CN104599291A - Structural similarity and significance analysis based infrared motion target detection method - Google Patents

Abstract

A structural similarity and significance analysis based infrared motion target detection method comprises the steps of utilizing GBVS to conduct significance analysis on a source image so as to obtain a significance region, distinguishing the source image into a severe change region and a slight change region through improved SSIM and adopting different learning rates to optimize a Gaussian mixing model, finally utilizing the optimized Gaussian mixing model to detect a closed region, wherein the region overlapped with the significance region, of the closed region, is a final motion target of a current frame. The problem of edge blur of a Gaussian mixing algorithm is solved by means of the structural similarity and significance analysis based infrared motion target detection method, and the structural similarity and significance analysis based infrared motion target detection method has good self adaptability and detection effect.

Description

The infrared motion target detection method of structure based similarity and significance analysis

Technical field

What the present invention relates to is a kind of technology of image processing field, the specifically infrared motion target detection method of a kind of structure based similarity and significance analysis.

Background technology

It is that succeeding target is followed the tracks of that moving target detects in real time, and the prerequisite of identification, its effect directly affects robustness and the accuracy of follow-up work.Moving object detection refers to the foreground target extracting in the video sequence and there is relative motion with background, for the more senior motion analysiss such as target following are subsequently prepared, be an important research direction of computer vision field, be used widely in the field such as intelligent monitor system, man-machine interactive system.

Moving object detection is analyzed by the image sequence photographed imageing sensor, detects moving target, for more high-rise behavior understanding lays good basis.Moving object detection is the Focal point and difficult point of video sequence analysis.In the studying for a long period of time of moving object detection, people propose to comprise many classic algorithm such as method of difference, Background difference, optical flow method.Wherein: most widely used general with background subtraction, by background extraction model and compare frame difference detect moving target.The precision of background model determines the validity of background subtraction, if the variation of background modeling process occurrence scene, the situations such as imaging device vibration, seriously can reduce contrast and the signal to noise ratio (S/N ratio) of image, affect the identification to infrared target.

Based on the background modeling method of gauss hybrid models, rely on continuous print gaussian component modeling background information, recycling background information Difference test goes out moving target, can solve multi-modal background problems preferably, especially be applicable to outdoor light and the little and fireballing moving object detection of Changes in weather.But when gauss hybrid models initialization, when new model is set up and learning rate does not mate, all diplopia phenomenon can be produced,

Through finding the retrieval of prior art, open (bulletin) the day 2014.05.21 of Chinese patent literature CN103810703A, disclose a kind of tunnel video moving object detection method based on image procossing, the method comprises the following steps: set up initial back-ground model; Set up and dynamically update background model in real time; Structure local structure similarity measure function; Structure local gray level statistical measurement function; Motion target area is extracted according to local structure similarity measure function and local gray-scale statistical measure function.But this technology merely determines whether moving target, it is discrete for extracting moving target by algorithm, does not comprise the characteristic information of original object, cannot provide support for the follow-up further process to moving target.

Summary of the invention

The present invention is directed to the change that traditional Gauss mixture model can not detect background in time, detect by it infrared target obtained and comprise false profile, the deficiency not easily accurately identified, infrared motion target detection method and the system of a kind of structure based similarity and significance analysis are proposed, combine by watershed algorithm the spatial information characterizing target area discrete pixels point and obtain enclosed region, again by Based PC NN (Pulse Coupled Neural Network, Pulse Coupled Neural Network) partitioning algorithm eliminate diplopia, finally detect complete infrared motion target, thus foreground target can be extracted preferably, there is good effect and stronger robustness.

The present invention is achieved by the following technical solutions:

The present invention relates to the infrared motion target detection method of a kind of structure based similarity and significance analysis, first utilize GBVS (Graph ?Based Visual Saliency, saliency analytical algorithm based on graph theory) to source images significance analysis, obtain salient region; Again source images is divided into change intense regions and change shoulder by the SSIM (Structural similarity, structural similarity algorithm) improved, and adopt the method optimizing gauss hybrid models of different learning rate; Finally utilize the gauss hybrid models after optimizing to detect closed region, the overlapping region of this closed region and salient region is the final moving target of present frame.

