CN102254394A - Antitheft monitoring method for poles and towers in power transmission line based on video difference analysis - Google Patents

Abstract

The invention discloses an antitheft monitoring method for poles and towers in a power transmission line based on video difference analysis. The method comprises the following steps of: acquiring power transmission line video signals by a camera, and transmitting the acquired power transmission line video signals to a monitoring center through a video server; intercepting digital images of the power transmission line needed to be monitored from a video stream by the monitoring center to get a monitored target image; performing pretreatment by a neighborhood averaging method; modeling via a Gaussian mixture method to generate a background image; dividing a moving target in the current frame through a background difference method to implement target detection; identifying the moving target via a method of a support vector machine to obtain a main monitoring target object; and tracking the target object to determine whether the target object has burglary behaviors. The monitoring method disclosed by the invention determines whether somebody wants to steal electric wires, tower materials and the like by detecting, identifying and tracking human bodies, non-human objects and the like, and gives an alarm in time; and devices used in the monitoring method are simple and economic.

Description

A kind of electric power line pole tower anti-thefting monitoring method based on the video variance analysis

Technical field

The invention belongs to video image and handle and transmission line of electricity on-line monitoring technique field, relate to a kind of electric power line pole tower anti-thefting monitoring method, be specifically related to a kind of electric power line pole tower anti-thefting monitoring method based on the video variance analysis.

Background technology

After domestic most of rural area and backcountry switched in succession, shaft tower was in the open air mostly, lacked special messenger's nurse, gave some lawless person's opportunities.Stolen cases such as the round steel, angle steel under the copper core, ground wire, ground lead in aerial high-voltage electric wire and the outdoor power transformer have taken place in a lot of areas.As the material base of electric power industry development, high-voltage power line frequently stolen brings heavy economic losses for power department and users.In order effectively to frighten criminal, stop to steal taking place frequently of electric power facility, just must develop effective transmission line of electricity guard against theft and alarm system.In recent years, many novel shaft tower guard against theft and alarm systems have been researched and developed both at home and abroad, as microwave induced formula burglary-resisting system, based on the acceleration transducer burglary-resisting system, based on the burglary-resisting system of vibration transducer and radar detedtor, induction type burglary-resisting system etc., the system of these systems constitutes similar substantially, and total system is by detecting extension set, Surveillance center, patrolling and examining composition of personnel.A detection extension set is installed on each basic shaft tower, detecting extension set is made up of front end sensors part and single-chip microcomputer or DSP processing section, the status information of mobile object or collection shaft tower are received the information of vibration around the detection shaft tower, use sound and light alarm to warn criminal in case of necessity.Center host monitoring software is in the background work pattern, when receiving the information that certain basic shaft tower sends, activates monitoring software, and the monitor staff determines stolen information by understanding short message content, in time informs the personnel of patrolling and examining.And the method needs a lot of sensors, complex structure, and it is cumbersome that system installs, and processing procedure is more complicated also.

Summary of the invention

The purpose of this invention is to provide a kind of electric power line pole tower anti-thefting monitoring method, solved the problem of existing shaft tower guard against theft and alarm system complex structure, installation trouble, complex disposal process based on the video variance analysis.

The technical solution adopted in the present invention is, a kind of electric power line pole tower anti-thefting monitoring method based on the video variance analysis is specifically implemented according to following steps:

Step 1:, the transmission line of electricity vision signal that collects is sent to Surveillance center by video server by camera collection transmission line of electricity vision signal; Surveillance center intercepts the digital picture of the transmission line of electricity that needs supervision, the target image that obtains monitoring from video flowing;

Step 2: the target image of the supervision that step 1 is obtained adopts neighborhood averaging to carry out pre-service;

Step 3: the pretreated image that step 2 is obtained adopts the mixed Gaussian method to carry out modeling, the generation background image;

Step 4: after treating the background image generation of step 3, adopt the background subtraction point-score that the moving target in the present frame is cut apart, realize target detection;

Step 5: the image after the target detection that the method that adopts support vector machine obtains step 4 carries out moving target identification, obtains main monitoring objective object;

Step 6: the image after the Target Recognition that step 5 is obtained carries out target following, determines the destination object behavior of whether committing theft.

