CN103049739A - Tree detection method for use in intelligent monitoring of power transmission line - Google Patents

Abstract

The invention discloses a tree detection method for use in intelligent monitoring of a power transmission line, which belongs to the technical field of image processing. The method comprises the following steps of: acquiring a video monitoring image of a high-voltage power transmission line protection region, and acquiring a tree image serving as training data; establishing visual characteristic vectors of tree pixel points; training a classifier through a machine learning method, and judging whether pixels in the image are trees or not; specific to a real-time video monitoring image, finding pixel points belonging to trees in the image; judging whether a large quantity of pixel points belong to trees or not in an interested region by adopting a region tracking method, and partitioning tree image regions; if a large quantity of pixel points belong to trees in the interested region, judging that trees exist, and transmitting alarm information; and meanwhile, transmitting an alarm signal to a remote monitoring client to remind monitoring personnel of paying attention to the growing states of trees. Due to the adoption of a region-based partitioning technology, local space information of images is utilized effectively, and the defect of discontinuity of image partitioning spaces existing in other methods can be overcome effectively.

Description

Trees detection method in a kind of power transmission line intelligent monitoring

Technical field

The invention belongs to technical field of image processing, relate in particular to a kind of intelligent control method for transmission line of electricity.

Background technology

Arboreal growth touches ultra-high-tension power transmission line, and being affects ultra-high-tension power transmission line (electrical network) safety and stable main outside destroy hidden danger.In summer, arboreal growth is very fast especially, and power load is high, and trees touch easily because of the higher and sagging hv transmission line of load, and the accident that causes power failure occurs.District of Shanghai at present to the maintenance work of ultra-high-tension power transmission line mainly by manually finishing, rely on and repeat at the scene to patrol and examine to prevent external force to the destruction of transmission line of electricity that the contradiction that transmission line of electricity quantity rapid growth and operations staff are equipped with between the deficiency increases day by day a large amount of artificial every days.And the transmission line of electricity accident that causes because of outside destroy in recent years rises year by year, has illustrated that also traditional tour mode can not satisfy existing demand for security, and manual inspection is difficult in time provide alert and alarm to appearing at the danger that reaches on the steel tower around the circuit.Therefore being badly in need of adopting new technological means to help the circuit operations staff increases work efficiency.

Active Eyes is incorporated in the maintenance of transmission line of electricity, circuit is carried out monitoring remote video and early warning, can make the long-range real-time monitoring images of checking circuit of O﹠M personnel, in time understand field data, accident is eliminated in bud, effectively reduced the transmission line of electricity accident that the outside destroy factor causes.The present invention is realizing on the video monitoring system basis; added the function of video image being carried out intellectual analysis; trees in the real-time detection line protected location; in the time that distance is less than safe distance between trees and the wire, automatically send early warning information; avoid manual inspection to be difficult to the problem that in time provides alert to appearing at circuit trees harm on every side, realized intelligent monitoring.

Image segmentation is the basic problem of image processing field, also is the committed step that image is processed and analyzed.Image segmentation is divided into the zone of each tool characteristic to image, and utilizes Partial Feature in the image information to extract technology and the process of some interesting targets in the image.Based on the video monitoring image of transmission line of electricity, for judging the distance between trees top branches and leaves and the hi-line, at first to cut apart Tree image, thereby judge that distance between trees and the hi-line is whether in safe range.Because the background of video image is complicated, the cutting apart first difficulty relatively of Tree image, therefore, Method of Tree Image Division is the key that realizes based on the ultra-high-tension power transmission line trees harm monitoring of video.

Compared to other images, Tree image has its singularity, and is simultaneously also more complicated.As in input picture, containing a large amount of non-target objects, such as the buildings of difformity, size, color, go back live wire, people, billboard, sky etc., their color, shape also have very large difference with the target trees.In image, also contain in addition the plant very similar to target, such as different seeds, green lawn etc.And for target Tree image itself, the also shape that can sum up or the color general character can be used as the accurate parameters of cutting apart not, the at present domestic research that Tree image is cut apart is at the early-stage.