Described salient region, utilizes GBVS to carry out significance analysis to infrared sequence image and obtains.

Described change intense regions and change shoulder, obtain in the following manner:

1) the SSIM algorithm improved is utilized to ask for the structural similarity of the local block of adjacent two frame figure in infrared sequence image;

2) adopt moving window mode to calculate the structural similarity statistical value of local block, and generate the statistical graph based on infrared sequence image on this basis;

3) utilize CDF mode counting statistics picture in picture as localized variation situation, the cumulative number Au of the some correspondence finding level and smooth rear distribution curve curvature maximum _max, be greater than Au by statistical graph _maxpoint form region divide into change intense regions, all the other for change shoulder.

Described optimization gauss mixture model refers to: adopt larger learning rate, the less learning rate of change shoulder employing for change intense regions.

The present invention relates to a kind of system realizing said method, comprise: salient region extraction module, structural similarity sort module, Gauss model update module and closed region detection module, wherein: image is divided into background area and target area by structural similarity sort module, Gauss model update module is connected with structural similarity sort module, different learning rate is adopted to upgrade background area and target area respectively, transmitting discrete moving target information, closed region detection module is connected with Gauss model Renewal model and transmits closed target information, salient region extraction module is connected with closed region detection module and transmits accurately moving target information.

Technique effect

Compared with prior art, the present invention is in model learning, for solving the replacement problem of background model, utilize structural similarity algorithm infrared image background is divided into graded obviously with region slowly, set different learning rates respectively and upgrade gauss hybrid models, ensure the Stability and veracity of model, with the position in quick obtaining infrared motion target region; In target detection, the watershed algorithm based on spatial information is utilized to obtain the enclosed region of target, vision significance (Graph ?Based Visual Salience, the GBVS) algorithm of recycling graphic based removes diplopia, finally obtains complete moving target.Experimental result shows, this method well solves the ill-defined problem of mixed Gaussian algorithm, has good adaptivity and Detection results.

Accompanying drawing explanation

Fig. 1 is cumulative distribution function figure;

Fig. 2 is schematic flow sheet of the present invention;

Fig. 3 is embodiment 1 effect schematic diagram;

In figure: (a), b () is respectively infrared sequence image the 18th, 19 frames, (c), d () is respectively the target area that constant learning rate and learning rate changing detect, (e) ?(h) be respectively the testing result obtained by frame difference method, background subtraction and the inventive method between watershed algorithm, neighbour.

Fig. 4 is for split by Fig. 3 (b) the accurate target region obtained by hand.

Embodiment

Elaborate to embodiments of the invention below, the present embodiment is implemented under premised on technical solution of the present invention, give detailed embodiment and concrete operating process, but protection scope of the present invention is not limited to following embodiment.

Embodiment 1

As shown in Figure 2, the present embodiment is A to infrared sequence image _t(t=1 ... n) following process is carried out:

The first step: utilize GBVS algorithm to infrared sequence image A _tcarry out significance analysis, obtain salient region B _t, concrete steps comprise:

1.1) defined nucleotide sequence image A _ttwo pixel m of characteristic pattern ₁(i ₁, j ₁) and m ₂(i ₂, j ₂) otherness:

$d(m_1,m_2)=\left|{\log }_2\frac{M_f\left(m_1\right)}{M_f\left(m_2\right)}\right|,$ Wherein: M _f() is the pixel in characteristic pattern.

1.2) construct the full figure connected after obtaining the otherness of any two points in image, in figure, each summit represents a pixel, and every bar limit represents the weight between two pixels, namely $w(m_1,m_2)=d(m_1,m_2)\·\exp (-\frac{{(i_1-i_2)}^2+{(j_1-j_2)}^2}{2{\mathrm{\σ}}^2}),$ Wherein: σ is scale factor, m ₁(i ₁, j ₁) and m ₂(i ₂, j ₂) be two pixels;

1.3) after all weights are normalized, salient region is obtained.

