CN104077786A - Moving object detection method based on self-adapting kernel density estimation model - Google Patents

Abstract

The invention discloses a moving object detection method based on a self-adapting kernel density estimation model. The method includes the steps that a background model based on kernel density estimation is built according to input videos, and accordingly the width of a probability density function of pixels of the input videos and the width of a kernel function are obtained; according to the width of a probability density function of gray values of the pixels of the input videos and the width of the kernel function, a probability value interval of a sample point is determined, then probabilities of the gray values of all the pixels are traversed, and accordingly a probability distribution histogram is formed; secondary linear interpolation and small threshold difference are carried out on the probability distribution histogram, and accordingly a difference histogram is obtained; a foreground threshold and a background threshold are solved in a self-adaptive mode according to the difference histogram; the built background model is updated according to the probability values of the gray values of the pixels of the input videos, the foreground threshold and the background threshold. The method has the advantages of being little in calculation quantity, good in instantaneity and few in error, and can be widely used in the field of computer visual analysis.

Description

A kind of moving target detecting method based on self-adaptive kernel Density estimating model

Technical field

The present invention relates to computer vision analysis field, especially a kind of moving target detecting method based on self-adaptive kernel Density estimating model.

Background technology

The moving object detection application of video sequence is very extensive, comprises the fields such as intelligent transportation, bank monitoring and man-machine interaction.Background subtraction point-score is a kind of moving target detecting method of commonly using, and it is mainly background to be rejected from present frame by " subtracting computing ", thereby obtains complete foreground moving target.Background modeling is very responsive to dynamic environment such as illumination, weathers, when the renewal of background model can not adapt to the variation of dynamic environment well, will affect to a great extent the testing result of moving target.Therefore, how to obtain stable and reliably background be the key point of the moving object detection based on background subtraction point-score.

Background modeling is to cut apart foreground target and background image according to the statistical property of background area.The modeling method of imparametrization becomes a study hotspot of background modeling because of its definition advantage that does not rely on model.The people such as Elgammal have proposed the nonparametric background model based on Density Estimator, the method makes full use of nearest historical frames information and represents background model, can adapt to complicated pixel distribution density, overcome the frequent variations that pixel value occurs at short notice, but because it adopts single threshold, inevitably bring error in classification, and its calculation cost is too high.The people such as Mao Yanfen adopt the Density Estimator method based on diversity sampling principle to set up background model, from former sample, extract and there is high frequency and multifarious compared with small sample, thereby reduce the calculated amount of background estimating, although improved time efficiency, but still cannot meet the requirement of real-time of system.The people such as Hu Min propose before Density Estimator, first to use inter-frame difference in conjunction with background subtraction method, filter out relatively static background dot, only typical motion pixel is carried out to Density Estimator, to reduce calculated amount, but introduce inter-frame difference in conjunction with background subtraction timesharing, also introduce its error, easily caused the sufficiently complete of foreground target extraction.

In sum, need in the industry at present that a kind of calculated amount is little badly, real-time better and the less moving target detecting method based on Density Estimator model of error.

Summary of the invention

In order to solve the problems of the technologies described above, the object of the invention is: provide that a kind of calculated amount is little, real-time better and the less moving target detecting method based on self-adaptive kernel Density estimating model of error.

The technical solution adopted for the present invention to solve the technical problems is: a kind of moving target detecting method based on self-adaptive kernel Density estimating model, comprising:

A. according to input video, build the background model based on Density Estimator, thereby obtain the probability density function of input video pixel and the width of kernel function;

B. according to the width of the probability density function of input video grey scale pixel value and kernel function, determine that the parameter probability valuing of sample point is interval, then travel through the probability of all pixel gray-scale values, thus formation probability distribution histogram;

C. probability distribution histogram is carried out to secondary linear interpolation and little threshold value difference, thereby obtain histogram of difference;

D. according to histogram of difference self-adaptation, ask for prospect threshold value and background threshold;

E. according to the probable value of input video pixel gray-scale value, prospect threshold value and background threshold, the background model building is upgraded.