Summary of the invention

Technical matters to be solved by this invention provides trees detection method in a kind of power transmission line intelligent monitoring, it adopts area tracking method to be partitioned into the trees zone from video image, area tracking is to seek the pixel group with similarity, this cutting techniques based on the zone has effectively utilized the local spatial information of image, can effectively overcome the discontinuous shortcoming in image segmentation space that additive method exists.

Technical scheme of the present invention is: trees detection method in a kind of power transmission line intelligent monitoring is provided, it is characterized in that described detection method may further comprise the steps:

A, obtain the video monitoring image of ultra-high-tension power transmission line protected location, gather Tree image as training data;

The visual feature vector of B, structure trees pixel;

C, by the method for machine learning, train a sorter, judge whether pixel is trees in the image;

D, for the real-time video monitoring image, in image, find the pixel that belongs to tree;

E, employing area tracking method, whether judge has a large amount of pixels to belong to trees, to be partitioned into the Tree image zone in the area-of-interest;

If there are a large amount of pixels to belong to trees in the F area-of-interest, namely there are trees, warning message then occurs;

G, simultaneously the remote monitoring client is sent alerting signal to remind the monitor staff to note observing the growth conditions of trees.

Concrete, in described step B, the visual feature vector of the trees pixel of its structure has 27 dimensions; Wherein, color characteristic 6 dimension, 3 Vc IE Lab color characteristics wherein, 3 dimension illumination invariant color characteristics; Textural characteristics 18 dimensions are the gabor feature of 3 yardsticks, 6 angles of each yardstick; 3 dimension entropy features.

In described step B, for each feature, the as a result linear normalization after all will calculating to [0,1), that is:

$\hat{f}=\frac{f-\min \left\{f\right\}}{\max \left\{f\right\}-\min \left\{f\right\}}.$

In described step C, adopt the sorter of LIBSVM training trees pixel identification; Adopt the step of LIBSVM training trees classify of image element device to be:

4-1) prepare data set according to the desired form of LIBSVM software package;

4-2) data are carried out simple zoom operations;

4-3) select kernel function, be generally RBF kernel function K (X _i, X _j)=exp (γ || X _i-X _j‖ ²), γ〉0;

4-4) adopt cross validation to select optimal parameter C and g;

4-5) adopt optimal parameter C and g that whole training set is trained and obtain supporting vector machine model;

4-6) utilize the model obtain to test and predict.

At described step D, in the real-time video monitoring image, each pixel is classified with the sorter of described step C training, judge whether this pixel is the trees pixel.

In described step e, the Tree image that adopts area tracking method to be partitioned in the image is regional, namely seeks the pixel group with similarity.

In described step F, according to the position of real-time video monitoring image mesohigh transmission pressure, and seek the Tree image zone be partitioned into, judge the distance between trees and the wire, when this distance less than safe distance the time, send warning message.

Compared with the prior art, advantage of the present invention is:

1. the employing area tracking method is sought the pixel group with similarity.With the trees classify of image element device that trains, in image, find the pixel that belongs to tree, whether area-of-interest in have a large amount of pixel belong to trees, the words that are then be partitioned into Tree image zone, when the distance between trees and the wire is sent warning message less than safe distance the time if then judging;

2. adopt area tracking method to be partitioned into the trees zone from video image, this cutting techniques based on the zone has effectively utilized the local spatial information of image, can effectively overcome the discontinuous shortcoming in image segmentation space that additive method exists.

Description of drawings

Fig. 1 is detection method schematic flow sheet of the present invention.

Embodiment

The present invention will be further described below in conjunction with accompanying drawing.

The purpose of this invention is to provide a kind of real-time video monitoring image based on the ultra-high-tension power transmission line protected location, the trees in the image are carried out intellectuality detect, judge the method for the distance between trees and the wire.

Gather the video monitoring image of high voltage power transmission protected location, the Tree image pixel is made up visual feature vector, visual signature is 27 dimensions altogether.Color characteristic 6 dimension, 3 Vc IE Lab color characteristics wherein, 3 dimension illumination invariant color characteristics; Textural characteristics 18 dimensions are the gabor feature of 3 yardsticks, 6 angles of each yardstick; 3 dimension entropy features:

(1) color characteristic.Comprise two class color characteristics:

One class is three passage L of CIE Lab color space, a, b feature.