Second step: utilize the SSIM algorithm improved to ask for the structural similarity of the local block of adjacent two frame figure in infrared sequence image, concrete steps comprise:

2.1) moving window of entire image by 7 × 7 pixel sizes is pressed from top to bottom, order is divided into some overlapping sub-blocks from left to right, moving window then carries out mirror-extended to image when exceeding image range and obtains 7 × 7 sub-blocks, if when image is of a size of m × n, then the sub-block number obtained is m × n;

2.2) from the structural similarity of the sub-block Block (x, y) of same position in image F and image L be:

$\mathrm{SSIM}\left(\mathrm{Block}(x,y)\right)=\frac{(2u_F{u}_L+C_1)(2{\mathrm{\σ}}_F{\mathrm{\σ}}_L+C_2)}{(u_F^2+u_L^2+C_1)({\mathrm{\σ}}_F^2+{\mathrm{\σ}}_L^2+C_2)}---\left(1\right)$

Wherein: (x, y) represents sub-block center pixel, u _f, u _l, σ _f, σ _lrepresent average and the variance of the sub-block Block (x, y) of image F and image L respectively; C ₁, C ₂be respectively the normal amount being tending towards 0.

3rd step: adopt moving window mode to calculate the structural similarity statistical value A of local block _t' (x, y), and generate the statistical graph based on infrared sequence image on this basis:

D _t(x,y)＝A _t'(x,y)-A _t(x,y)，

${A_t}^{\&prime;}(x,y)=\left\{\begin{array}{ccc}{A}_t(x,y),& \mathrm{if}& {\mathrm{SSIM}}_{t,j}\left(\mathrm{Block}(x,y)\right)\&GreaterEqual;0.9\\ A_t(x,y)+\mathrm{Ones}(x,y),& \mathrm{if}& {\mathrm{SSIM}}_{t,j}\left(\mathrm{Block}(x,y)\right)<0.9\end{array}\right.---\left(2\right)$

Wherein: Ones (x, y) is all 1's matrix of 7 × 7 sizes centered by (x, y), when local block structural similarity higher than 0.9 time, think that its intensity of variation is little, A _t(x, y) is constant, and when the structural similarity of sub-block is less than 0.9, centered by (x, y), the statistical value of matrix of 7 × 7 all adds 1.

Described statistical graph D _tthis regional change degree Shaoxing opera of region representation that in (x, y), gray-scale value is larger is strong.

4th step: utilize CDF (s) (Cumulative Distribution Function, cumulative distribution function) mode counting statistics figure D _timage local situation of change in (x, y), the cumulative number Au of the some correspondence finding level and smooth rear distribution curve curvature maximum _max, by statistical graph D _tau is greater than in (x, y) _maxpoint form region divide into change intense regions A _tr, all the other are change shoulder A _tg.

Described cumulative distribution function is the function of the localized variation degree of image, namely asks image D _tthe probability P (s) that in (x, y), pixel value s occurs, cumulative distribution function CDF (s)=P (S≤s) obtains smooth curve function through 10 difference matchings, as shown in Figure 1; Obtain D corresponding to the maximum point of this curvature of curve _tthe value Au of (x, y) _max, by D _tin (x, y), pixel value is greater than Au _maxsome composition region be change intense regions.

5th step: by infrared sequence image A _t(t=1 ... n) gauss hybrid models is upgraded by the mode of the corresponding learning rate of zones of different, namely gauss hybrid models upgrades weight, average, variance according to the image information of a new frame in learning process, and change intense regions upgrades with 10 times of learning rates, change shoulder keeps original learning rate α, and concrete steps comprise:

5.1) pixel representing moving target is calculated:

$\left\{\begin{array}{c}P\left(S_t(x,y)\right)=\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_{i,t}(x,y)\mathrm{\&eta;}(S_t(x,y),u_{i,t},{\mathrm{\&Sigma;}}_{i,t})\\ \mathrm{\&eta;}(X_t,{\mathrm{\&mu;}}_{i,t},{\mathrm{\&Sigma;}}_{i,t})=\frac{1}{{\left(2\mathrm{\&pi;}\right)}^{\frac{n}{2}}{\left|{\mathrm{\&Sigma;}}_{i,t}\right|}^{\frac{1}{2}}}{e}^{-\frac{1}{2}{(X_t-u_{i,t})}^T{\mathrm{\&Sigma;}}_{i,t}^{-1}(X_t-u_{i,t})}\\ {\mathrm{\&Sigma;}}_{i,t}={\mathrm{\&sigma;}}_{i,t}^{2}I\end{array}\right.---\left(\text{3}\right)$

Wherein: P (S _t) be S _tthe distribution function of (x, y), S _t(x, y) represents the pixel value of t width image, ω _i,t(x, y) represents that t width image slices vegetarian refreshments belongs to the weights of i-th Gaussian distribution, u _i,tfor Gauss model average, ∑ _i,tfor the covariance of Gauss model, σ _i,tfor standard deviation, I is unit matrix, and n is X _tdimension, the probability density function that η () is Gauss model.

5.2) following renewal is carried out to Gauss model:

$\left\{\begin{array}{c}{\mathrm{\&omega;}}_t=(1-\mathrm{\&alpha;}){\mathrm{\&omega;}}_{t-1}+\mathrm{\&alpha;}\left(M_t\right)\\ u_t=(1-\mathrm{\&rho;})u_{t-1}+\mathrm{\&rho;}{X}_t\\ {\mathrm{\&sigma;}}_t^2=(1-\mathrm{\&rho;}){\mathrm{\&sigma;}}_{t-1}^2+\mathrm{\&rho;}{(X_t-u_t)}^T(X_t-u_t)\end{array}\right.---\left(4\right)$

Wherein: α is learning rate, represent that moving target incorporates the speed of background, by pixel value S _t(x, y) mates with K Gauss model, and works as S _t(x, y) is less than 2.5 times of standard deviation sigma with the difference of i-th Gauss model _i,tin time, is thought and to mate with "current" model, i.e. M _tcoupling is represented, the M when not mating when=1 _t=0, and corresponding reduction weight and average and variance not being upgraded.

5.3) due to S _tthe Gauss model that (x, y) mates the most has maximum weights and minimum standard deviation, by K Gauss model according to ω _i,t/ σ _i,tratio descending sort, be in sequence top Gauss model most possibly Steady Background Light is described, and be in sequence bottom Gauss model Describing Motion target, therefore:

By Num before in K Gauss model model as a setting, τ represents the minimum scale that background Gauss model is shared in probability distribution, the i.e. number of background model $\mathrm{Num}=\arg \min (\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_{i,t}>\mathrm{\&tau;})---\left(5\right)$

Work as S _t(x, y) does not mate with K Gauss model, then use average to be x coming last Gauss model _t, standard deviation sigma _i,twith weights ω _i,tthe Gauss model being set to initial value respectively replaces; After each renewal completes, to weights ω _i,tbe normalized.

6th step: using the gray-scale value of each pixel as level height value, calculates P (S _t) spatial frequency SF _tobtain the edge of dispersive target, then adopt watershed algorithm connection edge to obtain closed region E _t.

Described spatial frequency refers to: the index of grey scale change severe degree in token image, i.e. the gradient of gray scale on plane space.

Described watershed algorithm is realized by following iterative manner:

$\left\{\begin{array}{c}X\left(h_{\min }\right)=\{p\&Element;D|I\left(p\right)=h_{\min }\}=T\left(h_{\min }\right)\\ X(h+1)=\mathrm{MIN}(h+1)\&cup;Z_{T(h+1)}\left(X\left(h\right)\right),h\&Element;[h_{\min },h_{\max }]\end{array}\right.,$ Wherein: X (h) is regional union of sets when level value is h, namely in image, gray-scale value is less than the set of the point of h, and I (p) represents the gray-scale value of image, h _minand h _maxrepresent that image is minimum with the highest gray-scale value; T (h _min) corresponding connected domain when representing that gray-scale value is minimum, MIN (h+1) is the associating of gray-scale value all Minimum Areas when h+1, Z _{t (h+1)}(X (h)) represents region X (h) the geodetic range of influence in connected domain T (h+1).