Further, described steps A, it comprises:

A1. according to input video and gaussian kernel function, obtain the probability density function of t pixel i constantly, the expression formula of the probability density function P of described moment t pixel i (x (t) i) is:

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

Wherein, N is the number of each pixel background sample, the dimension that d is pixel, the dimension that m is color space, the gray-scale value of m the Color Channel that x (t) im is i pixel, x (t) _{im, j}be m Color Channel gray-scale value of j sample point of i pixel, σ _mit is the cuclear density width of m Color Channel;

A2. according to the median of the absolute value of difference between consecutive frame gray-scale pixels sample, calculate the cuclear density width of gaussian kernel function, the expression formula of the cuclear density width cs of described gaussian kernel function is:

$\mathrm{\σ}=\frac{\mathrm{median}\left(\right|x_{i,j}-x_{i,j}+1\left|\right)}{0.68\sqrt{2}},$

Wherein, median function is for asking median function, x (t) _{i, j}be the gray-scale value of j background sample in i pixel, x (t) _{i, j+1}it is the gray-scale value of j+1 background sample in i pixel.

Further, described step B, it comprises:

B1. according to the minimum possibility of the width of kernel function value σ _mindetermine the maximum probability possibility value at t moment N sample point, the expression formula of the described maximum probability possibility value Pmax at t moment N sample point is:

$P_{\max }=\frac{1}{{\left(\sqrt{2\mathrm{\π}}{\mathrm{\σ}}_{\min }\right)}^d};$

B2. the minimum possibility of the probability value of t moment N sample point is taken as to 0, and the probability at t moment N sample point is multiplied by scale factor, wherein α > 1, thereby the interval of the probability at t moment N sample point is quantified as to [0, α Pmax];

B3. according to the interval after quantizing, travel through the probability of all pixels, thus formation probability distribution histogram Hp, and the expression formula of described probability distribution histogram Hp is:

Wherein, Hp (j) is the histogrammic ordinate of probability distribution, and Hp (j) represents that j sample is at the number of current time probability same pixel point, and Hp ' (j) represents that j sample is at the number of previous moment probability same pixel point.

Further, described step C, it comprises:

C1. adopt secondary linear interpolation method to carry out smoothly probability distribution histogram Hp, thus the histogram Hps after obtaining smoothly;

C2. adopt little threshold value to carry out difference to the histogram Hps after level and smooth, thereby obtain histogram of difference H _b, described histogram of difference H _bexpression formula be:

Wherein, H _b(j) be the ordinate of j sample in histogram of difference, b is adjacent two sample point minimum differences in the histogram after level and smooth; Min function is the function of minimizing, the number of probability same pixel point in j sample of histogram, a j+1 sample and j+2 sample after Hps (j), Hps (j+1) and Hps (j+2) represent respectively smoothly, Ts is default little threshold value.

Further, described step D, it comprises:

D1. by histogram of difference H _bfind the turning point P that histogram is delayed by abrupt change _z, described turning point P _zmeet:

$P_Z=\arg \underset{i}{\max }\left|\frac{H_b\left(i\right)-H_b(i+1)}{H_b(i+1)-H_b(i+2)}\right|;$

D2. at histogram of difference H _babove from Pz, start to find to the right the starting point P that variation tendency slows down _l;

D3. at histogram of difference H _bupper from P _lstart to find to the right the terminal P slowly changing _r, until the maximum probability point P of prospect _m;

D4. according to P _land P _rself-adaptation is calculated prospect threshold value and background threshold, described prospect threshold value T _fwith background threshold T _bcomputing formula be respectively:

T _f＝P _L-β(P _R-P _L)，

T _b＝P _R+β(P _R-P _L)，

Wherein, β is adaptive factor, and β < 1.

