CN103700116A - Background modeling method for movement target detection - Google Patents

Abstract

The invention relates to a background modeling method for movement target detection, and belongs to the technical field of photoelectric image processing. In order to establish an accurate background model in a complex scene, the method provides background statistical modeling methods utilizing inter-frame differential information, namely a median method and an average method, based on the problems that a background model is difficult to obtain for the background differential method and the background differential method is easily affected by external interference, the influence of a foreground target is removed, so that the established background model is more accurate and has high anti-interference performance. The median method and the average method have simple principles and high anti-noise performance and are easy to implement. In addition, according to the method, adjacent frames of video images are differenced, and an appropriate threshold value is selected, so that an accurate background model is guaranteed.

Description

A kind of background modeling method for moving object detection

Technical field

The present invention relates to opto-electronic image processing technical field, be specifically related to a kind of background modeling method for moving object detection.

Background technology

Background subtraction point-score is a kind of method the most frequently used in current moving object detection.It can provide characteristic information more completely, and complexity is low, real-time is better, but is generally difficult in actual applications obtain in advance complete background image; And the variation of background subtraction point-score to dynamic scene, as illumination gradual change, external environment disturbance, noise etc. disturb responsive especially.Therefore how according to scene changes, setting up background model accurately, realize the real-time update to background simultaneously, is the basis that follow-up whole motion analysis and behavior are understood.

In conjunction with above-mentioned background content, how to reject the impact of foreground image, make the background model image of foundation more accurate, be the direction that current those skilled in the art make great efforts exploitation.

Summary of the invention

(1) technical matters that will solve

The technical problem to be solved in the present invention is how under complex scene, to set up background model exactly.

(2) technical scheme

For solving the problems of the technologies described above, the invention provides a kind of background modeling method for moving object detection, the method comprises:

Step S1: read in continuous M two field picture, guarantee that the position of moving target in the first frame and last frame image is not overlapping;

Step S2: two two field pictures to some frames of being separated by subtract each other acquisition difference image;

Step S3: default dynamic partition threshold value, utilize described dynamic partition threshold value to carry out binaryzation operation to described difference image and obtain differentiated binary image, to distinguish background area and the target area in described difference image;

Step S4: adopt morphologic filtering method and regional connectivity method to remove noise and fill medium and small cavity, target area;

Step S5: reject target area, calculate intermediate value or the average of residue each pixel of background area gray scale on time shaft, obtain complete background image.

Wherein, in described step S1, M is greater than 10 integer.

Wherein, in described step S2, described in be separated by two two field pictures of some frames be two two field pictures of multiframe of being separated by.

Wherein, suppose I(x, y, m) be that m two field picture is positioned at the pixel value that coordinate (x, y) is located, D(x, y, m) be differentiated binary image, B(x, y) be the background image that will form; In described step S2, described in two two field pictures that two two field pictures of some frames are direct neighbor of being separated by;

According to formula (1), calculate the difference image of consecutive frame image:

A(x,y,m)＝|I(x,y,m+1)-I(x,y,m)|,1≤m＜M （1）。

Wherein, in described step S3, described dynamic partition threshold hypothesis is T, and it adopts the gray-scale value fractile of difference image; And carry out binaryzation difference image according to formula (2):

$D(x,y,m)=\left\{\begin{array}{c}1,A(x,y,m)\&GreaterEqual;T\\ 0,A(x,y,m)<T\end{array}\right.---\left(2\right).$

Wherein, in described step S3, described dynamic partition threshold value T adopts 95% fractile of difference image gray-scale value fractile.

Wherein, in described step S5, by calculating, remain each pixel of background area intermediate value of gray scale on time shaft and obtain the process of complete background image and be:

Step S501: first to each pixel (x, y) on binary sequence image, get indexed set and represent that difference is not less than the place frame of threshold value T; Described indexed set is obtained according to formula (3):

S _x,y(m)＝{m:D(x,y,m)＝1,1≤m＜M} （3）；

And, respectively by f _x,yand e _x,ybe designated as pixel D(x, y, m) be the label of 1 o'clock first with last; Described f _x,yand e _x,yaccording to formula (4), obtain:

f _x,y＝min{m:m∈S _x,y(m)}

e _x,y＝max{m:m∈S _x,y(m)} （4）；

Step S502: by indexed set S _x,y(m) and in conjunction with formula (5) can obtain background image:

In formula, median () represents to get intermediate value.

