CN101315701A - Movement destination image partition method - Google Patents

Abstract

The invention provides a method for segmenting a moving target image, belonging to the technical field of a computer of image processing; the method comprises the following steps of: adopting a minimum median variance method to obtain a background image; adopting an indirect difference function to obtain a difference image; step 3, selecting a segmentation threshold T and carrying out binarization for the difference image so as to obtaining a binarization foreground image; updating the ground image dynamically according to the updated image, the current background image and the binarization foreground image; using a morphological filtering to eliminate the noise in the binary image and fill defects in the moving target image; the invention can provide help for effective use of a monitoring system and reliable evaluation of monitoring effect, and can be applied in various public places such as police, fire fighting, customs, ports and stations, etc. widely, thus obtaining considerable social effect and improving public safety service.

Description

Movement destination image partition method

Technical field

The invention belongs to the pattern process computer technical field, relate to a kind of movement destination image partition method.

Background technology

Be that video monitoring image is handled the most basic link cutting apart of movement human image under the complex background, and it is intended to from monitoring gained video sequence image the movement human zone be extracted from background images.The post-processed of effectively cutting apart for monitoring objective such as classification, tracking and identification of moving region are extremely important.Yet because the dynamic change of background images, confusion such as the variation of the variation of weather, illumination condition, background is disturbed, blocking even the motion of video camera etc. between shadow, object and the environment of moving target or between object and the object, makes the detection of movement human image become the quite work of difficulty.Therefore also cause domestic and international many researchers' great interest, become the forward position focus that video image information detection range in recent years receives much concern.

According to algorithm characteristic, under the complex background in the video image movement human detection method can roughly be divided into motion segmentation and the static two class algorithms of cutting apart.The motion segmentation algorithm is a movement properties of utilizing moving target to have, and it is distinguished from the sequence image background.Static partitioning algorithm then be at first to single-frame images according to it information such as each regional gray scale, lines or profile gradient do static state and cut apart, between sequential frame image, adopt the estimation of taking exercises of similar Block Matching Algorithm afterwards, remerge the extraction that each cut zone is finished moving target at last.For stationary background, have only the video image of simple global scene motion, adopt differential motion detection algorithm or background modeling motion detection algorithm usually.The differential motion detection algorithm is that the image difference of adjacent two frames or three frames is done the moving region that thresholding extracts image.The background modeling motion detection algorithm is one of motion segmentation method of using always.Usually, this algorithm will utilize complete sequence of video images information to carry out statistical modeling to distinguish pixel class (background/moving region) and generation background image earlier; Again from every two field picture background correction to obtain moving target.In a word, the motion segmentation algorithm is numerous, and key is to search out appropriate methods, promptly realizes the target of cutting apart of expection under the situation of cost minimum.

Summary of the invention

The present invention is directed to public place fixed point monitoring gained video image and have the characteristics of stationary background and panorama motion, propose a kind of movement destination image partition method in the complex background video monitoring that can be used for.The present invention can be effective use of supervisory system and the reliable evaluation of monitoring effect is offered help, and can be widely used in various public places such as public security, fire-fighting, customs, harbour, station, obtains the lifting of considerable social benefit and public safety service.

The present invention adopts following technical scheme:

A kind of movement destination image partition method, each image segmentation adopts following step:

Step 1: adopt minimum meta variance method to obtain background image;

Step 2: adopt indirect difference function $f(a,b)=1-\frac{2\sqrt{(a+1)(b+1)}}{(a+1)+(b+1)}\frac{2\sqrt{(256-a)(256-b)}}{(256-a)+(256-b)}$ Obtain difference image, a wherein, b represent present image and background image gray scale (intensity) value in same pixel position respectively, and 0≤a, and b≤255,0≤f (a, b)≤1;

Step 3: select segmentation threshold T, $M_{(x,y)}=\left\{\begin{array}{cc}1& f(a_{(x,y)},b_{(x,y)})\&GreaterEqual;\mathrm{<figure-callout\; id="0"\; label="T"\; filenames="A20081005383000102.png"\; state="\{\{state\}\}">T</figure-callout>}\\ 0& \mathrm{Otherwise}\end{array}\right.$ With the difference image binaryzation, obtain the binaryzation foreground image;

Step 4: establish I _{(x, y)} ^tBe the sequence image that the N frame is gathered, B _{(x, y)} ^N+1And B _{(x, y)} ⁿBe respectively after the renewal that obtains according to step 1 and current background image, M _{(x, y)} ⁿFor the binaryzation foreground image that obtains according to step 3, according to formula

Dynamically update background image;

Step 5: utilize morphologic filtering to eliminate the noise in the bianry image and fill up disappearance in the movement destination image.