Further, described step e, it comprises:

E1. according to the renewal probability of the magnitude relationship calculating pixel point of the probable value of input video pixel gray-scale value and prospect threshold value, background threshold;

E2. according to the pixel calculating, upgrading probability upgrades the sample value of input video pixel according to first in first out.

Further, described step e 1, it is specially:

Probable value P (x (t) i) and prospect threshold value T according to input video pixel in the t moment _f, background threshold T _bthe renewal probability of magnitude relationship calculating pixel point, the renewal probability P of described pixel _bthe expression formula of (x (t) i) is:

Further, described step e 2, it is specially:

According to the pixel calculating, upgrade probability P _b(x (t) i) upgrades in t j sample value constantly input video pixel i according to first in first out, the sample value x after renewal (t) i, and the expression formula of j is:

x(t) _i，j＝P _b(x(t) _i)·(x(t) _i)+(1-P _b(x(t) _i))·(x(t-1) _i，j)，

Wherein, x (t-1) i, j is that input video pixel i is in t-1 j sample value constantly.

The invention has the beneficial effects as follows: by histogram analysis self-adaptation, ask for prospect threshold value and background threshold, and upgrade according to prospect threshold value and the background threshold asked for, method is simple, and calculated amount is little; The dual threshold method that has adopted prospect threshold value to combine with background threshold is carried out pixel classification, has overcome the error in classification of single threshold method, and error is less; The probable value of the prospect threshold value of asking for according to self-adaptation and background threshold, pixel is upgraded background model, has adopted the method based on probability to dynamically update background model, and dynamic property is better.

Accompanying drawing explanation

Lower and the invention will be further described in conjunction with the accompanying drawings and embodiments.

Fig. 1 is the flow chart of steps of a kind of moving target detecting method based on self-adaptive kernel Density estimating model of the present invention;

Fig. 2 is the process flow diagram of steps A of the present invention;

Fig. 3 is the process flow diagram of step B of the present invention;

Fig. 4 is the process flow diagram of step C of the present invention;

Fig. 5 is the process flow diagram of step D of the present invention;

Fig. 6 is the process flow diagram of step e of the present invention;

Fig. 7 is the probability histogram of bus stop of the present invention scene;

Fig. 8 is the present invention's histogram of difference corresponding with bus stop scene probability histogram.

Fig. 9 is the partial enlarged drawing of bus stop of the present invention scene histogram of difference.

Embodiment

With reference to Fig. 1, a kind of moving target detecting method based on self-adaptive kernel Density estimating model, comprising:

A. according to input video, build the background model based on Density Estimator, thereby obtain the probability density function of input video pixel and the width of kernel function;

B. according to the width of the probability density function of input video grey scale pixel value and kernel function, determine that the parameter probability valuing of sample point is interval, then travel through the probability of all pixel gray-scale values, thus formation probability distribution histogram;

C. probability distribution histogram is carried out to secondary linear interpolation and little threshold value difference, thereby obtain histogram of difference;

D. according to histogram of difference self-adaptation, ask for prospect threshold value and background threshold;

E. according to the probable value of input video pixel gray-scale value, prospect threshold value and background threshold, the background model building is upgraded.

With reference to Fig. 2, be further used as preferred embodiment, described steps A, it comprises:

A1. according to input video and gaussian kernel function, obtain the probability density function of t pixel i constantly, the expression formula of the probability density function P of described moment t pixel i (x (t) i) is:

$P\left(x{\left(t\right)}_i\right)=\frac{1}{N}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}\underset{m=1}{\overset{d}{\mathrm{\&Pi;}}}\frac{1}{\sqrt{2\mathrm{\&pi;}}{\mathrm{\&sigma;}}_m}{e}^{-\frac{{(x{\left(t\right)}_{\mathrm{im}}-x{\left(t\right)}_{\mathrm{im},j})}^2}{{{2\mathrm{\&sigma;}}_m}^2}},$