Characteristics of the present invention also are,

Wherein the neighborhood averaging in the step 2 is carried out pre-service, specifically implements according to following steps:

At image space, suppose original image f that a secondary N * N pixel is arranged (x, y), the operation of removing each the pixel point value in the alternative image with the mean value of several pixels in the neighborhood, through obtain after the smoothing processing sub-picture g (x, y):

$g(x,y)=\frac{1}{M}\underset{(m,n)\∈S}{\mathrm{\Σ}}f(m,n),$

In the formula: x, y=0,1,2 ..., N-1; S is that (x y) puts the set of the coordinate of field mid point, but wherein do not comprise (x, y) point; M is the sum of set internal coordinate point.

Wherein adopt the mixed Gaussian method to carry out modeling in the step 3, specifically implement according to following steps:

The gray-scale value of each pixel is described with K Gaussian distribution, and the K value gets 3～5, and the size of K value depends on that calculator memory reaches the rate request to algorithm, the K value is big more, the ability of handling grey scale change is strong more, and the corresponding required processing time is also just long more, definition pixel gray-scale value variable X _tExpression, its probability density function can be represented with following K three-dimensional Gaussian function:

$P(X_t=\underset{i=1}{\overset{k}{\mathrm{\Σ}}}{\mathrm{\ω}}_{i,t}\·\mathrm{\η}(X_t,{\mathrm{\μ}}_{i,t},{\mathrm{\Σ}}_{i,t}),$

ω in the formula _{I, t}Be i Gaussian distribution in t weight constantly, and have

η (X _t, μ _{I, t}, ∑ _{I, t}) be t i Gaussian distribution constantly, its average is μ _{I, t}, covariance is a ∑ _{I, t},

$\mathrm{\η}(X_t,{\mathrm{\μ}}_{i,t},{\mathrm{\Σ}}_{i,t})=\frac{1}{{\left(2\mathrm{\π}\right)}^{\frac{n}{2}}{\left|{\mathrm{\Σ}}_{i,t}\right|}^{1/2}}{e}^{-\frac{1}{2}{(X_t-{\mathrm{\μ}}_{i,t})}^T{\mathrm{\Σ}}_{i,t}^{-1}(X_t,{\mathrm{\μ}}_{i,t})},$

In the formula, i=1 ..., K; N represents X _tDimension, R, G, three passages of B are separate, and identical variance is arranged, and then have

The expression variance, I representation unit battle array.

Wherein adopt the background subtraction point-score that the moving target in the present frame is cut apart in the step 4, realize target detection, specifically implement according to following steps:

At first establish B _kBe background image, f _kBe current frame image, difference image is D _k, D then _k(x, y)=| f _x(x, y)-B _K-1(x, y) |, establish R _kBe bianry image after the difference; To R _kCarry out connectivity analysis, when the area in the zone of a certain connection greater than certain threshold value, think that then the target that detects occurs, and think that the zone of this connection is exactly detected target image,

The binary-state threshold of T for setting.

Image after the target detection that the method that wherein adopts support vector machine in the step 5 obtains step 4 carries out moving target identification, specifically implements according to following steps:

A. linear separability situation:

The general type g of space neutral line discriminant function (x, y)=ω ^TX+b, the classifying face equation is ω ^TX+b=0 carries out normalization with discriminant function, and all samples of two classes are all satisfied | g (x) | 〉=1, this moment is from the nearest sample of classifying face | g (x) |=1, and require classifying face to all samples can both be correct classification, require it to satisfy exactly:

y _i(ω ^Tx+b)-1≥0，i＝1，2，...，n，

Be that those samples that equal sign is set up are support vector in the formula, the gap size in the classification space of two class samples:

M?arg?in＝2/||w||，

Therefore, optimal classification face problem can be expressed as following about fasciculation problem, promptly at condition y _i(ω ^TX+b)-1 〉=0, i=1,2 ..., under the constraint of n, ask function

Minimum value, the Lagrange function that is defined as follows:

$L(w,b,\mathrm{\α})=\frac{1}{2}{w}^{T}w-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i[y_i(w^T{x}_i+b)-1],$

Wherein, α _i" 0 be the Lagrange coefficient, following formula respectively to w, b, α _iAsk partial differential and make them equal 0:

