CN101437113B - Apparatus and method for detecting self-adapting inner core density estimation movement - Google Patents

Abstract

The invention provides a self-adaptive kernel density estimation motion detection device and a method thereof. The self-adaptive kernel density estimation motion detection device comprises a receiving unit, an initializing unit, a probability calculation statistic unit and a classifying unit, wherein the receiving unit is used for receiving an inputted video frame; the initializing unit is used for initializing a background model and an interframe difference background model; the probability calculation statistic unit is used for calculating a kernel density estimation probability of each pixel aiming at the pixel of the current video frame according to a background sample, and calculating a column diagram of the kernel density estimation probability of the pixel of the current frame; andthe classifying unit is used for self adaptively calculating a foreground threshold and a background threshold of the column diagram of the kernel density estimation probability, if the kernel density estimation probability of a certain pixel is less than the foreground threshold, the pixel is determined to be a foreground pixel, if the kernel density estimation probability of the pixel is more than the background threshold, the pixel is determined to be a background pixel.

Description

Self-adapting inner core density is estimated motion detection apparatus and method

Technical field

The background model that the present invention relates in the video analysis field generates, and adopts the dual threshold system of selection of self adaptation prospect, background threshold to carry out the pixel classification.Technical scheme according to employing dual threshold of the present invention can overcome the deficiency that adopts single threshold value classification to exist, and the more important thing is that the selection of threshold value can be carried out adaptively, and can adapt to different scenes.On this basis, the present invention has further proposed the context update model based on probability, upgrades background according to the probability of pixel, can reduce the pollution level of prospect to background, helps moving object detection.Simultaneously, the present invention also utilizes inter-frame difference background model and inner core density to estimate that the classification results background of suddenling change detects, and existing moving target flase drop is surveyed problem when having solved the background sudden change preferably.

Background technology

Isolating moving target from video sequence, is the important research content in the computer vision, can be applied to fields such as Traffic monitoring, people's behavior identification and man-machine interaction.Background subtraction technique (Background subtraction) is a kind of widely used motion detection technique under the video camera quiescent conditions, and the researcher has proposed diverse ways (referring to list of references [1]) for this reason.Elgamma etc. have proposed to estimate that based on inner core density (this method can adapt to different scenes for Kernel DensityEstimation, nonparametric background model KDE) (referring to list of references [2]); Be different from gauss hybrid models (GMM), it makes full use of nearest historical frames information and comes update background module, can adapt to complicated pixel distribution density, overcomes the frequent variations that pixel value takes place at short notice, therefore can obtain estimated result more accurately.But, in list of references [2], by given vacation just rate (False Positive) select threshold value, this needs the priori of scene, needs manual intervention for different scenes, reselects.Anurag Mittal and Nikos Paragios are according to rate of false alarm (false alarmrate) and the rate of failing to report (miss probability) set, adjust threshold value (referring to list of references [3]) by training sample, use this threshold value under given condition, can obtain good relatively classification results.But this method adopts single threshold value, is difficult to solve the contradiction between rate of false alarm and the rate of failing to report like this, unavoidably brings error in classification.Particularly, when scene changes, need reselect the sample training, just can obtain being applicable to the threshold value of this scene.

Summary of the invention

The purpose of this invention is to provide a kind of adaptive inner core density and estimate the motion detection scheme.At first this scheme, probability histogram is analyzed, the dual threshold system of selection of a kind of self adaptation prospect threshold value and background threshold has been proposed, this method does not need the sample training, can be according to the difference of scene, self adaptation is adjusted threshold value, has overcome the error that adopts single threshold value classification to cause.At context update mechanism, the present invention proposes update background module method based on probability, can access more rational background, help motion target detection.In addition, the invention allows for a kind of sudden change background checkout gear and method of utilizing inter-frame difference background model and KDE classification results, the moving target flase drop that has solved preferably when background is undergone mutation is surveyed problem.

To achieve these goals, according to first scheme of the present invention, proposed a kind of self-adapting inner core density and estimated motion detection apparatus, having comprised: receiving element is used to receive the frame of video of input; Initialization unit is used for initialization background model and inter-frame difference background model; The probability calculation statistic unit is used for each pixel at current video frame, according to the background sample, calculates the inner core density estimated probability of this pixel, and the inner core density estimated probability histogram of the pixel of statistics current video frame; Taxon is used for asking for adaptively histogrammic prospect threshold value of inner core density estimated probability and background threshold, if the inner core density estimated probability of a certain pixel, determines then that this pixel is a foreground pixel less than described prospect threshold value; If the inner core density estimated probability of this pixel, determines then that this pixel is a background pixel greater than described background threshold.Described taxon at first internally Density Estimator probability histogram is carried out level and smooth and difference, obtains histogram of difference, on histogram of difference, seeks the starting point P that inner core density estimated probability histogram variation tendency slows down then _LTerminal point P with slow variation _R, at last according to P _LAnd P _R, determine described prospect threshold value and described background threshold.

Preferably, described self-adapting inner core density estimates that motion detection apparatus also comprises: the probabilistic background updating block is used for internal Density Estimator probability and carries out the probabilistic background renewal greater than described prospect threshold value and less than the pixel of described background threshold.

Preferably, described probabilistic background updating block is according to the principle sequential update background sample of first in first out.

Preferably, described self-adapting inner core density estimates that motion detection apparatus also comprises: sudden change context update unit, be used for result according to inter-frame difference background model and taxon, judge whether to have occurred the sudden change pixel, and in the sudden change pixel when satisfying predetermined condition, to the sudden change pixel context update that suddenlys change.