Color is that human visual system is for the sensing results of visible light.Rgb value is the first data that camera obtains, but its shortcoming is not directly perceived, and inhomogeneous, so only be difficult to the Color Cognition attribute of knowing that this value is represented from rgb value.The present invention at first with the RGB color space conversion to the CIE color space.CIE LAB is a kind of uniform color space, and it is the abbreviation of CIE 1976 LAB color spaces, is also referred to as CIE1976 L*a*b* (being abbreviated as CIE L*a*b*) color space.CIE LAB color system is most popular object color measure, and as the international standard of measuring color.Conversion from the RGB color space to CIE LAB color space will be passed through XYZ space, LAB space.

CIELAB uses L*, a* and b* coordinate axis definition CIE color space, and wherein: the L* value represents luminance brightness, and it is worth from 0(black)～100(white); A* and b* represent chromaticity coordinate, and wherein a* represents red-green axle, and b* represents Huang-blue axle, and their value represents colourless from 0～10, a*=b*=0; The L* representative is from black to white scale-up factor.RGB is a linear transformation to the transfer process of XYZ, and its conversion can be passed through one 3 * 3 matrix form:

$\left[\begin{array}{c}X\\ Y\\ Z\end{array}\right]=\left[\begin{array}{ccc}0.489989& 0.310008& 0.20\\ 0.176962& 0.812400& 0.01\\ 0.000000& 0.010000& 0.99\end{array}\right]\left[\begin{array}{c}R\\ G\\ B\end{array}\right]$

In the formula, X, Y and Z represent the values of three primary colours, calculate L*, and a* and b* calculate according to following transform:

$L^{*}=\left\{\begin{array}{cc}116{\left(\frac{Y}{Y_n}\right)}^{\frac{1}{3}}-16,& \left(\frac{Y}{Y_n}\right)>0.008856\\ 903.3\left(\frac{Y}{Y_n}\right),& \left(\frac{Y}{Y_n}\right)\≤0.008856\end{array}\right.$

$a^{*}=500[f\left(\frac{X}{X_n}\right)-f\left(\frac{Y}{Y_n}\right)]$

$b^{*}=200[f\left(\frac{Y}{Y_n}\right)-f\left(\frac{Z}{Z_n}\right)]$

Wherein,

$f\left(t\right)=\left\{\begin{array}{cc}{t}^{\frac{1}{3}},& \frac{Y}{Y_n}>0.008856\\ 7.787t+\frac{16}{116},& \frac{Y}{Y_n}\≤0.008856\end{array}\right.$

X _n, Y _nAnd Z _nBe the coordinate of CIE standard light source, value is X _n=90.45, Y _n=150, Z _n=100.665.

Equations of The Second Kind is a kind of color characteristic of illumination invariant, transforms to the CIE XYZ space from RGB first:

$\left[\begin{array}{c}X\\ Y\\ Z\end{array}\right]=\left[\begin{array}{ccc}0.489989& 0.310008& 0.20\\ 0.176962& 0.812400& 0.01\\ 0.000000& 0.010000& 0.99\end{array}\right]\left[\begin{array}{c}R\\ G\\ B\end{array}\right]$

Then do conversion by following formula,

Wherein:

$\stackrel{\→}{x}={(X,Y,Z)}^T$

$A=\left[\begin{array}{ccc}2.707439\×{10}^1& -2.280783\×{10}^1& -1.806681\\ -5.646736& -7.722125& 1.286503\×{10}^1\\ -4.163133& -4.579428& -4.576049\end{array}\right]$

$B=\left[\begin{array}{ccc}9.465229\×{10}^{-1}& 2.946927\×{10}^{-1}& -1.313419\×{10}^{-1}\\ -1.179179\×{10}^{-1}& 9.929960\×{10}^{-1}& 7.371554\×{10}^{-3}\\ 9.230461\×{10}^{-2}& -4.645794\×{10}^{-2}& 9.946464\×{10}^{-1}\end{array}\right]$

(2) textural characteristics

Ask for the Gabor wavelet character of each pixel, 3 yardsticks, 6 angles of each yardstick at gray level image.