Wherein, N is the number of each pixel background sample, the dimension that d is pixel, the dimension that m is color space, the gray-scale value of m the Color Channel that x (t) im is i pixel, x (t) _{im, j}be m Color Channel gray-scale value of j sample point of i pixel, σ _mit is the cuclear density width of m Color Channel;

A2. according to the median of the absolute value of difference between consecutive frame gray-scale pixels sample, calculate the cuclear density width of gaussian kernel function, the expression formula of the cuclear density width cs of described gaussian kernel function is:

$\mathrm{\&sigma;}=\frac{\mathrm{median}\left(\right|x_{i,j}-x_{i,j}+1\left|\right)}{0.68\sqrt{2}},$

Wherein, median function is for asking median function, x (t) _{i, j}be the gray-scale value of j background sample in i pixel, x (t) _{i, j+1}it is the gray-scale value of j+1 background sample in i pixel.

With reference to Fig. 3, be further used as preferred embodiment, described step B, it comprises:

B1. according to the minimum possibility of the width of kernel function value σ _mindetermine the maximum probability possibility value at t moment N sample point, the expression formula of the described maximum probability possibility value Pmax at t moment N sample point is:

$P_{\max }=\frac{1}{{\left(\sqrt{2\mathrm{\&pi;}}{\mathrm{\&sigma;}}_{\min }\right)}^d};$

B2. the minimum possibility of the probability value of t moment N sample point is taken as to 0, and the probability at t moment N sample point is multiplied by scale factor, wherein α > 1, thereby the interval of the probability at t moment N sample point is quantified as to [0, α Pmax];

B3. according to the interval after quantizing, travel through the probability of all pixels, thus formation probability distribution histogram Hp, and the expression formula of described probability distribution histogram Hp is:

Wherein, Hp (j) is the histogrammic ordinate of probability distribution, and Hp (j) represents that j sample is at the number of current time probability same pixel point, and Hp ' (j) represents that j sample is at the number of previous moment probability same pixel point.

Probability at t moment N sample point is multiplied by scale factor, and object is to make histogram more clear and directly perceived.

With reference to Fig. 4, be further used as preferred embodiment, described step C, it comprises:

C1. adopt secondary linear interpolation method to carry out smoothly probability distribution histogram Hp, thus the histogram Hps after obtaining smoothly;

C2. adopt little threshold value to carry out difference to the histogram Hps after level and smooth, thereby obtain histogram of difference H _b, described histogram of difference H _bexpression formula be:

Wherein, H _b(j) be the ordinate of j sample in histogram of difference, b is adjacent two sample point minimum differences in the histogram after level and smooth; Min function is the function of minimizing, the number of probability same pixel point in j sample of histogram, a j+1 sample and j+2 sample after Hps (j), Hps (j+1) and Hps (j+2) represent respectively smoothly, Ts is default little threshold value.

With reference to Fig. 5, be further used as preferred embodiment, described step D, it comprises:

D1. by histogram of difference H _bfind the turning point P that histogram is delayed by abrupt change _z, described turning point P _zmeet:

$P_Z=\arg \underset{i}{\max }\left|\frac{H_b\left(i\right)-H_b(i+1)}{H_b(i+1)-H_b(i+2)}\right|;$

D2. at histogram of difference H _babove from Pz, start to find to the right the starting point P that variation tendency slows down _l;

D3. at histogram of difference H _bupper from P _lstart to find to the right the terminal P slowly changing _r, until the maximum probability point P of prospect _m;

D4. according to P _land P _rself-adaptation is calculated prospect threshold value and background threshold, described prospect threshold value T _fwith background threshold T _bcomputing formula be respectively:

T _f＝P _L-β(P _R-P _L)，

T _b＝P _R+β(P _R-P _L)，

Wherein, β is adaptive factor, and β < 1.

Turning point P _zit is the corresponding point of argument maximal value in histogram of difference.

Starting point P _lin histogram of difference, the derivative of function curve levels off to corresponding starting point while stablizing.