Preferably, determining a pixel when the inter-frame difference background model is background, and taxon is when determining that this pixel is prospect, and described this pixel of sudden change context update unit judges is the sudden change pixel.

Preferably, described sudden change context update unit carries out the connected domain analysis to all sudden change pixels, extract a plurality of predetermined characteristic of connected domain, when the predetermined characteristic of connected domain that and if only if satisfied predetermined condition, described sudden change context update unit was just to the sudden change pixel relevant with this connected domain context update that suddenlys change.

Preferably, the predetermined a plurality of frame of video of described sudden change context update unit after with current video frame are upgraded the background sample.

Preferably, described self-adapting inner core density estimates that motion detection apparatus also comprises: the motion process unit, be used for detected foreground pixel is carried out the connected domain analysis, and in conjunction with the priori of moving target, merge for the zone that belongs to same target, cut apart for the zone that belongs to different target, thereby obtain complete moving target.

To achieve these goals, according to alternative plan of the present invention, proposed a kind of self-adapting inner core density and estimated method for testing motion, comprising: receiving step receives the frame of video of importing; Initialization step, initialization background model and inter-frame difference background model; The probability calculation statistic procedure at each pixel of current video frame, according to the background sample, is calculated the inner core density estimated probability of this pixel, and the inner core density estimated probability histogram of the pixel of statistics current video frame; Classification step is asked for histogrammic prospect threshold value of inner core density estimated probability and background threshold adaptively, if the inner core density estimated probability of a certain pixel, determines then that this pixel is a foreground pixel less than described prospect threshold value; If the inner core density estimated probability of this pixel, determines then that this pixel is a background pixel greater than described background threshold.In described classification step, at first internally the Density Estimator probability histogram carries out level and smooth and difference, obtains histogram of difference, on histogram of difference, seeks the starting point P that inner core density estimated probability histogram variation tendency slows down then _LTerminal point P with slow variation _R, at last according to P _LAnd P _R, determine described prospect threshold value and described background threshold.

Preferably, described self-adapting inner core density estimates that method for testing motion also comprises: the probabilistic background step of updating, internally the Density Estimator probability carries out the probabilistic background renewal greater than described prospect threshold value and less than the pixel of described background threshold.

Preferably, in described probabilistic background step of updating, according to the principle sequential update background sample of first in first out.

Preferably, described self-adapting inner core density estimates that method for testing motion also comprises: sudden change context update step, result according to inter-frame difference background model and classification step, judge whether to have occurred the sudden change pixel, and in the sudden change pixel when satisfying predetermined condition, to the sudden change pixel context update that suddenlys change.

Preferably, in described sudden change context update step, determining a pixel when the inter-frame difference background model is background, and classification step judges that this pixel is the sudden change pixel when determining that this pixel is prospect.

Preferably, in described sudden change context update step, all sudden change pixels are carried out the connected domain analysis, extract a plurality of predetermined characteristic of connected domain, when the predetermined characteristic of connected domain that and if only if satisfies predetermined condition, just to the sudden change pixel relevant context update that suddenlys change with this connected domain.

Preferably, in described sudden change context update step, upgrade the background sample with the predetermined a plurality of frame of video behind the current video frame.

Preferably, described self-adapting inner core density estimates that method for testing motion also comprises: the motion process step, detected foreground pixel is carried out the connected domain analysis, and in conjunction with the priori of moving target, merge for the zone that belongs to same target, cut apart for the zone that belongs to different target, thereby obtain complete moving target.

The present invention has the following advantages and technique effect: adopt dual threshold can overcome the deficiency of single threshold value, the more important thing is that the selection of threshold value does not need manual intervention, can carry out automatically, and can adapt to different scenes.On this basis, the present invention proposes context update model based on probability.This model utilizes the KDE probability, in conjunction with the classification results of prospect and background threshold, pixel is updated in the background according to different probability, thereby has alleviated the pollution level of prospect background, can access more rational background model, help motion target detection.In addition, the present invention also utilizes the rapidity of inter-frame difference, in conjunction with the KDE testing result, has solved problems such as the renewal of background sudden change and motion detection preferably.

Description of drawings

By below in conjunction with description of drawings the preferred embodiments of the present invention, will make above-mentioned and other purpose of the present invention, feature and advantage clearer, wherein:

Fig. 1 is the overall flow figure that estimates method for testing motion according to self-adapting inner core density of the present invention;

Fig. 2 shows the probability distribution and the histogram thereof of pixel;

Fig. 3 shows local probability histogram, histogram of difference and segmentation result;

Segmentation result when the probability that Fig. 4 shows pixel is in different range;

Fig. 5 shows background and the corresponding segmentation result that adopts the probability renewal to obtain;

Fig. 6 is a whole block diagram of estimating motion detection apparatus according to self-adapting inner core density of the present invention.

Embodiment

To a preferred embodiment of the present invention will be described in detail, having omitted in the description process is unnecessary details and function for the present invention with reference to the accompanying drawings, obscures to prevent that the understanding of the present invention from causing.

Fig. 1 shows the overall flow figure that estimates method for testing motion according to self-adapting inner core density of the present invention.Below with reference to Fig. 1, self-adapting inner core density according to the present invention is estimated each step of method for testing motion is described in detail.

1. initial model parameter (step S100 and S102)

To the frame of video (step S100) of input, initialization background model, inter-frame difference background model, other parameters of initialization system, for example: the video frame number of sampling is 50.(step S102).