Above-mentioned movement destination image partition method, in the step 1, can be according to formula $B_{(x,y)}=\arg \left\{{\min }_p\mathrm{me}{d}_t{(I_{(x,y)}^t-p)}^2\right\}$ Obtain the background image of R, G, three components of B respectively, p is that (x y) locates coloured image value to be determined to location of pixels, and t is the frame index value in the formula; Can adopt maximization inter-class variance method to determine threshold value T in the step 3; In the step 4, α can get 0.8.

The used vedio data of the present invention has the characteristics of stationary background and panorama motion, adopts dynamic background modeling and differential motion to detect the algorithm that combines and cuts apart movement destination image.From each stage result of Fig. 3, Fig. 4 and Fig. 5, it is suitable that the present invention adopts the overall plan of movement human contour images dividing method in the complex background video monitoring; Wherein static background modeling, dynamic background are upgraded, differential motion is cut apart with processing links such as morphology aftertreatment all be indispensable or; Its treatment scheme is reasonable substantially, and final segmentation result is more clearly, agile.

In modern society's supervisory system, automatically the method for obtaining the monitored object data roughly can be divided into two classes: a class is to utilize sensors such as piezoelectricity, infrared, annular magnetic inductive coil to obtain the parameter of monitored object itself, this class methods Tracking Recognition rate is higher, but damage easily, install also inconvenient; Also have a class just to be based on the method (the invention belongs to this kind) of Flame Image Process and pattern-recognition, overcome the limitation of front one class methods, because the progress of Flame Image Process recognition technology and the significantly raising of hardware cost performance make that method proposed by the invention is implemented.Compare with preceding a kind of method, the obvious technical effects that the present invention produces, adaptive capacity to environment is strong, the energy long-term stable operation, and can under the unwitting situation of monitored object, monitor, make the effect of security monitoring improve greatly, and can avoid tradition (first kind) watch-dog when monitoring and monitored object generation unnecessary friction and contradiction.

Description of drawings

Fig. 1 movement human contour images is cut apart flow process.

The difference function of Fig. 2 different background gray scale.

The motion segmentation result of Fig. 3 (a) R component.

The motion segmentation result of Fig. 3 (b) G component.

The motion segmentation result of Fig. 3 (c) B component.

Fig. 3 (d) a, b, c three width of cloth figure get the inclusive-OR operation result.

Fig. 4 morphologic filtering is handled the back result.

The last moving Object Segmentation result of Fig. 5.

The dilation operation example of Fig. 6 bianry image is a width of cloth bianry image (a), and (b) figure is structural element B, indicates the reference point of "+" representative structure element, (c) expansion results figure.

The erosion operation example of Fig. 7 bianry image is that width of cloth bianry image (b) figure is structural element B (a), indicates the reference point of "+" representative structure element, (c) Corrosion results figure.

Embodiment

Below in conjunction with drawings and Examples the present invention is done detailed description.

As embodiment, the entire image cutting procedure comprises that static background modeling, dynamic background upgrade, differential motion is cut apart and the morphology post-treating and other steps, as shown in Figure 1 with the movement human contour images dividing method in the complex background video monitoring in the present invention.Respectively to each step, do describing in further detail in conjunction with the embodiments below.

1. static background modeling

The present invention tests the gait data storehouse that used data are the issue of Beijing robotization research institute of the Chinese Academy of Sciences.It is a toy data base that contains 20 objects, and all data all use separate unit video camera (Panasonic NV-DX100EN) to gather and get, the video sequence data of each object three direction walking of picked-up (0 °, 45 °, 90 °) out of doors.Only use the data of 0 ° of direction among the present invention, i.e. people's body side surface walking sequence.