Terminal P _rit is corresponding some when the derivative of function curve starts to increase in histogram of difference

The maximum probability point P of prospect _mthat the derivative of function curve in histogram of difference levels off to corresponding point while stablizing again.

With reference to Fig. 6, be further used as preferred embodiment, described step e, it comprises:

E1. according to the renewal probability of the magnitude relationship calculating pixel point of the probable value of input video pixel gray-scale value and prospect threshold value, background threshold;

E2. according to the pixel calculating, upgrading probability upgrades the sample value of input video pixel according to first in first out.

Be further used as preferred embodiment, described step e 1, it is specially:

Probable value P (x (t) i) and prospect threshold value T according to input video pixel in the t moment _f, background threshold T _bthe renewal probability of magnitude relationship calculating pixel point, the renewal probability P of described pixel _bthe expression formula of (x (t) i) is:

Be further used as preferred embodiment, described step e 2, it is specially:

According to the pixel calculating, upgrade probability P _b(x (t) i) upgrades in t j sample value constantly input video pixel i according to first in first out, the sample value x after renewal (t) i, and the expression formula of j is:

x(t) _i，j＝P _b(x(t) _i)·(x(t) _i)+(1-P _b(x(t) _i))·(x(t-1) _i，j)，

Wherein, x (t-1) i, j is that input video pixel i is in t-1 j sample value constantly.

Below in conjunction with specific embodiment, the present invention is described in further detail

Embodiment mono-

The present embodiment is introduced the implementation procedure of a kind of moving target detecting method based on self-adaptive kernel Density estimating model of the present invention.

Moving target detecting method main process of the present invention comprises:

(1), background modeling

Background modeling method based on Density Estimator is according to the N of each pixel in an image sample, calculates the probability density of this grey scale pixel value by the method for non-parametric estmation.

Supposing has M pixel in input video frame, each pixel has N background sample, and in moment t frame of video, the gray-scale value of i pixel is x (t) i, and the gray-scale value of j the background sample that this pixel is corresponding is x (t) i, i, the probability of t pixel i is constantly:

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

Wherein K is kernel function, meets K (u) > 0, ∫ K (u) du=1.

General selection K is Gaussian function, Normal Distribution (0, σ ²), have:

$P\left(x{\left(t\right)}_i\right)=\frac{1}{N}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}\frac{1}{\sqrt{2\mathrm{\&pi;}}\mathrm{\&sigma;}}{e}^{-\frac{{(x{\left(t\right)}_i-x{\left(x\right)}_{i,j})}^2}{{{2\mathrm{\&sigma;}}_{}}^2}}---\left(2\right)$

Suppose that each Color Channel is separate, the probability density function of t pixel i is constantly:

$P\left(x{\left(t\right)}_i\right)=\frac{1}{N}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}\underset{m=1}{\overset{d}{\mathrm{\&Pi;}}}\frac{1}{\sqrt{2\mathrm{\&pi;}}{\mathrm{\&sigma;}}_m}{e}^{-\frac{{(x{\left(t\right)}_{\mathrm{im}}-x{\left(t\right)}_{\mathrm{im},j})}^2}{{{2\mathrm{\&sigma;}}_m}^2}}---\left(3\right)$

And the width cs of kernel function K can be calculated by the median of the absolute value of pixel consecutive frame gray scale sample difference, and then carry out linearization process, can make well sample distribution appropriateness smooth.Specifically can be tried to achieve by following formula (4) and (5):

m＝median(|x _i.j-x _i，j+1|) (4)

$\mathrm{\&sigma;}=\frac{m}{0.68\sqrt{2}}---\left(5\right)$

(2), probability histogram analysis

By step (), can estimate the wide σ of core, when the maximal value that core is wide may value σ _maxwith minimum possibility value σ _minafter determining, and make x=0, at t, the maximum probability possibility value Pmax of N sample point can be definite by following formula (6) constantly:

$P_{\max }=\frac{1}{{\left(\sqrt{2\mathrm{\&pi;}}{\mathrm{\&sigma;}}_{\min }\right)}^d}---\left(6\right)$

The minimum possibility of the probability value Pmin of N sample point, not only depends on σ _max, also the dimension d in formula (3) determines, its value can be taken as to 0, probability is multiplied by the interval that scale factor is quantified as probability and is quantified as [0, α Pmax],, then travel through the probability of all pixels, thus formation probability distribution histogram Hp.