2. calculate inner core density and estimate (KDE) probability (step S104)

Supposing has M pixel in the frame of video, each pixel has the background sample of N history samples as respective pixel, and then the pixel value of j background sample of i pixel is x _{I, j}, i=1...M, j=1...N.The pixel value of i pixel is x (t) in t moment frame of video _i, t pixel i probability P r (x (t) constantly then _i) can estimate by formula (1):

$\mathrm{Pr}\left(x{\left(t\right)}_i\right)=\frac{1}{N}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}K(x{\left(t\right)}_i-x{\left(t\right)}_{i,j})---\left(1\right)$

X (t) wherein _{I, j}Be the pixel value of j background sample of i pixel in the t moment frame of video, K is the nuclear estimator, if K gets normal distribution N (0, ∑), formula (1) is deformed into formula as follows (2):

$\mathrm{Pr}\left(x{\left(t\right)}_i\right)=\frac{1}{N}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}\underset{m=1}{\overset{d}{\mathrm{\Π}}}\frac{1}{\sqrt{2\mathrm{\π}{\mathrm{\σ}}_{i,m}^2}}{e}^{\frac{{(x{\left(t\right)}_{\mathrm{im}}-x{\left(t\right)}_{i,j})}^2}{2{\mathrm{\σ}}_{i,m}^2}}---\left(2\right)$

Wherein d is a color of pixel component intrinsic dimensionality, x (t) _ImBe x (t) _iM color component.For the coloured image of rgb format, the color of pixel component can be got R, G, B pixel value, supposes that R, G, the B of pixel is separate, and like this, the intrinsic dimensionality d of this moment is 3; σ _{I, m}, m=1 ... 3 is that the nuclear of R, G, B color component of i pixel correspondence is wide, then the wide matrix ∑ of the spatial nuclei of i the pixel correspondence of Gou Chenging _iFor:

${\mathrm{\Σ}}_i=\left(\begin{array}{ccc}{\mathrm{\σ}}_{i,1}& 0& 0\\ 0& {\mathrm{\σ}}_{i,2}& 0\\ 0& 0& {\mathrm{\σ}}_{i,3}\end{array}\right)$

3. statistical pixel distributes and obtains probability histogram (step S106)

According to formula (2), calculate the probability P r (x (t) of t moment M pixel in the input video frame _i), i=1...M for convenience of calculation, is revised as formula as follows (3) with formula (2), calculate N the sampling probability and:

$\mathrm{Pr}\left(x{\left(t\right)}_i\right)=\underset{j=1}{\overset{N}{\mathrm{\Σ}}}\underset{m=1}{\overset{d}{\mathrm{\Π}}}\frac{1}{\sqrt{2\mathrm{\π}{\mathrm{\σ}}_{i,m}^2}}{e}^{\frac{{(x{\left(t\right)}_{\mathrm{im}}-x{\left(t\right)}_{i,j})}^2}{2{\mathrm{\σ}}_{i,m}^2}}---\left(3\right)$

Estimate the wide σ of nuclear of i pixel according to the method in the list of references [2] _{I, m}, when all examine wide σ _{I, m}Maximum possible value σ _MaxWith minimum possibility value σ _MinAfter determining, t at any time, for N sampling of pixel, i pixel probability P r (x (t) in the t moment frame of video _i) possible maximum Pr _MaxCan determine by formula (4):

${\mathrm{Pr}}_{\max }=N*{(1/\sqrt{2\mathrm{\π}{\mathrm{\σ}}_{\min }^2})}^d---\left(4\right)$

Pr (x (t) _i) may value minimum value Pr _Min, not only depend on σ _Max, also, can get 0 by exponential term decision in the formula (3).[0, Pr _Max] multiply by scale factor β and quantize, obtain [0, Pr _h].Then, travel through the probability of all pixels, form probability distribution histogram hist _p(as shown in Equation (5)):

${\mathrm{hist}}_p\left(j\right)=\left\{\begin{array}{cc}{\mathrm{hist}}_p\left(j\right)+1& \mathrm{if}({\mathrm{Pr}}_h\left(i\right)=j)\\ {\mathrm{hist}}_p\left(j\right)& \mathrm{else}\end{array}\right.---\left(5\right)$

Fig. 2 shows the probability distribution and the histogram thereof of pixel.As shown in Figure 2, (a) be original image, moving target in the scene is the pedestrian, (b) is the probability distribution of (a), the Width of x direction of principal axis correspondence image, the short transverse of y direction of principal axis correspondence image, the z axle is represented the size of probable value, (c) is the vertical view of (b), the Width of x direction of principal axis correspondence image, the short transverse of y direction of principal axis correspondence image, (d) probability histogram hist _p, the probable value value of u axle remarked pixel, the v axle represents that probability equals the number of pixels of this value.