Wherein, in described step S5, by calculating, remain each pixel of background area average of gray scale on time shaft and obtain the process of complete background image and be:

Step S501 ': first binary image is done and corroded and expansion process, remove little noise agglomerate, larger binaryzation agglomerate is done to minimum boundary rectangle computing, and the weight w of the pixel in rectangle is got to 0, the pixel weight w outside rectangle gets 1; Wherein, the weight matrix of m frame is designated as W (x, y, m);

Step S502 ': to each pixel (x, y) on binary sequence image, get indexed set T according to formula (6) _x,y(m):

T _x,y(m)＝{m:W(x,y,m)＝1,1＜m≤M} （6）；

Step S503 ': by indexed set T _x,y(m) and in conjunction with formula (7) can obtain background image B(x, y):

In formula, mean () represents to average.

Wherein, if scene changes is relatively very fast, the method also comprises step 6: background is carried out to real-time update.

Wherein, described step 6 comprises:

According to formula (8), background is carried out to real-time update:

$B(x,y,m)=\left\{\begin{array}{cc}B(x,y,m-1),& D(x,y,m)=1\\ \mathrm{\&alpha;I}(x,y,m)+(1-\mathrm{\&alpha;})B(x,y,m-1),& D(x,y,m)=0\end{array}---\left(8\right)\right.$

In formula, α is for upgrading the factor, D(x, y, m) be:

$D(x,y,m)=\left\{\begin{array}{c}1,|I(x,y,m)-I(x,y,m-1)|\&GreaterEqual;T\\ 0,|I(x,y,m)-I(x,y,m-1)|<T\end{array}\right.---\left(9\right)$

Wherein, threshold value T chooses by fractile threshold method.

(3) beneficial effect

Compared with prior art, technical solution of the present invention possesses following some beneficial effect:

(1) it rejects the impact of foreground image by median method and averaging method, makes the background model image of foundation more accurate, and two Method And Principles are simple, are easy to realize, and have good noise immunity;

(2) by consecutive frame video image difference, choose suitable threshold value, utilize the method that image is processed to reject foreground target impact, thereby obtain background model accurately.

(3) for background subtraction point-score, be difficult for background extraction model, and be easily subject to the impact of external disturbance, the Background statistic modeling method of utilizing inter-frame difference information that the present invention proposes, can guarantee under complex background, to obtain background image accurately, and have good anti-interference.

Accompanying drawing explanation

The process flow diagram of the background modeling method for moving object detection that Fig. 1 provides for technical solution of the present invention.

Fig. 2 (a) and Fig. 2 (c) are respectively the difference image of adjacent two frames of applying in the invention process process.

Fig. 2 (b) and Fig. 2 (d) are respectively the grey level histogram of statistical difference partial image Fig. 2 (a) and Fig. 2 (c).

Fig. 3 (a)-Fig. 3 (h) is for carrying out the design sketch of the background image after modeling by technical solution of the present invention.

Fig. 3 (i) is median method background modeling effect.

Fig. 3 (j) is averaging method background modeling effect.

Fig. 4 (a) and Fig. 4 (b) are respectively the front and back frame as Fig. 2 (a) source.

Fig. 4 (c) and Fig. 4 (d) are respectively the front and back frame as Fig. 2 (c) source.

Embodiment

For making object of the present invention, content and advantage clearer, below in conjunction with drawings and Examples, the specific embodiment of the present invention is described in further detail.

In order to set up background model exactly under complex scene, the present invention is directed to background subtraction point-score and be difficult for background extraction model, and be easily subject to the impact of external disturbance, two kinds of Background statistic modeling methods of utilizing inter-frame difference information---median method and averaging method have been proposed, remove the impact of foreground target, made the background model of foundation more accurate.Based on above-mentioned two kinds of methods, the invention provides a kind of background modeling method for moving object detection, as shown in Figure 1, the method comprises:

Step S1: read in continuous M two field picture, guarantee that the position of moving target in the first frame and last frame image is not overlapping; Wherein, M is greater than 10 integer;

Step S2: two two field pictures to some frames of being separated by subtract each other acquisition difference image; Wherein, when target speed compares speed, two two field pictures that two two field pictures of some frames of being separated by described in can choosing are direct neighbor; And when target speed is slow, the number of image frames M that need to read in also can increase, now when step S1 makes calculus of differences, can not adopt the frame of adjacent two frames poor, but the n frame of being separated by is made first difference, can effectively reduce operand, now, described in be separated by two two field pictures of some frames be two two field pictures of multiframe of being separated by.

