CN106023076A - Splicing method for panoramic graph and method for detecting defect state of guard railing of high-speed railway - Google Patents

Abstract

The embodiment of the invention provides a panoramic graph splicing method based on a virtual sampling channel model and a method for detecting a defect state of a guard railing of a high-speed railway. The detection method comprises: a forward motion video recording a high-speed railway environment state is obtained; with a panoramic graph splicing method based on a virtual sampling channel model, the forward motion video is converted into a left panoramic graph and a right panoramic graph in a railway operating environment, wherein the left panoramic graph and the right panoramic graph are used as a guard railing panoramic graph; on the basis of a maximum entropy segmentation principle, positions of a guard railing and a background in the guard railing panoramic graph are located to obtain a binary code formed by 0 and 1; run coding is carried out on th binary code; according to the run coding, a pixel width Dcur of a background area between two adjacent guard railings is calculated; and according to the pixel width Dcur of the background area between two adjacent guard railings and a normal pixel distance d between the two adjacent guard railings, whether defects exist at the two adjacent guard railings is determined.

Description

The damage condition of the guard rail of the joining method of panorama sketch and detection high-speed railway Method

Technical field

The present invention relates to Computer Applied Technology field, particularly relate to joining method and the detection high-speed iron of a kind of panorama sketch The method of the damage condition of the guard rail on road.

Background technology

Safety is basis and the premise of Development of High Speed Railway.The factor affecting high-speed railway safety is a lot, relates to people, row Car, track and environment and the coupling effect between them.In order to ensure high-speed railway traffic safety, it is desirable to high-speed railway Running environment must realize omnidistance closing.It is a kind of means realizing enclosed environment that line of high-speed railway installs guard rail additional.

Within monitoring artificial destruction or climbing intrusion guard rail, and signalling arrangement and power supply in monitoring enclosed environment The fine status of the infrastructure such as system, at present, the fixed video equipment of main employing is monitored.Such as, in Beijing-Shanghai high speed On railway, more than 400 video monitoring equipments are arranged on Beijing-Shanghai High-Speed Railway roadbed, crossing, bridge, public affairs across ferrum, bottle-neck section etc. Position, it is ensured that vehicle safety runs.

In common video monitoring system, picture pick-up device is in fixed static state, and monitored object is likely to be at fortune Dynamic state.Using the relative motion of target and background in video sequence to detect the situation of target is conventional a kind of moving target Detection method, mainly has: the detection methods such as background subtraction method, time differencing method and optical flow method (Optical flow).Should The advantage of class method, without training, can detect online；And shortcoming is to search out respective objects, and cannot sentence Bright detection region is which kind of target concrete, needs the later stage to determine whether.

Background subtraction method depends primarily on the accuracy to background image modeling, but in actual environment, due to environment The factor impacts such as change, exacerbate extracting and the difficulty updated of background, add difficulty for accurately extracting target.Therefore, mesh Such method front uses statistical learning method to be analyzed successive video frames mostly, background modeling, and As time goes on Online updating background model, detects moving region by the difference between present frame and background model the most again.

Time difference method is a kind of relatively easy, method that operand is less, by calculating before and after video sequence two The pixel of frame correspondence position is poor, if more than the threshold value set, then it is assumed that be object pixel, be otherwise background pixel.Optical flow method is adopted The characteristic that can change over time with the vector of moving target detects moving target.Even if the advantage of the method is at video camera Moving target can also be detected in the case of motion.Its major defect: more sensitive to noise ratio, computationally intensive, be not suitable for The occasion that requirement of real-time is higher.

But, fixing point monitor mode is limited by gathering the visual field, it is impossible to control circuit and the whole circumstances along the line, because of This, utilize and be arranged on the car end surroundings monitoring apparatus of high speed comprehensive detection train, monitors front side line environment and along both side Guard rail state is a kind of effective method.

Special high speed comprehensive detection train is used to complete to affect the Detection task of train operating safety in the world.Detection Content generally comprises: contact net geometry, contact line abrasion, bow net effect, electric parameter, gauge, track geometry, rail profile and Undulatory wear, car body and axle box acceleration, wheel-rail force, track and bogie, communication check and location etc..In addition, High speed comprehensive detection train both domestic and external is equipped with the video equipment as drive recorder at its front and rear, transports with forward direction The mode of dynamic video is used for obtaining circumstance state information on the way, whether there is abnormal shape for later stage manual detection Along Railway environment State provides foundation.How from the multitude of video image obtained, quick obtaining sets in affecting environment closure and foreign body or circuit The standby abnormal information invading limit, and carry out correct early warning, it is high-speed railway problem demanding prompt solution.

Summary of the invention

The embodiment provides the joining method of a kind of panorama sketch and the defect of the guard rail of detection high-speed railway The method of state, carries out lossless information extraction by the video data of magnanimity with the panorama sketch form of lightweight, reduces video The storage of data mode and access expense.