4. the selection of self adaptation prospect threshold value and background threshold (step S108)

With gaussian kernel function to hist _pSmoothly obtain hist _Ps((d) among Fig. 2).To hist _Ps(i) and hist _Ps(i+1) carry out difference according to formula (6), further remove interference with predetermined little threshold value th, obtain histogram of difference hist at this _Diff(as shown in Equation (6)):

${\mathrm{hist}}_{\mathrm{diff}}\left(i\right)=\left\{\begin{array}{cc}\mathrm{diff}& \mathrm{ifdiff}>\mathrm{<figure-callout\; id="0"\; label="th"\; filenames="S2007101773100E00021.png,S2007101773100E00031.png"\; state="\{\{state\}\}">th</figure-callout>}\\ 0& \mathrm{else}\end{array}\right.---\left(6\right)$

diff＝abs(hist _ps(i+1)-hist _ps(i))

Fig. 3 shows local probability histogram, histogram of difference and segmentation result.As shown in Figure 3, (a) be the partial enlarged drawing of Fig. 2 (d), (b) for obtain by formula (6) with the corresponding histogram of difference of Fig. 2 (d), (c) be the partial enlarged drawing of (b), the probable value value of u axle remarked pixel among the figure, the v axle represents that probability equals the number of pixels of this value.(d) be to choose the segmentation result of first zero point of histogram of difference as prospect threshold value (this point is 6 for histogram changes the point that tends towards stability among Fig. 3 (c)), the Width of x direction of principal axis correspondence image, the short transverse of y direction of principal axis correspondence image.

For this reason, the following prospect threshold value Thf that chooses:

1) by histogram of difference hist _Diff((b) among Fig. 3) seeks histogram by the slow breakover point P of abrupt change _z(marking among Fig. 3 (c)), i.e. P _zSatisfy following formula:

$P_z=\underset{i}{\mathrm{Arg}}\max \left(\mathrm{abs}\left(\frac{{\mathrm{hist}}_{\mathrm{diff}}(i-1)-{\mathrm{hist}}_{\mathrm{diff}}\left(i\right)}{{\mathrm{hist}}_{\mathrm{diff}}\left(i\right)-{\mathrm{hist}}_{\mathrm{diff}}(i+1)}\right)\right)---\left(7\right)$

2) at histogram of difference hist _Diff((b) among Fig. 3) goes up from P _zThe starting point P that variation tendency slows down is sought in beginning to the right _L(marking among Fig. 3 (c));

3) at histogram of difference hist _Diff((b) among Fig. 3) goes up from P _LThe terminal point P that slowly changes is sought in beginning to the right _R(marking among Fig. 3 (c)) is up to P _Max(can be obtained by the priori in the scene, for target shared maximum ratio in whole scene, mark among Fig. 3 (c)) ends;

4) get P as Thf _LIn time, can bring and misses the part foreground target, when Thf gets P _RIn time, can survey part background flase drop for prospect, handles according to formula (8) for this reason, determines suitable prospect threshold value Thf, and wherein a is the coefficient greater than 1:

Thf＝P _L+(P _R-P _L)/a (8)

Choosing of adaptive background threshold value Thb is the probability that the prospect mistake is divided into background in order to reduce, thereby helps rationally upgrading background.If directly adopt Thf that background and prospect are divided into two classes, can be used as background to some pixels that belong to prospect, and be updated in the background sample, can cause the omission of foreground pixel to survey like this.Therefore, need suitably to improve the threshold value that is judged as background pixel.Therefore, in the present invention, P _RThreshold value as a setting:

Thb＝P _R (9)

Segmentation result when the probability that Fig. 4 shows pixel is in different range:

(a) 0≤Pr (x (t) _i)＜P _LThe time the result, as can be seen from the figure disappearance appears in the partial pixel of moving target (pedestrian), but noise is very little; (b) be P _L≤ Pr (x (t) _i)＜P _RThe time the result, former cause moves to static target (railing) and also is detected, and has more noise; (c) be Pr (x (t) _i) 〉=P _RThe time segmentation result, as can be seen from the figure have only the pixel that belongs to prospect on a small quantity to be detected as background; (d) for adopting formula (8) to ask for threshold value and carry out sorting result according to formula (10), moving target (pedestrian) omission is surveyed pixel seldom, has the pixel that belongs to background on a small quantity to be taken as prospect and detects.The Width of x direction of principal axis correspondence image among the figure, the short transverse of y direction of principal axis correspondence image.

5. pixel classification (step S110)

According to the probable value of pixel, classify according to following formula (10), if the probable value of a certain pixel, determines then that this pixel is the motion pixel less than prospect threshold value Thf, and export this pixel to the motion process unit and handle.

Pr(x(t) _i)＜Thf (10)

Otherwise carry out context update according to the 6th step.

6. based on the context update model (step S120, S122 and S124) of probability

By the pixel probability that formula (3) obtains, in fact reflected the similarity degree of this pixel and background.Therefore, the foundation that can upgrade probable value as a setting.Consideration is at t moment probability P r (x (t) _i), if obvious Pr (x (t) _i) less than prospect threshold value Thf, pixel x (t) _iBe prospect, when context update, do not do consideration; If Pr is (x (t) _i) greater than background threshold Thb, pixel x (t) _iBe background, can directly be updated to this pixel in the background and go; The pixel of probability between Thf and Thb can be carried out following processing according to formula (11), upgrades according to probability again:

${\mathrm{Pr}}_b\left(x{\left(t\right)}_i\right)=\left\{\begin{array}{cc}1& \mathrm{ifPr}\left(x{\left(t\right)}_i\right)\&GreaterEqual;\mathrm{<figure-callout\; id="0"\; label="Thb"\; filenames="S2007101773100E00021.png,S2007101773100E00031.png"\; state="\{\{state\}\}">Thb</figure-callout>}\\ 0& \mathrm{ifPr}\left(x{\left(t\right)}_i\right)<\mathrm{Thf}\\ \frac{\mathrm{Pr}\left(x{\left(t\right)}_i\right)-\mathrm{Thf}}{\mathrm{Thb}-\mathrm{Thf}}& \mathrm{others}\end{array}\right.---\left(11\right)$

