ (i, j) be N frame video frame images point (i, the j) average on, promptly $\mathrm{\μ}(i,j)=\frac{1}{N}\underset{k=1}{\overset{N}{\mathrm{\Σ}}}{f}_k(i,j),$ f _k(i j) is the k frame video image, and add up this N frame video image standard deviation sigma (i, j), that is:

$\mathrm{\σ}(i,j)=\frac{1}{N}\sqrt{\underset{k=1}{\overset{N}{\mathrm{\Σ}}}{(f_k(i,j)-\mathrm{\μ}(i,j))}^2}$ i＝1，2，…，m，j＝1，2，…，n。

3. the background modeling method that multimodality automatic updating according to claim 1 is replaced is characterized in that, the main background modeling b of described step 1 revised main background model in the step, specifically implemented according to following steps:

To the sequence of video images of N+1 frame, calculate revised main background model B by following to the 2N frame _Main=[b _Main(i, j)] _{M * n}, that is:

$b_{\mathrm{main}}(i,j)=\left\{\begin{array}{cc}{b}_{\mathrm{main}}(i,j)& \mathrm{if}|b_{\mathrm{main}}(i,j)-f_k(i,j)|\&GreaterEqual;{\mathrm{th}}_1(i,j)\\ (b_{\mathrm{main}}(i,j)+f_k(i,j))/2& \mathrm{if}|b_{\mathrm{main}}(i,j)-f_k(i,j)|<{\mathrm{th}}_1(i,j)\end{array}\right.,$ k＝N+1，N+2，…，2N，

Wherein, threshold value th ₁(i, j)=2 σ (i, j), σ (i, j) standard deviation of the preceding N frame video image that calculates for step 1a step.

4. the background modeling method that multimodality automatic updating according to claim 1 is replaced is characterized in that, the main background modeling c of described step 1 upgrades main background model threshold value in the step, specifically implements according to following steps:

Make σ ^Old(i, j)=th ₁(i, j)/2,

$\mathrm{\&sigma;}(i,j)=\left\{\begin{array}{cc}{\mathrm{\&sigma;}}^{\mathrm{old}}(i,j)& \mathrm{if}|b_{\mathrm{main}}(i,j)-f_k(i,j)|\&GreaterEqual;{\mathrm{th}}_1(i,j)\\ \sqrt{\mathrm{\&alpha;}\&CenterDot;{\mathrm{\&sigma;}}^{\mathrm{old}}{(i,j)}^2+(1-\mathrm{\&alpha;})\&CenterDot;{(f_k(i,j)-b_{\mathrm{main}}(i,j))}^2}& \mathrm{if}|b_{\mathrm{main}}(i,j)-f_k(i,j)|<{\mathrm{th}}_1(i,j)\end{array}\right.,$

Wherein, i=1,2 ..., m, j=1,2 ..., n, k=N+1, N+2 ..., 2N,

Th ₁(i, j)=2 σ (i, j), i=1,2 ..., m, j=1,2 ..., n, α are turnover rate.

5. the background modeling method that multimodality automatic updating according to claim 1 is replaced is characterized in that, described step 1 auxiliary background modeling a sets up alternate auxiliary background sequence in the step, specifically implements according to following steps:

In the sequence of video images of 3N frame, threshold series th is set at the 2N+1 frame ₂(i, j)=th ₁(i, j)+σ (i, j), th ₃(i, j)=th ₂(i, j)+σ (i, j) ..., th _K+1(i, j)=th _k(i, j)+σ (i, j), k is a positive integer,

If | f _k(i, j)-b _Main(i, j) |＞th ₁(i, j), and | f _k(i, j)-b _Main(i, j) |≤th ₂(i, j), then this frame is in that (i, j) pixel value on is classified as C ₁(i, j) class;

In like manner, | f _k(i, j)-b _Main(i, j) |＞th _m(i, j), and | f _k(i, j)-b _Main(i, j) |≤th _M+1(i, j), then this frame is in that (i, j) pixel value on is classified as C _m(i, j) class.