$\frac{\∂L}{\∂w}=0\⇒w=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i{y}_i{x}_i,$

$\frac{\∂L}{\∂b}=0\⇒\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i{y}_i=0,$

$\frac{\∂L}{{\∂\mathrm{\α}}_i}=0\⇒{\mathrm{\α}}_i[y_i(w^T{x}_i+b)-1]=0,$

More than three formulas add that former constraint condition can change into former problem the dual problem of convex quadratic programming

$\max \underset{i=1}{\overset{n}{\mathrm{\Σ}}}{a}_i-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i{\mathrm{\α}}_j{y}_i{y}_j\left(x_i^T{x}_j\right),$

A wherein _i" 0, i=1 ..., n,

If Be optimum solution, then

$w^{*}=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{a}_i^{*}{y}_i{x}_i,$

Non-vanishing sample is support vector, and therefore, the overall coefficient vector of optimal classification face is the linear combination of support vector, b ^*By constraint condition α _i[y _i(w ^Tx _i+ b)-1]=0 to find the solution, the optimal classification function of trying to achieve thus is:

$f\left(x\right)=\mathrm{sgn}({\left(w^{*}\right)}^{T}x+b^{*})=\mathrm{sgn}(\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{a}_i^{*}{y}_i{x}_i^{*}x+b^{*}),$

Sgn () is a sign function;

B. the non-linear situation of dividing:

In the time can not separating 2 class points fully, introduce slack variable ξ with a lineoid _i(ξ _i〉=0, i=1 n), makes lineoid ω ^TX+b=0 satisfies:

y _i(ω ^Tx _i+b)≥1-ξ _i，

As 0＜ξ _i＜1 o'clock sample point x _iStill correctly classified, and worked as ξ _i〉=1 o'clock sample point x _iDivided by wrong, introduce following objective function:

$\mathrm{\ψ}(w,\mathrm{\ξ})=\frac{1}{2}{w}^{T}w+C\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\ξ}}_i,$

Wherein C is a positive constant, is called penalty factor, and this moment, SVM realized by quadratic programming:

$\max \underset{i=1}{\overset{n}{\mathrm{\Σ}}}{a}_i-\frac{1}{2}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i{\mathrm{\α}}_j{y}_i{y}_j\left(x_i^T{x}_j\right),$

0≤a wherein _i≤ C, i=1 ..., n, $\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{a}_i{y}_i=0.$

Wherein the image after the Target Recognition that in the step 6 step 5 is obtained carries out target following, specifically implements according to following steps:

1) in video sequence image, detects new moving target and moving region;

2) moving target of new detection is cut apart;

3) extract the feature of new moving target, and set up the object matching template;

4) with the forecast model target of prediction in next position that constantly may occur, determine next hunting zone of target constantly;

5) To Template with previous moment carries out match search in the hunting zone of prediction, seeks best match position;

6) utilize the target image that matches, revise the template data of tracked target, alternately repeat said process.

The invention has the beneficial effects as follows, in to human body around the electric power line pole tower and non-human body movement monitoring on-line monitoring process, determine whether that by detection, identification, tracking the someone wants to implement theft electric wire, tower material etc., and in time report to the police, and equipment is simple, economical human body and non-human body etc.

Description of drawings

Fig. 1 is certain original image of physical activity around shaft tower constantly in the inventive method;

Fig. 2 handles the background image that the back generates through mixed Gauss model in the inventive method;

Fig. 3 is the principle of work of background subtraction point-score in the inventive method;

Fig. 4 is based on the architecture of the image classification of SVM in the inventive method;

Fig. 5 is the target image through monitoring behind background difference and the image classification in the inventive method;

Fig. 6 is the process flow diagram of motion target tracking in the inventive method;

Fig. 7 is one group of result figure of target following in the inventive method;

Fig. 8 is the process flow diagram of the inventive method.

Embodiment

The present invention is described in detail below in conjunction with the drawings and specific embodiments.