Obtaining the probability P r that present frame is used to upgrade background _b(x (t) _i) afterwards, just can upgrade the background sample according to probability.Adopt the principle of first-in first-out to carry out in proper order when upgrading the background sample.For example upgrade j-1 the background sample x (t) of i pixel constantly at t-1 _{I, j-1}, then upgrade j the background sample x (t) of i pixel constantly at t _{I, j}Suppose to upgrade constantly j-1 the background sample x (t) of i pixel at t-1 _{I, j-1}, then t sample constantly upgrades as shown in Equation (12), wherein i=1...M:

X (t) _{I, j}=Pr _b(x (t) _i) * x (t) _i+ (1-Pr _b(x (t) _i)) * x (t-1) _{I, j}(12), then reappraise the wide σ of nuclear of i pixel according to the method in the list of references [2] if upgraded the background sample of i pixel _{I, m}

For example, upgrade the 9th the background sample x (t) of i pixel constantly at t-1 _{I, 9}, then upgrade the 10th the background sample x (t) of i pixel constantly at t _{I, 10}Suppose x (t) _iBe 130, Pr _b(x (t) _i) be 0.8, x (t-1) _{I, 10}Be 120, the background sample after then upgrading is:

x(t) _i，10＝0.8*130+(1-0.8)*120＝128。

Fig. 5 shows and adopts probability to upgrade background and the foreground detection result who obtains:

(a) be a frame in the original video, (b) do not adopt probability to upgrade the background that obtains for this moment, (c) for adopting probability constantly, this upgrades the background that obtains, (d) segmentation result for adopting this paper algorithm to obtain, and the white pixel among the figure is detected moving target.(b) in the comparison diagram 5 and (c) in the heavy line square frame in the zone, adopt probability to upgrade the background that obtains, have only a spot of foreground area to be updated in the background and go, and do not adopt probability to upgrade, then more foreground area is updated in the background goes.

Adopt formula (12) reliably to upgrade to the pixel that belongs to background, the flase drop survey is upgraded with probability for the pixel of background for belonging to prospect, and the pixel that belongs to prospect is not upgraded, thereby has avoided the pollution to background, and can reflect the slow variation of background.

Carry out the background sudden change according to the 7th step and detect, and upgrade accordingly.

7. detection, the new model (step S114, S116 and S118) more of sudden change background

If B (i), C (i), L (i) be the pixel value of background pixel value, current frame pixel value and the former frame of correspondence position pixel i respectively, wherein th1 is a predefined little threshold value, and Cn (i) is used for adding up the counting variable (as shown in Equation (13)) that background changes:

When Cn (i) satisfies formula (14), think that then this pixel i is a background, need upgrade the background pixel value B (i) of correspondence position with current pixel value C (i), wherein th2 is a turnover rate:

$B\left(i\right)=\left\{\begin{array}{cc}C\left(i\right)& \mathrm{ifCn}\left(i\right)>\mathrm{th}2\\ B\left(i\right)& \mathrm{else}\end{array}\right.---\left(14\right)$

The speed of context update depends on turnover rate th2, owing to do not relate to complex calculations, so the speed of service is very fast.When the th2 value hour, can be updated to prospect in the background and go, thereby detect less than foreground target; When turnover rate was higher, the background of variation may not can be upgraded in time, and mistake is used as foreground target to background.In the present invention, just utilize the new features more fast of above-mentioned update scheme, the foundation that changes as a setting, so the selection of th2 is difficult for too highly, for example, can choose following numerical value: the integer value between 30～40, as 30,35,40 etc.

The pixel that belongs to background before background is not updated, may be detected to prospect by KDE, can utilize this specific character to come the marked change of detection background, and as shown in Equation (15), wherein B represents background pixel set, B _KdeBe used for representing the mask whether background undergos mutation, i=1 ... M:

If B _KdeBe 1, illustrate that sudden change has taken place background.Otherwise background is not undergone mutation.

By the pixel that the detected background of formula (15) is undergone mutation, should be immediately as the foundation of upgrading KDE background sample.The background of sudden change generally shows as regionality, and area is bigger.For improving the robustness of system, all sudden change pixels that satisfy formula (15) are carried out the connected domain analysis, features such as the area of extraction connected domain, number, maximum length, Breadth Maximum.The condition that satisfies background sudden change when these features (for example, the area of connected domain greater than the predetermined ratio of scene, number greater than the maximum target number in the scene, maximum length greater than the maximum target length in the scene, width greater than maximum target width in the scene etc.) time, think that just sudden change has taken place background really, disturb thereby reduce, improve system accuracy.

When background was undergone mutation really, the N two field picture after beginning with present frame upgraded the sample in the background, guarantees that background model is upgraded in time, improved the robustness that detects.Simultaneously, the present invention also can solve the deadlock situation (deadlock situations) (referring to list of references [2,4]) in the context update.

Fig. 6 is a whole block diagram of estimating motion detection apparatus according to self-adapting inner core density of the present invention.

Particularly, estimate that according to self-adapting inner core density of the present invention motion detection apparatus 500 comprises: receiving element 510, initialization unit 520, probability calculation statistic unit 530, taxon 540, sudden change context update unit 550, probabilistic background updating block 560 and motion process unit 570.

Receiving element 510 is used to receive the frame of video of input.Initialization unit 520 is used for initialization background model, inter-frame difference background model, and other parameters of initialization system, and for example: the video frame number of sampling is 50.

Probability calculation statistic unit 530 is used for each pixel at present frame, according to the background sample, calculates the inner core density estimated probability of this pixel, and the inner core density estimated probability histogram of the pixel of statistics present frame.