Step S3: default dynamic partition threshold value, utilize described dynamic partition threshold value to carry out binaryzation operation to described difference image and obtain differentiated binary image, to distinguish background area and the target area in described difference image; Wherein, if adopt static threshold, the method is not strong to the adaptability of the complex environments such as light variation, noise, and the present invention adopts the gray-scale value fractile of difference image for this reason, preferably can adopt 95% fractile to be used as dynamic partition threshold value; The grey level histogram of statistical difference partial image, as Fig. 2 shows, Fig. 2 (a) and Fig. 2 (c) are respectively the difference image of rigid body two frames adjacent with non-rigid motion; The primitive frame that Fig. 2 (a) originates is respectively Fig. 4 (a) and Fig. 4 (b), wherein, first Fig. 4 (a) and Fig. 4 (b) is carried out to black and white processing, then Fig. 4 of black and white (b) pattern is deducted to Fig. 4 (a), thereby obtains Fig. 2 (a); The primitive frame that Fig. 2 (c) originates is respectively Fig. 4 (c) and Fig. 4 (d), wherein, first Fig. 4 (c) and Fig. 4 (d) is carried out to black and white processing, then Fig. 4 of black and white (d) pattern is deducted to Fig. 4 (c), thereby obtains Fig. 2 (c); Fig. 2 (b) is respectively its corresponding difference image grey level histogram with Fig. 2 (d).Can find out, the most value of gray scale of difference image concentrates near 0, and this mainly causes due to external environment condition interference, and only has the value of minority to be significantly greater than 0, can think the result that target travel causes; Therefore, suitable fractile threshold value can be according to actual conditions self-adaptation value, thereby effectively reduces the probability of miscarriage of justice of classification;

Step S4: adopt morphologic filtering method and regional connectivity method to remove noise and fill medium and small cavity, target area; This step is conventional at present technological means, does not repeat them here;

Step S5: reject target area, calculate intermediate value or the average of residue each pixel of background area gray scale on time shaft, obtain complete background image.

(1), for median method in the situation that, specifically details are as follows for above-mentioned steps:

First, suppose I(x, y, m) be that m two field picture is positioned at the pixel value that coordinate (x, y) is located, D(x, y, m) be differentiated binary image, B(x, y) be the background image that will form; In described step S2, described in two two field pictures that two two field pictures of some frames are direct neighbor of being separated by;

According to formula (1), calculate the difference image of consecutive frame image:

A(x,y,m)＝|I(x,y,m+1)-I(x,y,m)|,1≤m＜M （1）

In described step S3, described dynamic partition threshold hypothesis is T, and it adopts the gray-scale value fractile of difference image; And carry out binaryzation difference image according to formula (2):

$D(x,y,m)=\left\{\begin{array}{c}1,A(x,y,m)\&GreaterEqual;\mathrm{<figure-callout\; id="0"\; label="T"\; filenames="DEST\_PATH\_HDA00002447408400022.png,DEST\_PATH\_HDA00002447408400024.png"\; state="\{\{state\}\}">T</figure-callout>}\\ 0,A(x,y,m)<T\end{array}\right.---\left(2\right).$

In described step S3, described dynamic partition threshold value T adopts 95% fractile of difference image gray-scale value fractile.

In described step S5, by calculating, remain each pixel of background area intermediate value of gray scale on time shaft and obtain the process of complete background image and be:

Step S501: first to each pixel (x, y) on binary sequence image, get indexed set and represent that difference is not less than the place frame of threshold value T; Described indexed set is obtained according to formula (3):

S _x,y(m)＝{m:D(x,y,m)＝1,1≤m＜M} （3）；

And, respectively by f _x,yand e _x,ybe designated as pixel D(x, y, m) be the label of 1 o'clock first with last; Described f _x,yand e _{x, y}according to formula (4), obtain:

f _x,y＝min{m:m∈S _x,y(m)}

e _x,y＝max{m:m∈S _x,y(m)} （4）；

Step S502: by indexed set S _x,y(m) and in conjunction with formula (5) can obtain background image:

In formula, median () represents to get intermediate value.

(2), for median method in the situation that, specifically details are as follows for above-mentioned steps:

It is a kind of comprehensive of multi-frame mean method and Surendra update method that the method can be considered as, and wherein, step S1-step S4 is identical with above-mentioned median method: for step S5,

By calculating, remaining each pixel of background area average of gray scale on time shaft obtains the process of complete background image and is:

Step S501 ': first binary image is done and corroded and expansion process, remove little noise agglomerate, larger binaryzation agglomerate is done to minimum boundary rectangle computing, and the weight w of the pixel in rectangle is got to 0, the pixel weight w outside rectangle gets 1; Wherein, the weight matrix of m frame is designated as W (x, y, m);

Step S502 ': to each pixel (x, y) on binary sequence image, get indexed set T according to formula (6) _x,y(m):

T _x,y(m)＝{m:W(x,y,m)＝1,1＜m≤M} （6）；

Step S503 ': by indexed set T _x,y(m) and in conjunction with formula (7) can obtain background image B(x, y):

In formula, mean () represents to average.