Described closed region $E_t=\left\{x\right|x\&Element;\mathrm{Wshed},\mathrm{and\; x}\&NotElement;{\mathrm{Wshed}}_{\max }\},$ Wshed _maxfor the enclosed region that area in image is maximum, Wshed (f)=D-X (h _max) namely at image D _tx (h in (x, y) _max) supplementary set.

Described geodetic range of influence refers to: when be divided into k the region B be connected _i, i=1 ..., k, then in A, subset B _igeodetic range of influence be defined as: { p ∈ A|=d _a(p, B _i) < d _a(p, B B _i), geodetic range of influence is called reception basin, and the border of reception basin then forms watershed divide.

7th step: by the gauss hybrid models after renewal in the 5th step to the closed region E in the 6th step _tdetect, the salient region B obtained in gained region and the first step _tlap region be the final moving target of present frame.

The result complete and accurate that the target area that this method obtains obtains than additive method, as shown in Figure 3, shown in being compared as follows with the standard picture shown in Fig. 4, wherein MI represents related coefficient, QAB/F represents edge gradient information, the moving target that the larger expression of value obtains and former target more close.

	MI	QAB/F
			This method	0.2577	0.5922
Background subtraction	0.2459	0.4003
			Method of difference between neighbour	0.0854	0.1374

Claims (10)

1. an infrared motion target detection method for structure based similarity and significance analysis, is characterized in that, first utilizes GBVS to source images significance analysis, obtains salient region; Again source images is divided into change intense regions and change shoulder by the SSIM improved, and adopt the method optimizing gauss hybrid models of different learning rate; Finally utilize the gauss hybrid models after optimizing to detect closed region, the overlapping region of this closed region and salient region is the final moving target of present frame.

2. method according to claim 1, is characterized in that, described salient region, utilizes GBVS algorithm to infrared sequence image A _tcarry out significance analysis to obtain, concrete steps comprise:

1.1) defined nucleotide sequence image A _ttwo pixel m of characteristic pattern ₁(i ₁, j ₁) and m ₂(i ₂, j ₂) otherness: wherein: M _f() is the pixel in characteristic pattern;

1.2) construct the full figure connected after obtaining the otherness of any two points in image, in figure, each summit represents a pixel, and every bar limit represents the weight between two pixels, namely $w(m_1,m_2)=d(m_1,m_2)\&CenterDot;\exp (-\frac{{(i_1-i_2)}^2+{(j_1-j_2)}^2}{{2\mathrm{\&sigma;}}^2}),$ Wherein: σ is scale factor, m ₁(i ₁, j ₁) and m ₂(i ₂, j ₂) be two pixels;

1.3) after all weights are normalized, salient region is obtained.

3. method according to claim 1, is characterized in that, described change intense regions and change shoulder, obtain in the following manner:

1) the SSIM algorithm improved is utilized to ask for the structural similarity of the local block of adjacent two frame figure in infrared sequence image;

2) moving window mode is adopted to calculate the structural similarity statistical value A of local block _t' (x, y), and generate the statistical graph D based on infrared sequence image on this basis _t(x, y);

3) CDF mode counting statistics figure D is utilized _timage local situation of change in (x, y), the cumulative number Au of the some correspondence finding level and smooth rear distribution curve curvature maximum _max, by statistical graph D _tau is greater than in (x, y) _maxpoint form region divide into change intense regions A _tr, all the other are change shoulder A _tg.