The present invention is based on the electric power line pole tower anti-thefting monitoring method of video variance analysis, as shown in Figure 8, specifically implement according to following steps:

Step 1: by camera collection transmission line of electricity vision signal, it is sent back Surveillance center by transmission channel in real time in the mode of video flowing through video server; Surveillance center intercepts the digital picture of the transmission line of electricity that is monitored from video flowing, the target image that obtains monitoring;

Step 2: the target image of the supervision that step 1 is obtained adopts neighborhood averaging to carry out pre-service to improve picture quality, for the succeeding target recognition and tracking is provided convenience.Specifically implement according to following steps:

At first the video image to input carries out simple time or spatial filtering, comprise level and smooth, the enhancing, recovery of image etc., purpose is to suppress unwanted distortion, noise or strengthens the characteristics of image that some helps subsequent treatment, improve picture quality, for the succeeding target recognition and tracking is provided convenience.In particular cases also need change frame resolution sizes and frame per second at some, or the synchronization video video recording that multiple-camera obtains is carried out image co-registration etc.This invention adopts neighborhood averaging that image is carried out pre-service.

The basic thought of neighborhood averaging is, at image space, suppose original image f that a secondary N * N pixel is arranged (x, y), the operation of removing each the pixel point value in the alternative image with the mean value of several pixels in the neighborhood.Through obtain after the smoothing processing sub-picture g (x, y).

$g(x,y)=\frac{1}{M}\underset{(m,n)\∈S}{\mathrm{\Σ}}f(m,n)---\left(1\right)$

In the formula: x, y=0,1,2 ..., N-1; S is that (x y) puts the set of the coordinate of field mid point, but wherein do not comprise (x, y) point; M is the sum of set internal coordinate point.

Step 3: the pretreated image that step 2 is obtained carries out modeling, and with the generation background image, its model parameter can adaptive updates, to handle multiple change of background simultaneously.What generation background partly adopted is that the mixed Gaussian method is carried out modeling, because the actual environment more complicated, so consideration is adopted a plurality of Gauss model hybrid modelings to these multi-modal situations, but and its model parameter adaptive updates, adapt to the background of various variations.Specifically implement according to following steps:

Because the shaft tower installation environment more complicated in the reality had both comprised also comprising night daytime, often also have the tree that shakes, with the wind and moving crops, the cloud that waves etc. on every side, complete static background is non-existent at all.For example, a pixel may be represented sky in certain frame, but then may represent leaf in another frame, and then may represent branch in other frames.Pixel brightness value under each state or color value are different, so consider to use a plurality of Gauss models to come hybrid modeling to these multi-modal situations.

The mixed Gauss model modeling principle is as follows: the gray-scale value of each pixel is described with K Gaussian distribution, usually the K value gets 3～5, the size of K value depends on that calculator memory reaches the rate request to algorithm, the K value is big more, the ability of handling grey scale change is strong more, and the corresponding required processing time is also just long more.Definition pixel gray-scale value variable X _tExpression, its probability density function can be represented with following K three-dimensional Gaussian function:

$P(X_t=\underset{i=1}{\overset{k}{\mathrm{\Σ}}}{\mathrm{\ω}}_{i,t}\·\mathrm{\η}(X_t,{\mathrm{\μ}}_{i,t},{\mathrm{\Σ}}_{i,t})---\left(2\right)$

ω in the formula _{I, t}Be i Gaussian distribution in t weight constantly, and have

η (X _t, μ _{I, t}, ∑ _{I, t}) be t i Gaussian distribution constantly, its average is μ _{I, t}, covariance is a ∑ _{I, t}

$\mathrm{\η}(X_t,{\mathrm{\μ}}_{i,t},{\mathrm{\Σ}}_{i,t})=\frac{1}{{\left(2\mathrm{\π}\right)}^{\frac{n}{2}}{\left|{\mathrm{\Σ}}_{i,t}\right|}^{1/2}}{e}^{-\frac{1}{2}{(X_t-{\mathrm{\μ}}_{i,t})}^T{\mathrm{\Σ}}_{i,t}^{-1}(X_t,{\mathrm{\μ}}_{i,t})}---\left(3\right)$

In the formula, i=1 ..., K; N represents X _tDimension, in order to reduce calculated amount, it has been generally acknowledged that R, G, a B3 passage are separate, and identical variance arranged then have

The expression variance, I representation unit battle array.As shown in Figures 1 and 2, be respectively certain original image of physical activity and handle the background image that the back generates around shaft tower constantly through mixed Gauss model.