The present invention adopts minimum meta variance method, and (Least median of squares LmedS) carries out modeling to background image.Minimum meta variance method (LmedS) is to be a kind of algorithm that theoretical foundation proposes with the robust statistics.Sane notion is meant the influence degree of indivedual exceptional values to statistic, and robust statistics is a kind of method for parameter estimation that has exceptional value in the sample cluster that is applicable to.Robust statistics has been subjected to paying close attention to widely in computer vision field, mainly is owing to usually disturbed by exceptional value in the input data of computer vision problem.

In robust statistics, in order to estimate the ability of certain algorithm opposing exceptional value disturbance, Hampel has proposed the notion of failpoint BP (Breakdown Point).Because originally it is an asymptotic result, inconvenience is calculated, so the BP that Donoho and Huber have defined for limited sample is:

Wherein T is an estimates of parameters, and n is the content of sample X, β (m; T, X) expression will be replaced the supremum of the difference of both sides, m some back estimates of parameters arbitrarily among the X.For the BP of limited sample, when expression is made parameter estimation with this method, the abnormal point numerical purpose minimum scale that estimated value was lost efficacy that exists in the permission data, when the abnormity point ratio in the data surpassed BP, it is very unstable that estimated value can become.

Disposal route to abnormity point has two kinds substantially, and a kind of is self-coordinating method (Accommodation-basedapproach), and promptly method itself can be born the interference of abnormity point, and another kind then is a rejecting abnormalities point at first, handles by classic method again.Minimum meta variance algorithm (LmedS) is the combination of the two, and under the self-coordinating prerequisite, (Least mean square LMS) estimates, so it can eliminate abnormity point and disturb, and can obtain estimation effect preferably simultaneously to use the least mean-square estimate algorithm again.Rousseeuw and Leroy are defined as about minimum meta variance method:

Known array X={x ₁, x ₂... x _i... x _NMiddle x _iBe N the observed reading of x, according to x _iX is estimated, suppose that estimated value is

Then:

$\widehat{\mathrm{\&theta;}}=\arg \left\{{\min }_{\mathrm{\&theta;}}{\mathrm{med}}_i{(x_i-\mathrm{\&theta;})}^2\right\}---\left(2\right)$

I=1 wherein, 2 ..., N.

Video sequence in the database used herein is that fixed cameras is taken gained, in theory, does not exist even fully and disturbs or other influences, and background is static.To the video sequence single-frame analysis, follow the tracks of wherein certain some gray scale change curve in time, this curve is held stationary substantially.Yet for moving object or other environmental interference factor origination points are arranged, its intensity profile curve can time to time change.The present invention attempts the gray scale of diverse location pixel in the gait sequence has been carried out tracing observation.

The static background modeling process is as described below.

If make I _{(x, y)} ^tThe sequence image of representing the N frame to gather, wherein t representative frame index value (t=1,2 ..., N), (x, y) ∈ I ^t, background B then _{(x, y)}For:

$B_{(x,y)}=\arg \left\{{\min }_p{\mathrm{med}}_t{(I_{(x,y)}^t-p)}^2\right\}---\left(3\right)$

In the formula p be location of pixels (x, y) locate coloured image to be determined (R, G, B) value, if each component is the image of 8 bits, then the span of p is 0～255; T is the frame index value, and it changes between 1～N.The idiographic flow of (arg represents to satisfy the value of the unknown number that requires in ()) algorithm is (being example with the R component only):

(i) a selected pixel position (x, y);

(ii)p＝0；

(iii) calculate (I successively _{(x, y)} ¹-p) ², (I _{(x, y)} ²-p) ²..., (I _{(x, y)} ^N-p) ²

(iv) to result of calculation ordering, if N is an even number, get the ordering back N/2 and (N+1)/2 mean value of number, if N is an odd number, then get the N/2 number, the result is saved among the array med, i.e. med ₀

(v) p=p+1 when p＜=255, returns (iii), repeats (iii), and (iv), (v), the result saves as med _p, otherwise carry out (vi);

(vi) find out med ₀, med ₁... med ₂₅₅Middle minimum value, corresponding p value is the background gray levels of this pixel position.

(vii) reselect the pixel position, return (ii) and repeat, all pixels all calculate and finish in image.

Consider that database images is rgb format, therefore here to R, G, the modeling respectively of three components of B, through the synthetic color background image that also can obtain R, G, B form.