To achieve these goals, this invention takes following technical scheme:

A kind of Panoramagram montage method based on virtual sampling channel model, including:

Obtain the propulsion video of record high-speed railway ambient condition；Extracting frame number from described propulsion video is The sequence of video images of N；

According to the railway scene structure determined via end point, every two field picture sets external sampling rectangle OR_m；

According to the speed of train, every two field picture sets internal sample rectangle IR_m；

By by the described external sampling rectangle OR of every two field picture_mWith described internal sample rectangle IR_mThe rectangular ring of composition is adopted Sample annulus, is divided into four strips mosaic region S_t,S_b,S_l,S_r；

By described four strips mosaic region S of every two field picture_t,S_b,S_l,S_rBy image volume around, be corrected to rule square Shape strips S_t',S_b',S_l',S_r'；

By 4 × N number of corrected described shape of rectangular ribbon S_t',S_b',S_l',S_r', carry out respectively according to respective sample plane Split, generates the panorama sketch of 4 planes of railway scene.

A kind of method of the damage condition of the guard rail of detection high-speed railway based on described Panoramagram montage method, bag Include:

Step one, obtains the propulsion video of record high-speed railway ambient condition；

Step 2, Panoramagram montage method based on virtual sampling channel model, by described propulsion video conversion be The left and right panorama sketch of railway operation environment, as guard rail panorama sketch；

Step 3, according to maximizing entropy segmentation principle, enters the position of the guardrail in described guard rail panorama sketch and background Row location, represents the position of guard rail, represents the position of background at F (j)=0 at F (j)=1, obtain the two-value of 0 and 1 composition Code；J is the row sequence number of guard rail panorama sketch；

Step 4, carries out run-length encoding to described two-value code；

Step 5, according to described run-length encoding, calculates pixel wide D of background area between adjacent two guardrails_cur；

Step 6, according to pixel wide D of the background area between described adjacent two guardrails_curWith described adjacent two guardrails Between normal pixel spacing d, it is judged that whether described adjacent two guardrails exist defect, generate judged result；

Step 7, exports described judged result.

The technical scheme that thered is provided by embodiments of the invention described above it can be seen that in the embodiment of the present invention, regarding magnanimity The panorama sketch form of frequency lightweight according to this carries out lossless information extraction, and the storage and the access that reduce video data form are opened Pin.

Aspect and advantage that the present invention adds will part be given in the following description, and these will become from the following description Obtain substantially, or recognized by the practice of the present invention.

Accompanying drawing explanation

In order to be illustrated more clearly that the technical scheme of the embodiment of the present invention, required use in embodiment being described below Accompanying drawing be briefly described, it should be apparent that, below describe in accompanying drawing be only some embodiments of the present invention, for this From the point of view of the those of ordinary skill of field, on the premise of not paying creative work, it is also possible to obtain other according to these accompanying drawings Accompanying drawing.

Fig. 1 is the schematic flow sheet of Panoramagram montage method based on virtual sampling channel model of the present invention；

Fig. 2 is the damage condition of the guard rail of detection high-speed railway based on Panoramagram montage method of the present invention Method；

Fig. 3 is high speed railway scene sampling ring geometry of the present invention；

Fig. 4 is the image registration in the present invention between adjacent two frames；

Fig. 5 is the histogram distribution of three-dimensional feature F (j) of pixel column in panorama sketch in the present invention；

Fig. 6 is Threshold segmentation based on Mean-Variance-gradient in the present invention；

Fig. 7 is that in the present invention, the run-length encoding of guardrail panorama sketch represents；

Fig. 8 is from closing the panorama sketch produced Railway Environment in the present invention: (a) left side panorama sketch；(b) right side panorama Figure；(c) bottom panorama sketch；(d) top panorama sketch

Fig. 9 is the video image that in the present invention, comprehensive detection train gathers: (a) guardrail N/D；B () guardrail has defect

Figure 10 is part railway panorama sketch in the present invention: the panorama sketch on the left of (a) railway；B protection that () intercepts from (a) Hurdle panorama sketch

Figure 11 is the binarization segmentation result of guardrail panorama sketch in the present invention: (a) gray scale one dimensional histograms maximum entropy threshold Segmentation；B () Gray Level-Gradient (GLGM) two-dimensional histogram maximum entropy threshold is split；C combination gray average that () present invention proposes- The stereogram maximum entropy threshold segmentation of variance-gradient

Figure 12 is the binarization segmentation result of high ferro guardrail panorama sketch in the present invention: the guardrail panorama sketch of (a) high-speed railway； B () gray scale one dimensional histograms maximum entropy threshold is split；C () Gray Level-Gradient (GLGM) two-dimensional histogram maximum entropy threshold is split； The stereogram maximum entropy threshold segmentation of d combination gray average-variance-gradient that () present invention proposes

Figure 13 is high speed railway protective hurdle of the present invention defect method for quick flow chart of steps.