Step 4: after the background image for the treatment of step 3 generates, need cut apart to the moving target in the present frame, can be according to the background subtraction point-score etc., subtract each other with current environment background estimating image and current frame image, realize target detection.Specifically implement according to following steps:

On the basis of computing before, we have well generated the background area of image, next are exactly to utilize the method for background difference to detect target.The background subtraction separating method subtracts each other with current environment background estimating image and current frame image exactly, realizes target detection.

As shown in Figure 3, its principle is as follows: at first establishing Bk is background image, f _kBe current frame image, difference image is D _k, D then _k(x, y)=| f _x(x, y)-B _K-1(x, y) |, establish R _kBe bianry image after the difference.To R _kCarry out connectivity analysis, when the area in the zone of a certain connection greater than certain threshold value, think that then the target that detects occurs, and think that the zone of this connection is exactly detected target image.

The binary-state threshold of T for setting.

Step 5: the image after the target detection that step 4 is obtained carries out moving target identification: at first carry out filtering by numerous nontarget areas also being taken as for some reason the image that foreground detection comes out, all section objects are carried out area selection, gray scale selection, morphology processing etc., extract the moving target of present frame.But, usually still have being detected of other like this with the similar especially foreground area of target prospect, some wild animals such as wild boar for example, the activity of automobile etc., so its feature of moving target selective extraction of again needs being discerned, carry out image classification identification with the sorter that trains, to determine main monitoring objective object accurately.Specifically implement according to following steps:

Method with support vector machine is carried out the target image Classification and Identification, and support vector machine realizes it being by certain Nonlinear Mapping (kernel function) of selecting in advance input vector to be mapped to a high-dimensional feature space, structure optimal classification lineoid in this space.

A. linear separability situation: SVM is that optimal classification face under the linear separability situation proposes.So-called optimal classification face requires the classifying face not only can two class sample points are faultless separately exactly, and will make the classification gap maximum of two classes.D be space neutral line discriminant function general type g (x, y)=ω ^TX+b, the classifying face equation is ω ^TX+b=0, we carry out normalization with discriminant function, and all samples of two classes are all satisfied | g (x) | 〉=1, this moment is from the nearest sample of classifying face | g (x) |=1, and require classifying face to all samples can both be correct classification, require it to satisfy exactly

y _i(ω ^Tx+b)-1≥0，i＝1，2，...，n (5)

Be that those samples that equal sign is set up are called support vector in the formula.The gap size in the classification space of two class samples:

M?arg?in＝2/||w|| (6)

Therefore, optimal classification face problem can be expressed as following about fasciculation problem, promptly under the constraint of condition (5), asks function

Minimum value.For this reason, the Lagrange function that can be defined as follows:

$L(w,b,\mathrm{\α})=\frac{1}{2}{w}^{T}w-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i[y_i(w^T{x}_i+b)-1]---\left(7\right)$

Wherein, α _i" 0 be the Lagrange coefficient, our problem is that w and b are asked Lagrange minimum of a function value.Wushu (7) is respectively to w, b, α _iAsk partial differential and make them equal 0:

$\frac{\∂L}{\∂w}=0\⇒w=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i{y}_i{x}_i---\left(8\right)$

$\frac{\∂L}{\∂b}=0\⇒\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i{y}_i=0---\left(9\right)$

$\frac{\∂L}{{\∂\mathrm{\α}}_i}=0\⇒{\mathrm{\α}}_i[y_i(w^T{x}_i+b)-1]=0---\left(10\right)$

More than three formulas add that former constraint condition can change into former problem the dual problem of convex quadratic programming:

$\max \underset{i=1}{\overset{n}{\mathrm{\Σ}}}{a}_i-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i{\mathrm{\α}}_j{y}_i{y}_j\left(x_i^T{x}_j\right)---\left(11\right)$

A wherein _i" 0, i=1 ..., n, $\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{a}_i{y}_i=0,$

This is with quadratic function mechanism problem under the individual inequality constrain, has unique optimum solution.If

Be optimum solution, then

$w^{*}=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{a}_i^{*}{y}_i{x}_i---\left(12\right)$

Non-vanishing sample is support vector, and therefore, the overall coefficient vector of optimal classification face is the linear combination of support vector.b ^*Can be by constraint condition α _i[y _i(w ^Tx _i+ b)-1]=0 to find the solution, the optimal classification function of trying to achieve thus is

$f\left(x\right)=\mathrm{sgn}({\left(w^{*}\right)}^{T}x+b^{*})=\mathrm{sgn}(\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{a}_i^{*}{y}_i{x}_i^{*}x+b^{*})---\left(13\right)$

Sgn () is a sign function.