1. differential motion is cut apart

In order to determine moving target, most popular method is present image and background model to be subtracted each other the difference image that obtains carry out Threshold Segmentation again.A very big deficiency of this method is to soft image, will be difficult to determine segmentation threshold because of its grey scale change is too little, promptly is difficult to moving target is fully clearly extracted from background.For this reason, the present invention uses an indirect difference function instead and carries out difference operation.Being expressed as of this difference function:

$f(a,b)=1-\frac{2\sqrt{(a+1)(b+1)}}{(a+1)+(b+1)}\frac{2\sqrt{(256-a)(256-b)}}{(256-a)+(256-b)}---\left(5\right)$

A wherein, b represent present image and background image gray scale (intensity) value in same pixel position respectively, and 0≤a, and b≤255,0≤f (a, b)≤1.When a=b, and f (a, b)=0; Work as a, b not simultaneously, (a, b) and a, the difference between the b is directly proportional f.Simultaneously, the sensitivity of difference function can change automatically with background gray levels.With b=5 and b=100 is example, and its difference function as shown in Figure 2.B difference as can be seen from Figure, and difference function f (a, b) also different, when b less (b=5), difference function is along with the growth of a becomes big rapidly, and this explanation sensitivity meeting of difference function under low-contrast circumstances improves automatically, and this adaptivity has improved the accuracy of image segmentation.

Behind the difference value of having determined background image and present image, just need to select segmentation threshold T (0≤T≤1).The present invention takes Otsu method (source of paper: Otsu N.A threshold selection method from grey-levelhistograms.In:IEEE Trans.Systems, Man and Cybernetics, 1979, SMC-9 (1), 62 ~ 66), promptly determine threshold value T by the maximization inter-class variance.Its binaryzation process can be expressed as:

$M_{(x,y)}=\left\{\begin{array}{cc}1& f(a_{(x,y)},b_{(x,y)})\&GreaterEqual;\mathrm{<figure-callout\; id="0"\; label="T"\; filenames="A20081005383000102.png"\; state="\{\{state\}\}">T</figure-callout>}\\ 0& \mathrm{Otherwise}\end{array}\right.---\left(6\right)$

2. dynamic background upgrades

Though more than set up the static background model by minimum meta variance method because the influence of noise and illumination variation, the background of actual video sequence is not to be constantly to keep static constant.Cut apart (also being background subtraction) effect in order to obtain accurate more differential motion, must dynamically upgrade background.The present invention adopts the kalman filter method of Karmann and Brandt to carry out the (source of paper: Karmann K that dynamically updates of background, Brandt A.Moving object recognitionusing an adaptive background memory.In:Cappellini V ed.Time-varying ImageProcessing and Moving Object Recognition.2.Elsevier, Amsterdam, The Netherlands, 1990).If B _{(x, y)} ^N+1And B _{(x, y)} ⁿBe respectively and upgrade back and current background, according to present frame I _{(x, y)} ⁿIn finish the binaryzation foreground image M that motion detection is obtained _{(x, y)} ⁿ, the context update process is:

Wherein α is a weighting coefficient, the present invention analyze by experiment draw α get 0.8 comparatively suitable.Before upgrading rear backdrop and upgrading between the background enough near the time, stop to upgrade.

The result who carries out motion segmentation with the background of automatic renewal as shown in Figure 3.

3. morphology aftertreatment

Can there be noise unavoidably in the image after the motion segmentation, has a small amount of point in the moving target simultaneously and be mistaken for background, therefore also need image is carried out aftertreatment, to obtain best segmentation effect.The present invention uses morphologic filtering to eliminate the noise in the bianry image and fills up disappearance in the moving target.

In morphology, dilation operation and erosion operation are the most basic morphological transformations.

1. dilation operation (Dilation)

Dilation operation also claims to expand computing, uses symbol Expression, X expands with B and is designated as X

B is defined as

Expansion process can be described below: set B at first is the Mapping B ^ about initial point, and translation x forms set (B^) x then, and last set of computations (B^) x and set X are not the set of the structural element reference point of empty set.In other words, be the set of the displacement of the B^ reference point locations of structural element B when having at least a nonzero element to intersect with expand set that X obtains of B with set X.