4. method according to claim 3, is characterized in that, the structural similarity of described local block, obtains in the following manner:

I) moving window of entire image by 7 × 7 pixel sizes is pressed from top to bottom, order is divided into some overlapping sub-blocks from left to right, moving window then carries out mirror-extended to image when exceeding image range and obtains 7 × 7 sub-blocks, if when image is of a size of m × n, then the sub-block number obtained is m × n;

Ii) from the structural similarity of the sub-block Block (x, y) of same position in image F and image L be:

$\mathrm{SSIM}\left(\mathrm{Block}(x,y)\right)=\frac{({2u}_F{u}_L+C_1)({2\mathrm{\&sigma;}}_F{\mathrm{\&sigma;}}_L+C_2)}{(u_F^2+u_L^2+C_1)({\mathrm{\&sigma;}}_F^2+{\mathrm{\&sigma;}}_L^2+C_2)},$ Wherein: (x, y) represents sub-block center pixel, u _f, u _l, σ _f, σ _lrepresent average and the variance of the sub-block Block (x, y) of image F and image L respectively; C ₁, C ₂be respectively the normal amount being tending towards 0.

5. method according to claim 4, is characterized in that, the described statistical graph based on infrared sequence image, adopts moving window mode to calculate the structural similarity statistical value A of local block _t' (x, y), and generate D on this basis _t(x, y)=A _t' (x, y)-A _t(x, y),

${A_t}^{\&prime;}(x,y)=\left\{\begin{array}{c}{A}_t(x,y),\mathrm{if}{\mathrm{SSIM}}_{t,j}\left(\mathrm{Block}(x,y)\right)\&GreaterEqual;0.9\\ A_t(x,y)+\mathrm{Ones}(x,y),\mathrm{if}{\mathrm{SSIM}}_{t,j}\left(\mathrm{Block}(x,y)\right)<0.9\end{array}\right.,$ Wherein: Ones (x, y) is 1 matrix of 7 × 7 sizes centered by (x, y), when local block structural similarity higher than 0.9 time, think that its intensity of variation is little, A _t(x, y) is constant, and when the structural similarity of sub-block is less than 0.9, centered by (x, y), the statistical value of matrix of 7 × 7 all adds 1.

6. method according to claim 3, is characterized in that, described image local situation of change, and by calculating cumulative distribution function, namely the function of the localized variation degree of image obtains: counting statistics figure D _tthe probability P (s) that in (x, y), pixel value occurs, cumulative distribution function CDF (s)=P (S≤s) obtains smooth curve function through 10 difference matchings, D corresponding to the point that this curvature of curve is maximum _tthe value Au of (x, y) _max, by D _tin (x, y), pixel value is greater than Au _maxsome composition region be change intense regions.

7. method according to claim 2, it is characterized in that, described optimization gauss mixture model refers to: infrared sequence image is upgraded gauss hybrid models by the mode of the corresponding learning rate of zones of different, namely gauss hybrid models upgrades weight, average, variance according to the image information of a new frame in learning process, and change intense regions upgrades with 10 times of learning rates, change shoulder keeps original learning rate α.

8. the method according to claim 2 or 7, is characterized in that, described optimization gauss mixture model, specifically comprises:

I) pixel representing moving target is calculated: $\left\{\begin{array}{c}P\left(S_t(x,y)\right)=\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_{i,t}(x,y)\mathrm{\&eta;}(S_t(x,y),u_{i,t},{\mathrm{\&Sigma;}}_{i,t})\\ \mathrm{\&eta;}(X_i,{\mathrm{\&mu;}}_{i,t},{\mathrm{\&Sigma;}}_{i,t})=\frac{1}{{\left(2\mathrm{\&pi;}\right)}^{\frac{n}{2}}{\left|{\mathrm{\&Sigma;}}_{i,t}\right|}^{\frac{1}{2}}}{e}^{-\frac{1}{2}{(X_T-u_{i,t})}^T{\mathrm{\&Sigma;}}_{i,t}^{-1}(X_t-u_{i,t})}\\ {\mathrm{\&Sigma;}}_{i,t}={\mathrm{\&sigma;}}_{i,t}^{2}I\end{array}\right.$ Wherein: P (S _t) be S _tthe distribution function of (x, y), S _t(x, y) represents the pixel value of t width image, ω _i,t(x, y) represents that t width image slices vegetarian refreshments belongs to the weights of i-th Gaussian distribution, u _i,tfor Gauss model average, ∑ _i,tfor the covariance of Gauss model, σ _i,tfor standard deviation, I is unit matrix, and n is X _tdimension, the probability density function that η () is Gauss model;