B. the non-linear situation of dividing

When can not be with a lineoid fully separately the time 2 class points (have only base point by mistake minute), can introduce slack variable ξ _i(ξ _i〉=0, i=1 n), makes lineoid ω ^TX+b=0 satisfies:

y _i(ω ^Tx _i+b)≥1-ξ _i (14)

As 0＜ξ _i＜1 o'clock sample point x _iStill correctly classified, and worked as ξ _i〉=1 o'clock sample point x _iDivided by wrong.For this reason, introduce following objective function:

$\mathrm{\ψ}(w,\mathrm{\ξ})=\frac{1}{2}{w}^{T}w+C\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\ξ}}_i---\left(15\right)$

Wherein C is a positive constant, is called penalty factor, and this moment, SVM can realize by quadratic programming (dual program):

$\max \underset{i=1}{\overset{n}{\mathrm{\Σ}}}{a}_i-\frac{1}{2}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i{\mathrm{\α}}_j{y}_i{y}_j\left(x_i^T{x}_j\right)---\left(16\right)$

0≤a wherein _i≤ C, i=1 ..., n,

Figure 4 shows that architecture based on the image classification of SVM.Figure 5 shows that the target image that monitors through behind background difference and the image classification.

Step 6: the image after the Target Recognition that step 5 is obtained carries out target following, after monitoring destination object, because destination object is in constantly moving, the movement position of destination object is inequality in every two field picture, then need to adopt series of algorithms etc. that target image is followed the tracks of, to determine the destination object behavior of whether committing theft.As shown in Figure 6, specifically implement according to following steps:

Motion target tracking is exactly to find out the position sequence of wherein interested moving target in continuous every two field picture in one section sequence image.The purpose of motion target tracking is exactly by analyzing and researching to sequence image, calculate displacement, movement velocity, the quantity of moving target and the kinematic parameters such as movement locus of moving target of moving target in sequential frame image, for the high level of image is understood and analysis provides foundation.The principle of target following is found out the target accurate location exactly in the next frame image.General tracking is the image that at first extracts tracked target, sets up a template, carries out the full figure coupling then in the next frame image, and the ferret out image is up to the position of finding coupling.Its detailed tracking process prescription of one group of result figure that is illustrated in figure 7 as target following is as follows:

(1) in video sequence image, detects new moving target and moving region;

(2) moving target of new detection is cut apart;

(3) extract the feature of new moving target, and set up the object matching template;

(4) position that may occur in next moment with the forecast model target of prediction is to determine next hunting zone of target constantly.

(5) To Template with previous moment carries out match search in the hunting zone of prediction, seek best match position, can in entire image, search for object like this, and only just can search target near the search window predicted position, thereby can improve search precision, quickening search speed greatly.If when in estimation range, not finding target, will carry out special circumstances and handle.(as blocking situation, pause conditions etc.)

(6) utilize the target image that matches, revise the template data of tracked target, this process constantly alternately repeats.

Step 7: the result after the tracking that obtains according to step 6, if monitor the whereabouts of same target in the continuous n width of cloth image, then start and report to the police, the monitoring personnel will carry out key monitoring to it, have larceny to notify the personnel of patrolling and examining to rush towards the scene immediately in case find it.Otherwise repeat above each step.

The present invention is based on the digital remote image processing techniques, with the digital picture that intercepts in the video flowing that is sent to Surveillance center is research object, adopt a kind of new algorithm mainly in order to the activity of the human body in the piece image around monitoring, the identification shaft tower, when finding that the pedestrian that monitors commits theft suspicion, in time report to the police, for guaranteeing that the electric power enterprise production safety provides a kind of new accurate means directly perceived, has very important realistic meaning.