(3) entropy feature

Centered by each pixel, get the window of a N * N size, add up the grey level histogram (number of bin gets 64) in each window, the entropy of bin in the calculation window then is specially:

$H=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}-p_i\log p_i,$

P wherein _iBe the probability that i bin occurs, M is the number of bin, and window size N is taken as respectively 5,9,11.For speed-up computation, when asking for each window histogrammic, adopted integration histogram (integralhistogram).

For each feature, the as a result linear normalization after all will calculating to [0,1), that is:

$\hat{f}=\frac{f-\min \left\{f\right\}}{\max \left\{f\right\}-\min \left\{f\right\}}$

Adopt support vector machine (Support Vector Machine, SVM) to train the sorter of trees pixel.

Support vector machine (Support Vector Machine, SVM) be a kind of based on structural risk minimization (Structure Risk Minimization, abbreviation SRM) general learning algorithm, it is that basic thought is to construct an optimum lineoid at sample input control or feature space, so that lineoid reaches maximum to the distance between the two class sample sets, thereby obtain best generalization ability.Construct optimum lineoid and can be converted into a quadratic programming problem.Be different from neural network, the solution of support vector machine is global optimum, and support vector machine does not need artificial planned network structure.SVM Machine Learning Problems under Small Sample Size has good application, and is fit to very much the classification of two class problems.

In the present invention, adopt the sorter of LIBSVM training trees pixel identification.LIBSVM is a software package of increasing income, and is that doctor Lin Zhiren of Taiwan Univ. waits exploitation, and the three class machine learning basic problems of mentioning above can solving provide linearity, polynomial expression, radial basis and four kinds of kernel functions commonly used of sigmoid function selective.

Adopt the step of LIBSVM training trees classify of image element device to be:

1) prepares data set according to the desired form of LIBSVM software package;

2) data are carried out simple zoom operations;

3) select kernel function, be generally RBF kernel function K (X _i, X _j)=exp (γ || X _i-X _j‖ ²), γ〉0;

4) adopt cross validation to select optimal parameter C and g;

5) adopt optimal parameter C and g that whole training set is trained and obtain supporting vector machine model;

6) utilize the model obtain to test and predict.

Tree image in the real time video image is cut apart.

The image segmentation definition: make the I presentation video, H represents to have the predicate of same nature, and image segmentation resolves into n regional R to I _i, i=1,2 ..., n, satisfy:

(1) $\underset{i=1}{\overset{N}{\∪}}{R}_i=I,R_i\∩R_i=\mathrm{\φ},\∀i,j,i\≠j$

(2) $\∀i,i=\mathrm{1,2},\·\·\·,n,H\left(R_i\right)=\mathrm{True}$

(3) $\∀i,j,i\≠j,H(R_i\∪R_i)=\mathrm{False}$

Compared to other images, Tree image has its singularity, and is simultaneously also more complicated.As in input picture, containing a large amount of non-target objects, such as the buildings of difformity, size, color, go back live wire, people, billboard, sky etc., their color, shape also have very large difference with the target trees.In image, also contain in addition the plant very similar to target, such as different seeds, green lawn etc.

Among Fig. 1; described from the video image of the ultra-high-tension power transmission line protected location that video camera 100 is taken; gather Tree image data 101; by making up trees pixel visual feature vector 102; the Tree image data as training data; training obtains trees classify of image element device 103; for the real-time video monitoring image; pixel in the 103 pairs of images of sorter that obtain with training is classified; judge whether pixel is the trees pixel; if in area-of-interest, there are a large amount of trees pixels; then can cut apart 105 to the Tree image zone, judge 106 according to the distance between Tree image zone and the high voltage electricity transmission conductive wire, if this distance is less than safe distance; then send abnormal event alarming 107, according to testing result to monitor client 108; or 109; or 110 send early warning signal.

Concrete steps are as follows:

1, gather the video monitoring image of high voltage power transmission protected location, the Tree image pixel is made up visual feature vector, visual signature is 27 dimensions altogether.Color characteristic 6 dimension, 3 Vc IE Lab color characteristics wherein, 3 dimension illumination invariant color characteristics; Textural characteristics 18 dimensions are the gabor feature of 3 yardsticks, 6 angles of each yardstick; 3 dimension entropy features:

(1) color characteristic.Comprise two class color characteristics,

One class is three passage L of CIE Lab color space, a, b feature.