Taxon 540 is used for asking for adaptively histogrammic prospect threshold value of inner core density estimated probability and background threshold, if the inner core density estimated probability of a certain pixel is less than described prospect threshold value, determine that then this pixel is a foreground pixel, and export this pixel to motion process unit 570; If the inner core density estimated probability of this pixel, determines then that this pixel is a background pixel, directly is updated to this pixel in the background greater than described background threshold; If the inner core density estimated probability of this pixel then exports this pixel to probabilistic background updating block 560 greater than described prospect threshold value and less than described background threshold.

Probabilistic background updating block 560 is used for internal Density Estimator probability and carries out the probabilistic background renewal greater than described prospect threshold value and less than the pixel of described background threshold.In described probabilistic background renewal process, probabilistic background updating block 560 is according to the principle sequential update background sample of first in first out.

The result that sudden change context update unit 550 is used for according to inter-frame difference background model and taxon judges whether to have occurred the sudden change pixel, and when the sudden change pixel satisfies predetermined condition, to the sudden change pixel context update that suddenlys change.

Motion process unit 570 is used for detected foreground pixel is carried out the connected domain analysis, and in conjunction with the priori of moving target, as the minimum and maximum width of target, minimum and maximum height, minimum and maximum area etc., merge for the zone that belongs to same target, cut apart for the zone that belongs to different target, thereby obtain complete moving target.

So far invention has been described in conjunction with the preferred embodiments.Should be appreciated that those skilled in the art can carry out various other change, replacement and interpolations under the situation that does not break away from the spirit and scope of the present invention.Therefore, scope of the present invention is not limited to above-mentioned specific embodiment, and should be limited by claims.

The list of references tabulation

[1]M?Piccardi，Background?subtraction?techniques：areview.Systems.Man?and?Cybernetics[J]，2004.4：3099-3104；

[2]A.Elgammal，D.Hanvood，and?L.S.Davis，Non-parametric?model?for?background?subtraction[C].Proc.ECCV?2000.751-767；

[3]Anurag?Mittal，Nikos?Paragios.?Motion-BasedBackground?Subtraction?using?Adaptive?Kernel?Density?Estimation.[C].CVPR，2004.II(2)：302-309；

[4]Hanzi?Wang，David?Suter.A?consensus-based?method?fortracking：Modeling?background?scenario?and?foregroundappearance[J].Pattern?Recognition.2007.40：1091-1105.

Claims (44)

1. a self-adapting inner core density is estimated motion detection apparatus, comprising:

Receiving element is used to receive the frame of video of input;

Initialization unit is used for initialization background model and inter-frame difference background model;

The probability calculation statistic unit is used for each pixel at current video frame, according to the background sample, calculates the inner core density estimated probability of this pixel, and the inner core density estimated probability histogram of the pixel of statistics current video frame;

Taxon is used for asking for adaptively histogrammic prospect threshold value of inner core density estimated probability and background threshold, if the inner core density estimated probability of a certain pixel, determines then that this pixel is a foreground pixel less than described prospect threshold value; If the inner core density estimated probability of this pixel, determines then that this pixel is a background pixel greater than described background threshold,

Wherein, described taxon at first internally Density Estimator probability histogram is carried out level and smooth and difference, obtains histogram of difference, on histogram of difference, seeks the starting point P that inner core density estimated probability histogram variation tendency slows down then _LTerminal point P with slow variation _R, at last according to P _LAnd P _R, determine described prospect threshold value and described background threshold.

2. self-adapting inner core density according to claim 1 is estimated motion detection apparatus, it is characterized in that also comprising:

The probabilistic background updating block is used for internal Density Estimator probability and carries out the probabilistic background renewal greater than described prospect threshold value and less than the pixel of described background threshold.

3. self-adapting inner core density according to claim 2 is estimated motion detection apparatus, it is characterized in that the principle sequential update background sample of described probabilistic background updating block according to first in first out.

4. self-adapting inner core density according to claim 1 is estimated motion detection apparatus, it is characterized in that also comprising:

Sudden change context update unit is used for the result according to inter-frame difference background model and taxon, judges whether to have occurred the sudden change pixel, and when the sudden change pixel satisfies predetermined condition, to the sudden change pixel context update that suddenlys change.

5. self-adapting inner core density according to claim 4 is estimated motion detection apparatus, it is characterized in that determining that when the inter-frame difference background model pixel is a background, and taxon is when determining that this pixel is prospect, and described this pixel of sudden change context update unit judges is the sudden change pixel.

6. self-adapting inner core density according to claim 4 is estimated motion detection apparatus, it is characterized in that described sudden change context update unit carries out the connected domain analysis to all sudden change pixels, extract a plurality of predetermined characteristic of connected domain, when the predetermined characteristic of connected domain that and if only if satisfied predetermined condition, described sudden change context update unit was just to the sudden change pixel relevant with this connected domain context update that suddenlys change.

7. self-adapting inner core density according to claim 6 is estimated motion detection apparatus, it is characterized in that the predetermined a plurality of frame of video after described sudden change context update unit is with current video frame are upgraded the background sample.

8. self-adapting inner core density according to claim 5 is estimated motion detection apparatus, it is characterized in that described sudden change context update unit carries out the connected domain analysis to all sudden change pixels, extract a plurality of predetermined characteristic of connected domain, when the predetermined characteristic of connected domain that and if only if satisfied predetermined condition, described sudden change context update unit was just to the sudden change pixel relevant with this connected domain context update that suddenlys change.