Ii) following renewal is carried out to Gauss model: $\left\{\begin{array}{c}{\mathrm{\&omega;}}_t=(1-\mathrm{\&alpha;}){\mathrm{\&omega;}}_{t-1}+\mathrm{\&alpha;}\left(M_t\right)\\ u_t=(1-\mathrm{\&rho;})u_{t-1}+\mathrm{\&rho;}{X}_t\\ {\mathrm{\&sigma;}}_t^2=(1-\mathrm{\&rho;}){\mathrm{\&sigma;}}_{t-1}^2+\mathrm{\&rho;}{(X_t-u_t)}^T(X_t-u_t)\end{array}\right.,$ Wherein: α is learning rate, represent that moving target incorporates the speed of background, by pixel value S _t(x, y) mates with K Gauss model, and works as S _t(x, y) is less than 2.5 times of standard deviation sigma with the difference of i-th Gauss model _i,tin time, is thought and to mate with "current" model, i.e. M _tcoupling is represented, the M when not mating when=1 _t=0, and corresponding reduction weight and average and variance not being upgraded;

Iii) due to S _tthe Gauss model that (x, y) mates the most has maximum weights and minimum standard deviation, by K Gauss model according to ω _i,t/ σ _i,tratio descending sort, the Gauss model being in sequence top most possibly describes Steady Background Light, and be in the Gauss model Describing Motion target of sequence bottom, by Num before in K Gauss model model as a setting, τ represents the minimum scale that background Gauss model is shared in probability distribution, the i.e. number of background model

Iv) S is worked as _t(x, y) does not mate with K Gauss model, then use average to be x coming last Gauss model _t, standard deviation sigma _i,twith weights ω _i,tthe Gauss model being set to initial value respectively replaces; After each renewal completes, to weights ω _i,tbe normalized;

Iv) using the gray-scale value of each pixel as level height value, P (S is calculated _t) spatial frequency SF _tobtain the edge of dispersive target, then adopt watershed algorithm connection edge to obtain closed region E _t.

9. method according to claim 8, is characterized in that, described closed region

E _t=x|x ∈ Wshed, wshed _maxfor the enclosed region that area in image is maximum, i.e. watershed region, Wshed (f)=D _t(x, y)-X (h _max), wherein: image D _tin (x, y), X (h _max) for level value be h _maxtime regional union of sets, namely in image, gray-scale value is less than h _maxthe set of point;

Described watershed divide is formed by the border of geodetic range of influence, obtains especially by watershed algorithm iterative computation:

$\left\{\begin{array}{c}X\left(h_{\min }\right)=\{p\&Element;D|I\left(p\right)=h_{\min }\}=T\left(h_{\min }\right)\\ X(h+1)=\mathrm{MIN}(h+1)\&cup;Z_{T(h+1)}\left(X\left(h\right)\right),h\&Element;[h_{\min },h_{\max }]\end{array}\right.,$ Wherein: X (h) is regional union of sets when level value is h, namely in image, gray-scale value is less than the set of the point of h, and I (p) represents the gray-scale value of image, h _minand h _maxrepresent that image is minimum with the highest gray-scale value; T (h _min) corresponding connected domain when representing that gray-scale value is minimum, MIN (h+1) is the associating of gray-scale value all Minimum Areas when h+1, Z _{t (h+1)}(X (h)) represents region X (h) the geodetic range of influence in connected domain T (h+1).

10. one kind realizes the system of method described in above-mentioned arbitrary claim, it is characterized in that, comprise: salient region extraction module, structural similarity sort module, Gauss model update module and closed region detection module, wherein: image is divided into background area and target area by structural similarity sort module, Gauss model update module is connected with structural similarity sort module, different learning rate is adopted to upgrade background area and target area respectively, transmitting discrete moving target information, closed region detection module is connected with Gauss model Renewal model and transmits closed target information, salient region extraction module is connected with closed region detection module and transmits accurately moving target information.