CIELAB uses L*, a* and b* coordinate axis definition CIE color space, and wherein: the L* value represents luminance brightness, and it is worth from 0(black)～100(white); A* and b* represent chromaticity coordinate, and wherein a* represents red-green axle, and b* represents Huang-blue axle, and their value represents colourless from 0～10, a*=b*=0; The L* representative is from black to white scale-up factor.RGB is a linear transformation to the transfer process of XYZ, and its conversion can be passed through one 3 * 3 matrix form:

$\left[\begin{array}{c}X\\ Y\\ Z\end{array}\right]=\left[\begin{array}{ccc}0.489989& 0.310008& 0.20\\ 0.176962& 0.812400& 0.01\\ 0.000000& 0.010000& 0.99\end{array}\right]\left[\begin{array}{c}R\\ G\\ B\end{array}\right]$

In the formula, X, Y and Z represent the values of three primary colours, calculate L*, and a* and b* calculate according to following transform:

$L^{*}=\left\{\begin{array}{cc}116{\left(\frac{Y}{Y_n}\right)}^{\frac{1}{3}}-16,& \left(\frac{Y}{Y_n}\right)>0.008856\\ 903.3\left(\frac{Y}{Y_n}\right),& \left(\frac{Y}{Y_n}\right)\≤0.008856\end{array}\right.$

$a^{*}=500[f\left(\frac{X}{X_n}\right)-f\left(\frac{Y}{Y_n}\right)]$

$b^{*}=200[f\left(\frac{Y}{Y_n}\right)-f\left(\frac{Z}{Z_n}\right)]$

Wherein,

$f\left(t\right)=\left\{\begin{array}{cc}{t}^{\frac{1}{3}},& \frac{Y}{Y_n}>0.008856\\ 7.787t+\frac{16}{116},& \frac{Y}{Y_n}\≤0.008856\end{array}\right.$

X _n, Y _nAnd Z _nBe the coordinate of CIE standard light source, value is X _n=90.45, Y _n=150, Z _n=100.665.

Equations of The Second Kind is a kind of color characteristic of illumination invariant, transforms to the CIE XYZ space from RGB first:

$\left[\begin{array}{c}X\\ Y\\ Z\end{array}\right]=\left[\begin{array}{ccc}0.489989& 0.310008& 0.20\\ 0.176962& 0.812400& 0.01\\ 0.000000& 0.010000& 0.99\end{array}\right]\left[\begin{array}{c}R\\ G\\ B\end{array}\right]$

Then do conversion by following formula,

Wherein:

$\stackrel{\→}{x}={(X,Y,Z)}^T$

$A=\left[\begin{array}{ccc}2.707439\×{10}^1& -2.280783\×{10}^1& -1.806681\\ -5.646736& -7.722125& 1.286503\×{10}^1\\ -4.163133& -4.579428& -4.576049\end{array}\right]$

$B=\left[\begin{array}{ccc}9.465229\×{10}^{-1}& 2.946927\×{10}^{-1}& -1.313419\×{10}^{-1}\\ -1.179179\×{10}^{-1}& 9.929960\×{10}^{-1}& 7.371554\×{10}^{-3}\\ 9.230461\×{10}^{-2}& -4.645794\×{10}^{-2}& 9.946464\×{10}^{-1}\end{array}\right]$

(2) textural characteristics

Ask for the Gabor wavelet character of each pixel, 3 yardsticks, 6 angles of each yardstick at gray level image.

(3) entropy feature

Centered by each pixel, get the window of a N * N size, add up the grey level histogram (number of bin gets 64) in each window, the entropy of bin in the calculation window then is specially:

$H=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}-p_i\log p_i,$

P wherein _iBe the probability that i bin occurs, M is the number of bin, and window size N is taken as respectively 5,9,11.For speed-up computation, when asking for each window histogrammic, adopted integration histogram (integralhistogram).