Fig. 7 is the probability histogram of certain frame in the scene video of bus stop, and wherein U axle represents the probable value value of pixel, and V axle represents that probability equals the pixel number of this value.Local peaking's number in Fig. 7 has reflected the complicacy of background, and the size of peak value has reflected the otherness of the number of pixels that belongs to this value.This otherness has reflected the complexity of background, and background is more complicated, and difference is more remarkable.

(3), difference histogram analysis

In this step, first by secondary linear interpolation, Hp is smoothly obtained to Hps.Three adjacent pixels are carried out to difference according to formula (8), and further remove interference with little threshold value Ts, thereby obtain histogram of difference H _b, H _bexpression formula suc as formula shown in (9).

(8)

b＝min(|H _ps(j)-H _ps(j+1)|，|H _ps(j)-H _ps(j+2)|)

Fig. 8 is the histogram of difference that Fig. 7 is corresponding, and wherein transverse axis U represents the value of probability, and Z-axis V represents to belong to the number of the pixel of this probability.Histogram of difference in Fig. 8 has reflected the pixel number difference between two adjacent pixel probable values, and number difference is larger, more close to sport foreground; Otherwise, more close to background.

(4), self-adaptation prospect threshold value and background threshold asks for

Fig. 9 is the partial enlarged drawing of Fig. 8.Wherein transverse axis U represents the value of probability, and Z-axis V represents to belong to the number of the pixel of this probability.

The detailed process that the present invention asks for self-adaptation prospect threshold value and background threshold is:

A. by histogram of difference H _bfind the turning point P that histogram is delayed by abrupt change _z, P _zmeet following formula (10):

$P_Z=\arg \underset{i}{\max }\left|\frac{H_b\left(i\right)-H_b(i+1)}{H_b(i+1)-H_b(i+2)}\right|---\left(10\right)$

B. at histogram of difference H _babove from Pz, start to find to the right the starting point P that variation tendency slows down _l;

C. at histogram of difference H _bupper from P _lstart to find to the right the terminal P slowly changing _r, until the maximum probability point P of prospect _m;

D. as prospect threshold value T _fget P _ztime can miss part foreground target, when getting P _ltime can be prospect part background error detection, prospect threshold value T of the present invention for this reason _fcomputing formula be:

T _f＝P _L-β(P _R-P _L)

(11)

Wherein, introduce β (P _r-P _l) this is in order to reduce detection error.

E. choosing of adaptive background threshold value is the probability that prospect mistake is divided into background in order to reduce, thereby is conducive to rationally upgrade background.So background threshold T _bcomputing formula be:

T _b＝P _R+β(P _R-P _L)

(12)

(5), the renewal of background model

The pixel probability being obtained by formula (3), has in fact reflected the similarity degree of this pixel and background, so the foundation that probable value can be upgraded as a setting.Consideration is at moment t probability P (x (t) i), if obviously probability P (x (t) i) is less than prospect threshold value T _f, (x (t) i is prospect to pixel, now when context update, does not consider this pixel; If probability P (x (t) i) is greater than background threshold T _b, (x (t) i is background to pixel; Probability is between T _fand T _bbetween pixel, with new probability P _b(x (t) i) upgrades.Have:

And j sample value of i pixel upgraded according to the principle of first in first out in moment t background model, have:

x(t) _i.j＝P _b(x(t) _i)·(x(t) _i)+(1-P _b(x(t) _i))·(x(t-1) _i (14)

Wherein, P _b(x (t) i) reflected the updating ability of model to background.If P _b(x (t) i) is larger, and background detection is subject to the influence degree of new samples larger.