Claims (6)

1. the electric power line pole tower anti-thefting monitoring method based on the video variance analysis is characterized in that, specifically implements according to following steps:

Step 1:, the transmission line of electricity vision signal that collects is sent to Surveillance center by video server by camera collection transmission line of electricity vision signal; Surveillance center intercepts the digital picture of the transmission line of electricity that needs supervision, the target image that obtains monitoring from video flowing;

Step 2: the target image of the supervision that step 1 is obtained adopts neighborhood averaging to carry out pre-service;

Step 3: the pretreated image that step 2 is obtained adopts the mixed Gaussian method to carry out modeling, the generation background image;

Step 4: after treating the background image generation of step 3, adopt the background subtraction point-score that the moving target in the present frame is cut apart, realize target detection;

Step 5: the image after the target detection that the method that adopts support vector machine obtains step 4 carries out moving target identification, obtains main monitoring objective object;

Step 6: the image after the Target Recognition that step 5 is obtained carries out target following, determines the destination object behavior of whether committing theft.

2. the electric power line pole tower anti-thefting monitoring method based on the video variance analysis according to claim 1 is characterized in that the neighborhood averaging in the described step 2 is carried out pre-service, specifically implements according to following steps:

At image space, suppose original image f that a secondary N * N pixel is arranged (x, y), with the operation of each the pixel point value in the mean value alternative image of several pixels in the neighborhood, through obtain after the smoothing processing sub-picture g (x, y):

$g(x,y)=\frac{1}{M}\underset{(m,n)\∈S}{\mathrm{\Σ}}f(m,n),$

In the formula: x, y=0,1,2 ..., N-1; S is that (x y) puts the set of the coordinate of field mid point, but wherein do not comprise (x, y) point; M is the sum of set internal coordinate point.

3. the electric power line pole tower anti-thefting monitoring method based on the video variance analysis according to claim 1 is characterized in that, adopts the mixed Gaussian method to carry out modeling in the described step 3, specifically implements according to following steps:

The gray-scale value of each pixel is described with K Gaussian distribution, and the K value gets 3～5, and the size of K value depends on that calculator memory reaches the rate request to algorithm, the K value is big more, the ability of handling grey scale change is strong more, and the corresponding required processing time is also just long more, definition pixel gray-scale value variable X _tExpression, its probability density function can be represented with following K three-dimensional Gaussian function:

$P(X_t=\underset{i=1}{\overset{k}{\mathrm{\Σ}}}{\mathrm{\ω}}_{i,t}\·\mathrm{\η}(X_t,{\mathrm{\μ}}_{i,t},{\mathrm{\Σ}}_{i,t}),$

ω in the formula _{I, t}Be i Gaussian distribution in t weight constantly, and have

η (X _t, μ _{I, t}, ∑ _{I, t}) be t i Gaussian distribution constantly, its average is μ _{I, t}, covariance is a ∑ _{I, t},

$\mathrm{\η}(X_t,{\mathrm{\μ}}_{i,t},{\mathrm{\Σ}}_{i,t})=\frac{1}{{\left(2\mathrm{\π}\right)}^{\frac{n}{2}}{\left|{\mathrm{\Σ}}_{i,t}\right|}^{1/2}}{e}^{-\frac{1}{2}{(X_t-{\mathrm{\μ}}_{i,t})}^T{\mathrm{\Σ}}_{i,t}^{-1}(X_t,{\mathrm{\μ}}_{i,t})},$

In the formula, i=1 ..., K; N represents X _tDimension, R, G, three passages of B are separate, and identical variance is arranged, and then have

The expression variance, I representation unit battle array.

4. the electric power line pole tower anti-thefting monitoring method based on the video variance analysis according to claim 1, it is characterized in that, adopt the background subtraction point-score that the moving target in the present frame is cut apart in the described step 4, realize target detection, specifically implement according to following steps:

At first establish B _kBe background image, f _kBe current frame image, difference image is D _k, D then _k(x, y)=| f _x(x, y)-B _K-1(x, y) |, establish R _kBe bianry image after the difference; To R _kCarry out connectivity analysis, when the area in the zone of a certain connection greater than certain threshold value, think that then the target that detects occurs, and think that the zone of this connection is exactly detected target image,

The binary-state threshold of T for setting.