Example 1. dilation operation examples

As Fig. 6 (a) is a width of cloth bianry image, and dash area is represented the zone of gray-scale value for high (being generally 1), and it is the zone of low (being generally 0) that white portion is represented gray-scale value, and its upper left corner volume coordinate is (0,0).(b) figure is structural element B, indicates the reference point of "+" representative structure element.The result who expands is shown in figure (c), and wherein black is the part of the expansion of expanding.X as a result

B compares discovery with X, and X is according to the form of the B certain limit that expanded.Therefore, this computing is by the expansion that is of name.

2. erosion operation (Erosion)

Erosion operation also claims to corrode computing, uses symbol

Expression, X corrodes with B and is designated as X

B is defined as

Corrosion process can be described below: the still set of the structural element reference point in set X behind the set B translation x.In other words, corrode the set that set that X obtains is B reference point locations of B when being included among the set X fully with B.

Example 2 erosion operation examples

As Fig. 7 (a) is a width of cloth bianry image, and (b) figure is structural element B, indicates "+" and represents reference point.The result of corrosion is shown in figure (c), and wherein black is the part that stays after the corrosion.X as a result

B compares discovery with X, and the regional extent of X is reduced, and as seen, part that can not the contained structure element all has been corroded.

In morphology, opening operation A ο B is meant the result who expands with B again after A is by the B corrosion, that is:

represents erosion operation in the formula,

The expression dilation operation.Opening operation can be deleted the object that can not comprise structural element fully, as the cam contour of smooth object, disconnect narrow connection, remove tiny jut.Closed operation is just opposite with opening operation, and its definition is meant the result of corroding with B again after A is expanded by B, that is:

But the hole that closed operation packing ratio structural element is little is as the concave contour of smooth object, connect into elongated curved mouthful with long and narrow breach.Can utilize the above-mentioned character of opening operation and closed operation, realize the function of filtering and filling cavity.Fig. 4 has provided the result that morphologic filtering is handled.

After morphologic filtering is handled, noise might not be eliminated fully, and the clutter noise that has may form piece not of uniform size, and moving target is maximum in these pieces often, therefore can carry out the connected domain analysis to image, purpose is only to keep moving target in image.The step of connected domain analysis is:

(i) mark connection matrix;

(ii) calculate each matrix pixel number;

(iii) find out the maximum matrix of pixel;

(iv) determine moving target.

The moving Object Segmentation of Huo Deing as shown in Figure 5 at last.

Claims (1)

1. movement destination image partition method, each image segmentation adopts following step:

Step 1: adopt minimum meta variance method to obtain background image;

Step 2: adopt indirect difference function $f(a,b)=1-\frac{2\sqrt{(a+1)(b+1)}}{(a+1)+(b+1)}\frac{2\sqrt{(256-a)(256-b)}}{(256-a)+(256-b)}$ Obtain difference image, a wherein, b represent present image and background image gray scale (intensity) value in same pixel position respectively, and 0≤a, and b≤255,0≤f (a, b)≤1;

Step 3: select segmentation threshold T, $M_{(x,y)}=\left\{\begin{array}{cc}1& f(a_{(x,y)},b_{(x,y)})\&GreaterEqual;T\\ 0& \mathrm{Otherwise}\end{array}\right.$ With the difference image binaryzation, obtain the binaryzation foreground image;

Step 4: establish I _{(x, y)} ^tBe the sequence image that the N frame is gathered, B _{(x, y)} ^N+1And B _{(x, y) n}Be respectively after the renewal that obtains according to step 1 and current background image, M _{(x, y)} ⁿFor the binaryzation foreground image that obtains according to step 3, according to formula

Dynamically update background image;

Step 5: utilize morphologic filtering to eliminate the noise in the bianry image and fill up disappearance in the movement destination image.

Movement destination image partition method according to claim 1 is characterized in that, in the step 1, according to formula $B_{(x,y)}=\arg \left\{{\min }_p{\mathrm{med}}_t{(I_{(x,y)}^t-p)}^2\right\}$ Obtain the background image of R, G, three components of B respectively, p is that (x y) locates coloured image value to be determined to location of pixels, and t is the frame index value in the formula.

Movement destination image partition method according to claim 1 is characterized in that, adopts maximization inter-class variance method to determine threshold value T in the step 3.

Movement destination image partition method according to claim 1 is characterized in that, in the step 4, α gets 0.8.

Images