Employing formula (14) upgrade can: (1) is reliably upgraded belonging to the pixel of background; (2) pixel that error detection is background for belonging to prospect is upgraded with probability; (3) to belonging to the pixel of prospect, do not upgrade; Avoid foreground pixel point pollution background, and can reflect the slow change procedure of background.

Compared with prior art, the present invention asks for prospect threshold value and background threshold by histogram analysis self-adaptation, and upgrades according to prospect threshold value and the background threshold asked for, and method is simple, and calculated amount is little; The dual threshold method that has adopted prospect threshold value to combine with background threshold is carried out pixel classification, has overcome the error in classification of single threshold method, and error is less; The probable value of the prospect threshold value of asking for according to self-adaptation and background threshold, pixel is upgraded background model, has adopted the method based on probability to dynamically update background model, and dynamic property is better.

More than that better enforcement of the present invention is illustrated, but the invention is not limited to described embodiment, those of ordinary skill in the art also can make all equivalent variations or replacement under the prerequisite without prejudice to spirit of the present invention, and the distortion that these are equal to or replacement are all included in the application's claim limited range.

Claims (8)

1. the moving target detecting method based on self-adaptive kernel Density estimating model, is characterized in that: comprising:

A. according to input video, build the background model based on Density Estimator, thereby obtain the probability density function of input video pixel and the width of kernel function;

B. according to the width of the probability density function of input video grey scale pixel value and kernel function, determine that the parameter probability valuing of sample point is interval, then travel through the probability of all pixel gray-scale values, thus formation probability distribution histogram;

C. probability distribution histogram is carried out to secondary linear interpolation and little threshold value difference, thereby obtain histogram of difference;

D. according to histogram of difference self-adaptation, ask for prospect threshold value and background threshold;

E. according to the probable value of input video pixel gray-scale value, prospect threshold value and background threshold, the background model building is upgraded.

2. a kind of moving target detecting method based on self-adaptive kernel Density estimating model according to claim 1, is characterized in that: described steps A, and it comprises:

A1. according to input video and gaussian kernel function, obtain the probability density function of t pixel i constantly, the expression formula of the probability density function P of described moment t pixel i (x (t) i) is:

$P\left(x{\left(t\right)}_i\right)=\frac{1}{N}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}\underset{m=1}{\overset{d}{\mathrm{\&Pi;}}}\frac{1}{\sqrt{2\mathrm{\&pi;}}{\mathrm{\&sigma;}}_m}{e}^{-\frac{{(x{\left(t\right)}_{\mathrm{im}}-x{\left(t\right)}_{\mathrm{im},j})}^2}{{{2\mathrm{\&sigma;}}_m}^2}},$

Wherein, N is the number of each pixel background sample, the dimension that d is pixel, the dimension that m is color space, the gray-scale value of m the Color Channel that x (t) im is i pixel, x (t) _im, j is m Color Channel gray-scale value of j sample point of i pixel, σ _mit is the cuclear density width of m Color Channel;

A2. according to the median of the absolute value of difference between consecutive frame gray-scale pixels sample, calculate the cuclear density width of gaussian kernel function, the expression formula of the cuclear density width cs of described gaussian kernel function is:

$\mathrm{\&sigma;}=\frac{\mathrm{median}\left(\right|x_{i,j}-x_{i,j}+1\left|\right)}{0.68\sqrt{2}},$

Wherein, median function is for asking median function, x (t) _{i, j}be the gray-scale value of j background sample in i pixel, x (t) _{i, j}+ 1 is the gray-scale value of j+1 background sample in i pixel.