6. the background modeling method that multimodality automatic updating according to claim 1 is replaced is characterized in that, described step 1 auxiliary background modeling b added up grouped data in the step, specifically implemented according to following steps:

Add up each classification C _k(i, j), k=1,2 ..., the number of pixels among the m, note is done

K=1,2 ..., m if the auxiliary background that is provided with is L, then from m classification, selects the maximum L class of number of pixels as auxiliary background, and not selected classification is then deleted, and calculates the average of this L classification respectively according to following formula:

${\mathrm{\&mu;}}_k(i,j)=\frac{1}{N_{C_k}}\underset{f_k(i,j)\&Element;C_k}{\mathrm{\&Sigma;}}{f}_k(i,j)$ i＝1，2，…，m，j＝1，2，…，n，k＝1，2，…，L，

Afterwards, add up the pixel distribution standard deviation in the selected classification, that is:

${\mathrm{\&sigma;}}_k(i,j)=\frac{1}{N}\sqrt{\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}{(f_k(i,j)-{\mathrm{\&mu;}}_k(i,j))}^2}$ i＝1，2，…，m，j＝1，2，…，n，k＝1，2，…，L。

7. the background modeling method that multimodality automatic updating according to claim 1 is replaced is characterized in that, described step 1 auxiliary background modeling c upgrades the threshold value of auxiliary background model in the step, specifically implement according to following steps:

According to the pixel distribution standard deviation that step 1 auxiliary background modeling b obtained in the step, upgrade the judgment threshold of auxiliary background according to following formula:

th _k(i，j)＝2σ _k(i，j)i＝1，2，…，m，j＝1，2，…，n，k＝1，2，…，L。

8. the background modeling method that multimodality automatic updating according to claim 1 is replaced is characterized in that the calculating of described step 2 background to be updated is specifically implemented according to following steps:

When background is when slowly changing, according to following formula:

$b_s(i,j)=\left\{\begin{array}{cc}{b}_s(i,j)& \mathrm{if}|b_s(i,j)-f_k(i,j)|\&GreaterEqual;{\mathrm{th}}_s(i,j)\\ (b_s(i,j)+f_k(i,j))/2& \mathrm{if}|b_s(i,j)-f_k(i,j)|<{\mathrm{th}}_s(i,j)\end{array}\right.,$

Wherein, b _s(i j) is a main background and L auxiliary background, th _s(i j) is the judgment threshold of a main background and L auxiliary background,

Simultaneously, the pixel distribution number of background model is added up, that is:

$N_s(i,j)=\left\{\begin{array}{cc}{N}_s(i,j)& \mathrm{if}|b_s(i,j)-f_k(i,j)|\&GreaterEqual;{\mathrm{th}}_s(i,j)\\ N_s(i,j)+1& \mathrm{if}|b_s(i,j)-f_k(i,j)|<{\mathrm{th}}_s(i,j)\end{array}\right.,$

Higher limit N is set _{Lim it}, work as N _s(i, j) 〉=N _{Lim it}The time, force N _s(i, j)=N _{Lim it}, at first, upgrade the standard deviation of each background model according to following formula:

${\mathrm{\&sigma;}}_s(i,j)=\left\{\begin{array}{cc}{\mathrm{\&sigma;}}_s(i,j)& \mathrm{if}|b_s(i,j)-f_k(i,j)|\&GreaterEqual;{\mathrm{th}}_s(i,j)\\ \sqrt{0.9\&CenterDot;{{\mathrm{\&sigma;}}^2}_s(i,j)+0.1\&CenterDot;{(f_k(i,j)-b_s(i,j))}^2}& \mathrm{if}|b_s(i,j)-f_k(i,j)|<{\mathrm{th}}_s(i,j)\end{array}\right.,$

Afterwards, upgrade the judgment threshold of background:

th _s(i，j)＝2σ _s(i，j)，

Wherein, σ _s(i j) is the standard deviation of a main background and L auxiliary background model;

When saltus step takes place in the part point in the background, if background to be updated and the main background b that is provided with before _Main(i, j), and the difference between L auxiliary background all surpassed setting threshold, is judged as the foreground point, its value is copied to background to be updated, that is:

b _renew(i，j)＝f _k(i，j)，

Simultaneously, threshold value th is set _New(i, j)=min{th ₁(i, j), th _s(i, j) (s=1,2 ..., L) }, th ₁(i j) is the judgment threshold of main background, th _s(i j) is the judgment threshold of s auxiliary background, to this background to be updated after the P frame in add up, if in the P frame Q frame is arranged, Q is no less than the lower limit that sets in advance, and satisfies

|b _renew(i，j)-f _j(i，j)|≤th _new，j＝k+1，…，P

Then be judged as and belong to background to be updated, be i.e. the number of pixels N of background to be updated _Renew(i j) is:

N _renew(i，j)＝Q，

Then the average with this Q frame deposits in the update background module, upgrades the judgment threshold of model to be updated, otherwise, this point is judged as prospect, from background to be updated, delete, wait for rejudging of next time, if N _Renew(i, j) greater than the pixel distribution number of some auxiliary background in L the auxiliary background, then background to be updated is replaced less one of statistical value in L the auxiliary background, as new auxiliary background, the auxiliary background deletion that replaces, also delete simultaneously background to be updated, wait for rejudging of next time.