(4) for each feature, the as a result linear normalization after all will calculating to [0,1), that is:

$\hat{f}=\frac{f-\min \left\{f\right\}}{\max \left\{f\right\}-\min \left\{f\right\}}$

2, adopt the sorter of open source software bag LIBSVM training trees pixel identification, concrete steps are:

1) prepares data set according to the desired form of LIBSVM software package;

2) data are carried out simple zoom operations;

3) select kernel function, be generally RBF kernel function K (X _i, X _j)=exp (γ || X _i-X _j‖ ²), γ〉0;

4) adopt cross validation to select optimal parameter C and g;

5) adopt optimal parameter C and g that whole training set is trained and obtain supporting vector machine model;

6) utilize the model obtain to test and predict.

3, adopt area tracking method, utilize the sorter that trains, the Tree image in the real time video image is cut apart.

4, the distance between judgement trees zone and the high voltage electricity transmission conductive wire.

5, when the distance between trees zone and the high voltage electricity transmission conductive wire less than safe distance the time, send abnormal event alarming information.

The present invention adopts area tracking method, seeks the pixel group with similarity.With the trees classify of image element device that trains, in image, find the pixel that belongs to tree, whether area-of-interest in have a large amount of pixel belong to trees, the words that are then be partitioned into Tree image zone, when the distance between trees and the wire is sent warning message less than safe distance the time if then judging.

The present invention can be widely used in the field of the trees in the ultra-high-tension power transmission line protected location being carried out intelligent monitoring.

Claims (7)

1. trees detection method during a power transmission line intelligent is monitored is characterized in that described detection method may further comprise the steps:

A, obtain the video monitoring image of ultra-high-tension power transmission line protected location, gather Tree image as training data;

The visual feature vector of B, structure trees pixel;

C, by the method for machine learning, train a sorter, judge whether pixel is trees in the image;

D, for the real-time video monitoring image, in image, find the pixel that belongs to tree;

E, employing area tracking method, whether judge has a large amount of pixels to belong to trees, to be partitioned into the Tree image zone in the area-of-interest;

If there are a large amount of pixels to belong to trees in the F area-of-interest, namely there are trees, warning message then occurs;

G, simultaneously the remote monitoring client is sent alerting signal to remind the monitor staff to note observing the growth conditions of trees.

2. according to trees detection method in the power transmission line intelligent monitoring claimed in claim 1, it is characterized in that in described step B the visual feature vector of the trees pixel of its structure has 27 dimensions; Wherein, color characteristic 6 dimension, 3 Vc IE Lab color characteristics wherein, 3 dimension illumination invariant color characteristics; Textural characteristics 18 dimensions are the gabor feature of 3 yardsticks, 6 angles of each yardstick; 3 dimension entropy features.

3. according to trees detection method in the power transmission line intelligent claimed in claim 1 monitoring, it is characterized in that in described step B, for each feature, the as a result linear normalization after all will calculating arrive [0,1), that is:

$\hat{f}=\frac{f-\min \left\{f\right\}}{\max \left\{f\right\}-\min \left\{f\right\}}.$

4. according to trees detection method in the power transmission line intelligent monitoring claimed in claim 1, it is characterized in that in described step C, adopt the sorter of LIBSVM training trees pixel identification; Adopt the step of LIBSVM training trees classify of image element device to be:

4-1) prepare data set according to the desired form of LIBSVM software package;

4-2) data are carried out simple zoom operations;

4-3) select kernel function, be generally RBF kernel function K (X _i, X _j)=exp (γ || X _i-X _j‖ ²), γ〉0;

4-4) adopt cross validation to select optimal parameter C and g;

4-5) adopt optimal parameter C and g that whole training set is trained and obtain supporting vector machine model;

4-6) utilize the model obtain to test and predict.

5. according to trees detection method in the power transmission line intelligent monitoring claimed in claim 1, it is characterized in that at described step D, in the real-time video monitoring image, each pixel is classified with the sorter of described step C training, judge whether this pixel is the trees pixel.

6. according to trees detection method in the power transmission line intelligent monitoring claimed in claim 1, it is characterized in that in described step e the Tree image that adopts area tracking method to be partitioned in the image is regional, namely seek the pixel group with similarity.

7. according to trees detection method in the power transmission line intelligent monitoring claimed in claim 1, it is characterized in that in described step F, position according to real-time video monitoring image mesohigh transmission pressure, and it is regional to seek the Tree image that is partitioned into, judge the distance between trees and the wire, when this distance less than safe distance the time, send warning message.

Images