In sum, according to median method and averaging method, reject the impact of foreground image, make the background model image of foundation more accurate.Two Method And Principles are simple, are easy to realize, and have good noise immunity.

In addition,, aspect the renewal of background model, if scene changes is slow, can take above-described median method or averaging method periodically to upgrade background; If scene changes is relatively very fast, the method also comprises step 6: background is carried out to real-time update;

Described step 6 comprises:

According to formula (8), background is carried out to real-time update:

$B(x,y,m)=\left\{\begin{array}{cc}B(x,y,m-1),& D(x,y,m)=1\\ \mathrm{\&alpha;I}(x,y,m)+(1-\mathrm{\&alpha;})B(x,y,m-1),& D(x,y,m)=0\end{array}---\left(8\right)\right.$

In formula, α is for upgrading the factor, D(x, y, m) be:

$D(x,y,m)=\left\{\begin{array}{c}1,|I(x,y,m)-I(x,y,m-1)|\&GreaterEqual;\mathrm{<figure-callout\; id="0"\; label="T"\; filenames="DEST\_PATH\_HDA00002447408400022.png,DEST\_PATH\_HDA00002447408400024.png"\; state="\{\{state\}\}">T</figure-callout>}\\ 0,|I(x,y,m)-I(x,y,m-1)|<T\end{array}\right.---\left(9\right)$

Wherein, threshold value T chooses by fractile threshold method.

For the validity of checking technical solution of the present invention, for rigid motion and non-rigid motion, utilize technical solution of the present invention and traditional Surendra algorithm to carry out the contrast experiment of background modeling, effect contrast figure is as shown in Fig. 3 (a)-Fig. 3 (j).Wherein, Fig. 3 (a)-Fig. 3 (e) is for rigid motion situation, the 1114th frame that Fig. 3 (a) is sequence image, Fig. 3 (b) is the 1149th frame, Fig. 3 (c) is traditional Surendra algorithm background modeling effect, Fig. 3 (d) is median method background modeling effect, and Fig. 3 (e) is averaging method background modeling effect.As can be seen from Figure, for the situation of rigid motion, three kinds of methods all have good effect, but the background that the median method based on statistics and averaging method are recovered is cleaner than Surendra algorithm.This be because, median method and averaging method are that the overall situation of the picture frame of all participation background modelings is considered, the moving region that any frame detects, reject by reasonable manner in capital, and can not affect obtaining of background image in this region, even if inter-frame difference is not accurately cut apart moving region, also can be by getting the tactful correction obtaining to a certain extent of intermediate value or average; And the context update of Surendra algorithm must depend on the testing result of the moving region of current frame image, if present frame has target originally, but inter-frame difference does not detect, will produce larger adverse effect to the context update in this target area, this limitation of Surendra algorithm shows more obvious in little target.

Fig. 3 (f)-Fig. 3 (J) is for non-rigid motion situation, the 1st frame that Fig. 3 (f) is sequence image, Fig. 3 (g) is the 50th frame, Fig. 3 (h) is traditional Surendra algorithm background modeling effect, Fig. 3 is (i) median method background modeling effect, and Fig. 3 (j) is averaging method background modeling effect.As can be seen from Figure, for the situation of non-rigid motion, the successful of Surendra algorithm is not as other two kinds of methods.This be because, Surendra algorithm must rely on the moving region testing result of next frame image background is upgraded, when in image, the regional area of moving target keeps motionless within a certain period of time, this region will be mistaken as background, thereby causes context update to be made mistakes; The equilibrium based on all picture frames of median method and averaging method is considered can reduce this impact as far as possible, makes background image more accurate.

The above is only the preferred embodiment of the present invention; it should be pointed out that for those skilled in the art, do not departing under the prerequisite of the technology of the present invention principle; can also make some improvement and distortion, these improvement and distortion also should be considered as protection scope of the present invention.

Claims (10)

1. for a background modeling method for moving object detection, it is characterized in that, the method comprises:

Step S1: read in continuous M two field picture, guarantee that the position of moving target in the first frame and last frame image is not overlapping;

Step S2: two two field pictures to some frames of being separated by subtract each other acquisition difference image;

Step S3: default dynamic partition threshold value, utilize described dynamic partition threshold value to carry out binaryzation operation to described difference image and obtain differentiated binary image, to distinguish background area and the target area in described difference image;

Step S4: adopt morphologic filtering method and regional connectivity method to remove noise and fill medium and small cavity, target area;

Step S5: reject target area, calculate intermediate value or the average of residue each pixel of background area gray scale on time shaft, obtain complete background image.