3. a kind of moving target detecting method based on self-adaptive kernel Density estimating model according to claim 2, is characterized in that: described step B, and it comprises:

B1. according to the width minimum of kernel function, may determine the maximum probability possibility value at t moment N sample point by value σ min, the expression formula of the described maximum probability possibility value Pmax at t moment N sample point is:

$P_{\max }=\frac{1}{{\left(\sqrt{2\mathrm{\&pi;}}{\mathrm{\&sigma;}}_{\min }\right)}^d};$

B2. the minimum possibility of the probability value of t moment N sample point is taken as to 0, and the probability at t moment N sample point is multiplied by scale factor, wherein α > 1, thereby the interval of the probability at t moment N sample point is quantified as to [0, α Pmax];

B3. according to the interval after quantizing, travel through the probability of all pixels, thus formation probability distribution histogram Hp, and the expression formula of described probability distribution histogram Hp is:

Wherein, Hp (j) is the histogrammic ordinate of probability distribution, and Hp (j) represents that j sample is at the number of current time probability same pixel point, and Hp ' (j) represents that j sample is at the number of previous moment probability same pixel point.

4. a kind of moving target detecting method based on self-adaptive kernel Density estimating model according to claim 3, is characterized in that: described step C, and it comprises:

C1. adopt secondary linear interpolation method to carry out smoothly probability distribution histogram Hp, thus the histogram Hps after obtaining smoothly;

C2. adopt little threshold value to carry out difference to the histogram Hps after level and smooth, thereby obtain histogram of difference H _b, described histogram of difference H _bexpression formula be:

Wherein, H _b(j) be the ordinate of j sample in histogram of difference, b is adjacent two sample point minimum differences in the histogram after level and smooth; Min function is the function of minimizing, the number of probability same pixel point in j sample of histogram, a j+1 sample and j+2 sample after Hps (j), Hps (j+1) and Hps (j+2) represent respectively smoothly, Ts is default little threshold value.

5. a kind of moving target detecting method based on self-adaptive kernel Density estimating model according to claim 4, is characterized in that: described step D, and it comprises:

D1. by histogram of difference H _bfind the turning point P that histogram is delayed by abrupt change _z, described turning point P _zmeet:

$P_Z=\arg \underset{i}{\max }\left|\frac{H_b\left(i\right)-H_b(i+1)}{H_b(i+1)-H_b(i+2)}\right|;$

D2. at histogram of difference H _babove from Pz, start to find to the right the starting point P that variation tendency slows down _l;

D3. at histogram of difference H _bupper from P _lstart to find to the right the terminal P slowly changing _r, until the maximum probability point P of prospect _m;

D4. according to P _land P _rself-adaptation is calculated prospect threshold value and background threshold, described prospect threshold value T _fwith background threshold T _bcomputing formula be respectively:

T _f＝P _L-β(P _R-P _L)，

T _b＝P _R+β(P _R-P _L)，

Wherein, β is adaptive factor, and β < 1.

6. a kind of moving target detecting method based on self-adaptive kernel Density estimating model according to claim 5, is characterized in that: described step e, and it comprises:

E1. according to the renewal probability of the magnitude relationship calculating pixel point of the probable value of input video pixel gray-scale value and prospect threshold value, background threshold;

E2. according to the pixel calculating, upgrading probability upgrades the sample value of input video pixel according to first in first out.

7. a kind of moving target detecting method based on self-adaptive kernel Density estimating model according to claim 6, is characterized in that: described step e 1, and it is specially:

Probable value P (x (t) i) and prospect threshold value T according to input video pixel in the t moment _f, background threshold T _bthe renewal probability of magnitude relationship calculating pixel point, the renewal probability P of described pixel _bthe expression formula of (x (t) i) is:

8. a kind of moving target detecting method based on self-adaptive kernel Density estimating model according to claim 7, is characterized in that: described step e 2, and it is specially:

According to the pixel calculating, upgrade probability P _b(x (t) i) upgrades in t j sample value constantly input video pixel i according to first in first out, the sample value x after renewal (t) i, and the expression formula of i is:

x(t) _i，j=P _b(x(t) _i)·(x(t) _i)+(1-P _b(x(t) _i))·(x(t-1) _i，j)，

Wherein, x (t-1) i, j is that input video pixel i is in t-1 j sample value constantly.