9. self-adapting inner core density according to claim 8 is estimated motion detection apparatus, it is characterized in that the predetermined a plurality of frame of video after described sudden change context update unit is with current video frame are upgraded the background sample.

10. self-adapting inner core density according to claim 2 is estimated motion detection apparatus, it is characterized in that also comprising:

Sudden change context update unit is used for the result according to inter-frame difference background model and taxon, judges whether to have occurred the sudden change pixel, and when the sudden change pixel satisfies predetermined condition, to the sudden change pixel context update that suddenlys change.

11. self-adapting inner core density according to claim 10 is estimated motion detection apparatus, it is characterized in that determining that when the inter-frame difference background model pixel is a background, and taxon is when determining that this pixel is prospect, and described this pixel of sudden change context update unit judges is the sudden change pixel.

12. self-adapting inner core density according to claim 10 is estimated motion detection apparatus, it is characterized in that described sudden change context update unit carries out the connected domain analysis to all sudden change pixels, extract a plurality of predetermined characteristic of connected domain, when the predetermined characteristic of connected domain that and if only if satisfied predetermined condition, described sudden change context update unit was just to the sudden change pixel relevant with this connected domain context update that suddenlys change.

13. self-adapting inner core density according to claim 12 is estimated motion detection apparatus, it is characterized in that the predetermined a plurality of frame of video after described sudden change context update unit is with current video frame are upgraded the background sample.

14. self-adapting inner core density according to claim 11 is estimated motion detection apparatus, it is characterized in that described sudden change context update unit carries out the connected domain analysis to all sudden change pixels, extract a plurality of predetermined characteristic of connected domain, when the predetermined characteristic of connected domain that and if only if satisfied predetermined condition, described sudden change context update unit was just to the sudden change pixel relevant with this connected domain context update that suddenlys change.

15. self-adapting inner core density according to claim 14 is estimated motion detection apparatus, it is characterized in that the predetermined a plurality of frame of video after described sudden change context update unit is with current video frame are upgraded the background sample.

16. self-adapting inner core density according to claim 3 is estimated motion detection apparatus, it is characterized in that also comprising:

Sudden change context update unit is used for the result according to inter-frame difference background model and taxon, judges whether to have occurred the sudden change pixel, and when the sudden change pixel satisfies predetermined condition, to the sudden change pixel context update that suddenlys change.

17. self-adapting inner core density according to claim 16 is estimated motion detection apparatus, it is characterized in that determining that when the inter-frame difference background model pixel is a background, and taxon is when determining that this pixel is prospect, and described this pixel of sudden change context update unit judges is the sudden change pixel.

18. self-adapting inner core density according to claim 16 is estimated motion detection apparatus, it is characterized in that described sudden change context update unit carries out the connected domain analysis to all sudden change pixels, extract a plurality of predetermined characteristic of connected domain, when the predetermined characteristic of connected domain that and if only if satisfied predetermined condition, described sudden change context update unit was just to the sudden change pixel relevant with this connected domain context update that suddenlys change.

19. self-adapting inner core density according to claim 18 is estimated motion detection apparatus, it is characterized in that the predetermined a plurality of frame of video after described sudden change context update unit is with current video frame are upgraded the background sample.

20. self-adapting inner core density according to claim 17 is estimated motion detection apparatus, it is characterized in that described sudden change context update unit carries out the connected domain analysis to all sudden change pixels, extract a plurality of predetermined characteristic of connected domain, when the predetermined characteristic of connected domain that and if only if satisfied predetermined condition, described sudden change context update unit was just to the sudden change pixel relevant with this connected domain context update that suddenlys change.

21. self-adapting inner core density according to claim 20 is estimated motion detection apparatus, it is characterized in that the predetermined a plurality of frame of video after described sudden change context update unit is with current video frame are upgraded the background sample.

22. estimate motion detection apparatus according to the described self-adapting inner core density of one of claim 1～21, it is characterized in that also comprising:

The motion process unit, be used for detected foreground pixel is carried out the connected domain analysis, and, merge for the zone that belongs to same target in conjunction with the priori of moving target, cut apart for the zone that belongs to different target, thereby obtain complete moving target.

23. a self-adapting inner core density is estimated method for testing motion, comprising:

Receiving step receives the frame of video of importing;

Initialization step, initialization background model and inter-frame difference background model;

The probability calculation statistic procedure at each pixel of current video frame, according to the background sample, is calculated the inner core density estimated probability of this pixel, and the inner core density estimated probability histogram of the pixel of statistics current video frame;

Classification step is asked for histogrammic prospect threshold value of inner core density estimated probability and background threshold adaptively, if the inner core density estimated probability of a certain pixel, determines then that this pixel is a foreground pixel less than described prospect threshold value; If the inner core density estimated probability of this pixel, determines then that this pixel is a background pixel greater than described background threshold,

Wherein, in described classification step, at first internally the Density Estimator probability histogram carries out level and smooth and difference, obtains histogram of difference, on histogram of difference, seeks the starting point P that inner core density estimated probability histogram variation tendency slows down then _LTerminal point P with slow variation _R, at last according to P _LAnd P _R, determine described prospect threshold value and described background threshold.

24. self-adapting inner core density according to claim 23 is estimated method for testing motion, it is characterized in that also comprising:

The probabilistic background step of updating, internally the Density Estimator probability carries out the probabilistic background renewal greater than described prospect threshold value and less than the pixel of described background threshold.

25. self-adapting inner core density according to claim 24 is estimated method for testing motion, it is characterized in that in described probabilistic background step of updating, according to the principle sequential update background sample of first in first out.