5. the electric power line pole tower anti-thefting monitoring method based on the video variance analysis according to claim 1, it is characterized in that, image after the target detection that the method that adopts support vector machine in the described step 5 obtains step 4 carries out moving target identification, specifically implements according to following steps:

A. linear separability situation:

The general type g of space neutral line discriminant function (x, y)=ω ^TX+b, the classifying face equation is ω ^TX+b=0 carries out normalization with discriminant function, and all samples of two classes are all satisfied | g (x) | 〉=1, this moment is from the nearest sample of classifying face | g (x) |=1, require classifying face to all samples can both be correct classification, require it to satisfy exactly:

y _i(ω ^Tx+b)-1≥0，i＝1，2，...，n，

Be that those samples that equal sign is set up are support vector in the formula, the gap size in the classification space of two class samples:

M?arg?in＝2/||w||，

Therefore, optimal classification face problem can be expressed as following about fasciculation problem, promptly at condition y _i(ω ^TX+b)-1 〉=0, i=1,2 ..., under the constraint of n, ask function

Minimum value, the Lagrange function that is defined as follows:

$L(w,b,\mathrm{\α})=\frac{1}{2}{w}^{T}w-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i[y_i(w^T{x}_i+b)-1],$

Wherein, α _i" 0 be the Lagrange coefficient, following formula respectively to w, b, α _iAsk partial differential and make them equal 0:

$\frac{\∂L}{\∂w}=0\⇒w=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i{y}_i{x}_i,$

$\frac{\∂L}{\∂b}=0\⇒\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i{y}_i=0,$

$\frac{\∂L}{{\∂\mathrm{\α}}_i}=0\⇒{\mathrm{\α}}_i[y_i(w^T{x}_i+b)-1]=0,$

More than three formulas add that former constraint condition can change into former problem the dual problem of convex quadratic programming:

$\max \underset{i=1}{\overset{n}{\mathrm{\Σ}}}{a}_i-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i{\mathrm{\α}}_j{y}_i{y}_j\left(x_i^T{x}_j\right),$

A wherein _i" 0, i=1 ..., n,

If

Be optimum solution, then

$w^{*}=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{a}_i^{*}{y}_i{x}_i,$

Non-vanishing sample is support vector, and therefore, the overall coefficient vector of optimal classification face is the linear combination of support vector, b ^*By constraint condition α _i[y _i(w ^Tx _i+ b)-1]=0 to find the solution, the optimal classification function of trying to achieve thus is:

$f\left(x\right)=\mathrm{sgn}({\left(w^{*}\right)}^{T}x+b^{*})=\mathrm{sgn}(\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{a}_i^{*}{y}_i{x}_i^{*}x+b^{*}),$

Sgn () is a sign function;

B. the non-linear situation of dividing:

In the time can not separating 2 class points fully, introduce slack variable ξ with a lineoid _i(ξ _i〉=0, i=1 n), makes lineoid ω ^TX+b=0 satisfies:

y _i(ω ^Tx _i+b)≥1-ξ _i，

As 0＜ξ _i＜1 o'clock sample point x _iStill correctly classified, and worked as ξ _i〉=1 o'clock sample point x _iDivided by wrong, introduce following objective function:

$\mathrm{\ψ}(w,\mathrm{\ξ})=\frac{1}{2}{w}^{T}w+C\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\ξ}}_i,$

Wherein C is a positive constant, is called penalty factor, and this moment, SVM realized by quadratic programming:

$\max \underset{i=1}{\overset{n}{\mathrm{\Σ}}}{a}_i-\frac{1}{2}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i{\mathrm{\α}}_j{y}_i{y}_j\left(x_i^T{x}_j\right),$

0≤a wherein _i≤ C, i=1 ..., n, $\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{a}_i{y}_i=0.$

6. the electric power line pole tower anti-thefting monitoring method based on the video variance analysis according to claim 1 is characterized in that the image after the Target Recognition that in the described step 6 step 5 is obtained carries out target following, specifically implements according to following steps:

1) in video sequence image, detects new moving target and moving region;

2) moving target of new detection is cut apart;

3) extract the feature of new moving target, and set up the object matching template;

4) with the forecast model target of prediction in next position that constantly may occur, determine next hunting zone of target constantly;

5) To Template with previous moment carries out match search in the hunting zone of prediction, seeks best match position;

6) utilize the target image that matches, revise the template data of tracked target, alternately repeat said process.

Images