2. the background modeling method for moving object detection as claimed in claim 1, is characterized in that, in described step S1, M is greater than 10 integer.

3. the background modeling method for moving object detection as claimed in claim 1, is characterized in that, in described step S2, described in be separated by two two field pictures of some frames be two two field pictures of multiframe of being separated by.

4. the background modeling method for moving object detection as claimed in claim 1, is characterized in that, supposes I(x, y; m) be that m two field picture is positioned at the pixel value that coordinate (x, y) is located, D(x, y; m) be differentiated binary image, B(x, y) be the background image that will form; In described step S2, described in two two field pictures that two two field pictures of some frames are direct neighbor of being separated by;

According to formula (1), calculate the difference image of consecutive frame image:

A(x,y,m)＝|I(x,y,m+1)-I(x,y,m)|,1≤m＜M （1）。

5. the background modeling method for moving object detection as claimed in claim 4, is characterized in that, in described step S3, described dynamic partition threshold hypothesis is T, and it adopts the gray-scale value fractile of difference image; And carry out binaryzation difference image according to formula (2):

$D(x,y,m)=\left\{\begin{array}{c}1,A(x,y,m)\&GreaterEqual;T\\ 0,A(x,y,m)<T\end{array}\right.---\left(2\right).$

6. the background modeling method for moving object detection as claimed in claim 5, is characterized in that, in described step S3, described dynamic partition threshold value T adopts 95% fractile of difference image gray-scale value fractile.

7. the background modeling method for moving object detection as claimed in claim 5, is characterized in that, in described step S5, remains each pixel of background area intermediate value of gray scale on time shaft obtain the process of complete background image and be by calculating:

Step S501: first to each pixel (x, y) on binary sequence image, get indexed set and represent that difference is not less than the place frame of threshold value T; Described indexed set is obtained according to formula (3):

S _x,y(m)＝{m:D(x,y,m)＝1,1≤m＜M} （3）；

And, respectively by f _x,yand e _{x, y}be designated as pixel D(x, y, m) be the label of 1 o'clock first with last; Described f _x,yand e _x,yaccording to formula (4), obtain:

f _x,y＝min{m:m∈S _x,y(m)}

e _x,y＝max{m:m∈S _x,y(m)} （4）；

Step S502: by indexed set S _x,y(m) and in conjunction with formula (5) can obtain background image:

In formula, median () represents to get intermediate value.

8. the background modeling method for moving object detection as claimed in claim 5, is characterized in that, in described step S5, remains each pixel of background area average of gray scale on time shaft obtain the process of complete background image and be by calculating:

Step S501 ': first binary image is done and corroded and expansion process, remove little noise agglomerate, larger binaryzation agglomerate is done to minimum boundary rectangle computing, and the weight w of the pixel in rectangle is got to 0, the pixel weight w outside rectangle gets 1; Wherein, the weight matrix of m frame is designated as W (x, y, m);

Step S502 ': to each pixel (x, y) on binary sequence image, get indexed set T according to formula (6) _x,y(m):

T _x,y(m)＝{m:W(x,y,m)＝1,1＜m≤M} （6）；

Step S503 ': by indexed set T _x,y(m) and in conjunction with formula (7) can obtain background image B(x, y):

In formula, mean () represents to average.

9. the background modeling method for moving object detection as claimed in claim 5, is characterized in that, if scene changes is relatively very fast, the method also comprises step 6: background is carried out to real-time update.

10. the background modeling method for moving object detection as claimed in claim 9, is characterized in that, described step 6 comprises:

According to formula (8), background is carried out to real-time update:

$B(x,y,m)=\left\{\begin{array}{cc}B(x,y,m-1),& D(x,y,m)=1\\ \mathrm{\&alpha;I}(x,y,m)+(1-\mathrm{\&alpha;})B(x,y,m-1),& D(x,y,m)=0\end{array}---\left(8\right)\right.$

In formula, α is for upgrading the factor, D(x, y, m) be:

$D(x,y,m)=\left\{\begin{array}{c}1,|I(x,y,m)-I(x,y,m-1)|\&GreaterEqual;T\\ 0,|I(x,y,m)-I(x,y,m-1)|<T\end{array}\right.---\left(9\right)$

Wherein, threshold value T chooses by fractile threshold method.

Images