26. self-adapting inner core density according to claim 23 is estimated method for testing motion, it is characterized in that also comprising:

Sudden change context update step according to the result of inter-frame difference background model and classification step, judges whether to have occurred the sudden change pixel, and when the sudden change pixel satisfies predetermined condition, to the sudden change pixel context update that suddenlys change.

27. self-adapting inner core density according to claim 26 is estimated method for testing motion, it is characterized in that in described sudden change context update step, determining a pixel when the inter-frame difference background model is background, and classification step judges that this pixel is the sudden change pixel when determining that this pixel is prospect.

28. self-adapting inner core density according to claim 26 is estimated method for testing motion, it is characterized in that in described sudden change context update step, all sudden change pixels are carried out the connected domain analysis, extract a plurality of predetermined characteristic of connected domain, when the predetermined characteristic of connected domain that and if only if satisfies predetermined condition, just to the sudden change pixel relevant context update that suddenlys change with this connected domain.

29. self-adapting inner core density according to claim 28 is estimated method for testing motion, it is characterized in that upgrading the background sample with the predetermined a plurality of frame of video behind the current video frame in described sudden change context update step.

30. self-adapting inner core density according to claim 27 is estimated method for testing motion, it is characterized in that in described sudden change context update step, all sudden change pixels are carried out the connected domain analysis, extract a plurality of predetermined characteristic of connected domain, when the predetermined characteristic of connected domain that and if only if satisfies predetermined condition, just to the sudden change pixel relevant context update that suddenlys change with this connected domain.

31. self-adapting inner core density according to claim 30 is estimated method for testing motion, it is characterized in that upgrading the background sample with the predetermined a plurality of frame of video behind the current video frame in described sudden change context update step.

32. self-adapting inner core density according to claim 24 is estimated method for testing motion, it is characterized in that also comprising:

Sudden change context update step according to the result of inter-frame difference background model and classification step, judges whether to have occurred the sudden change pixel, and when the sudden change pixel satisfies predetermined condition, to the sudden change pixel context update that suddenlys change.

33. self-adapting inner core density according to claim 32 is estimated method for testing motion, it is characterized in that in described sudden change context update step, determining a pixel when the inter-frame difference background model is background, and classification step judges that this pixel is the sudden change pixel when determining that this pixel is prospect.

34. self-adapting inner core density according to claim 32 is estimated method for testing motion, it is characterized in that in described sudden change context update step, all sudden change pixels are carried out the connected domain analysis, extract a plurality of predetermined characteristic of connected domain, when the predetermined characteristic of connected domain that and if only if satisfies predetermined condition, just to the sudden change pixel relevant context update that suddenlys change with this connected domain.

35. self-adapting inner core density according to claim 34 is estimated method for testing motion, it is characterized in that upgrading the background sample with the predetermined a plurality of frame of video behind the current video frame in described sudden change context update step.

36. self-adapting inner core density according to claim 33 is estimated method for testing motion, it is characterized in that in described sudden change context update step, all sudden change pixels are carried out the connected domain analysis, extract a plurality of predetermined characteristic of connected domain, when the predetermined characteristic of connected domain that and if only if satisfies predetermined condition, just to the sudden change pixel relevant context update that suddenlys change with this connected domain.

37. self-adapting inner core density according to claim 36 is estimated method for testing motion, it is characterized in that upgrading the background sample with the predetermined a plurality of frame of video behind the current video frame in described sudden change context update step.

38. self-adapting inner core density according to claim 25 is estimated method for testing motion, it is characterized in that also comprising:

Sudden change context update step according to the result of inter-frame difference background model and classification step, judges whether to have occurred the sudden change pixel, and when the sudden change pixel satisfies predetermined condition, to the sudden change pixel context update that suddenlys change.

39. estimate method for testing motion according to the described self-adapting inner core density of claim 38, it is characterized in that in described sudden change context update step, determining a pixel when the inter-frame difference background model is background, and classification step judges that this pixel is the sudden change pixel when determining that this pixel is prospect.

40. estimate method for testing motion according to the described self-adapting inner core density of claim 38, it is characterized in that in described sudden change context update step, all sudden change pixels are carried out the connected domain analysis, extract a plurality of predetermined characteristic of connected domain, when the predetermined characteristic of connected domain that and if only if satisfies predetermined condition, just to the sudden change pixel relevant context update that suddenlys change with this connected domain.

41. estimate method for testing motion according to the described self-adapting inner core density of claim 40, it is characterized in that in described sudden change context update step, upgrading the background sample with the predetermined a plurality of frame of video behind the current video frame.

42. estimate method for testing motion according to the described self-adapting inner core density of claim 39, it is characterized in that in described sudden change context update step, all sudden change pixels are carried out the connected domain analysis, extract a plurality of predetermined characteristic of connected domain, when the predetermined characteristic of connected domain that and if only if satisfies predetermined condition, just to the sudden change pixel relevant context update that suddenlys change with this connected domain.

43. estimate method for testing motion according to the described self-adapting inner core density of claim 42, it is characterized in that in described sudden change context update step, upgrading the background sample with the predetermined a plurality of frame of video behind the current video frame.

44. estimate method for testing motion according to the described self-adapting inner core density of one of claim 23～43, it is characterized in that also comprising:

The motion process step is carried out the connected domain analysis to detected foreground pixel, and in conjunction with the priori of moving target, merges for the zone that belongs to same target, cuts apart for the zone that belongs to different target, thereby obtains complete moving target.

Images