Detailed description of the invention

Embodiments of the present invention are described below in detail, and the example of described embodiment is shown in the drawings, the most ad initio Represent same or similar element to same or similar label eventually or there is the element of same or like function.Below by ginseng The embodiment examining accompanying drawing description is exemplary, is only used for explaining the present invention, and is not construed as limiting the claims.

As described in Figure 1, for a kind of Panoramagram montage method based on virtual sampling channel model of the present invention, bag Include:

Step 11, obtains the propulsion video of record high-speed railway ambient condition；Carry from described propulsion video Take the sequence of video images that frame number is N；

Step 12, according to the railway scene structure determined via end point, sets external sampling rectangle in every two field picture OR_m；

Step 13, according to the speed of train, sets internal sample rectangle IR in every two field picture_m；

Step 14, by by the described external sampling rectangle OR of every two field picture_mWith described internal sample rectangle IR_mThe square of composition Shape ring-type sampling annulus, is divided into four strips mosaic region S_t,S_b,S_l,S_r；

Step 15, by described four strips mosaic region S of every two field picture_t,S_b,S_l,S_rBy image volume around, be corrected to The shape of rectangular ribbon S of rule_t',S_b',S_l',S_r'；

Step 16, by 4 × N number of corrected described shape of rectangular ribbon S_t',S_b',S_l',S_r', according to respective sample plane Carry out split respectively, generate the panorama sketch of 4 planes of railway scene.

Determine the step of described end point particularly as follows:

End point coordinate in image coordinate system is (x₀,y₀)^T；The analytic expression of i-th line section is x+k_iy+b_i=0, k_i For the slope of i-th line section, b_iFor intercept；Weight w_iIt is the length of i-th line section；N is total row of pixel in guard rail panorama sketch Number；

$x_0=\frac{g_1{m}_{22}-g_2{m}_{12}}{m_{11}{m}_{22}-m_{12}^2},y_0=\frac{g_2{m}_{11}-g_1{m}_{12}}{m_{11}{m}_{22}-m_{12}^2};$

Wherein,

$m_{11}=\underset{i=1}{\overset{N}{\Σ}}\frac{w_i}{1+k_i^2},m_{12}=\underset{i=1}{\overset{N}{\Σ}}\frac{w_i{k}_i}{1+k_i^2},m_{22}=\underset{i=1}{\overset{N}{\Σ}}\frac{w_i{k}_i^2}{1+k_i^2},$

$g_1=-\underset{i=1}{\overset{N}{\Σ}}\frac{w_i{b}_i}{1+k_i^2},g_2=-\underset{i=1}{\overset{N}{\Σ}}\frac{w_i{k}_i{b}_i}{1+k_i^2}.$

In the embodiment of the present invention, the video data of magnanimity is carried out lossless information with the panorama sketch form of lightweight and takes out Take, reduce the storage of video data form and access expense, may be used for subsequent treatment.

As described in Figure 2, lacking for the guard rail detecting high-speed railway based on Panoramagram montage method of the present invention The method of damage state, including:

Step 21, obtains the propulsion video of record high-speed railway ambient condition；

Step 22, Panoramagram montage method based on virtual sampling channel model, by described propulsion video conversion be The left and right panorama sketch of railway operation environment, as guard rail panorama sketch；

Step 23, according to maximizing entropy segmentation principle, enters the position of the guardrail in described guard rail panorama sketch and background Row location, represents the position of guard rail, represents the position of background at F (j)=0 at F (j)=1, obtain the two-value of 0 and 1 composition Code；J is the row sequence number of guard rail panorama sketch；

Step 24, carries out run-length encoding to described two-value code；

Step 25, according to described run-length encoding, calculates pixel wide D of background area between adjacent two guardrails_cur；

Step 26, according to pixel wide D of the background area between described adjacent two guardrails_curWith described adjacent two guardrails Between normal pixel spacing d, it is judged that whether described adjacent two guardrails exist defect, generate judged result；Wherein, described step 26 include: judge D_curWhether more than k d, wherein, k is that experience regulates parameter.

Step 27, exports described judged result.

In the embodiment of the present invention, the video data of magnanimity is carried out lossless information with the panorama sketch form of lightweight and takes out Take, reduce the storage of video data form and access expense, and, video is converted into one be more suitable for manually inspecting with And the form of computer analyzing and processing, solve by the undistorted rapid translating of propulsion video to panorama sketch.

Described step 22 includes:

Extracting frame number from described propulsion video is the sequence of video images of N；

According to the railway scene structure determined via end point, every two field picture sets external sampling rectangle OR_m；

According to the speed of train, every two field picture sets internal sample rectangle IR_m；

By every two field picture by described external sampling rectangle OR_mWith described internal sample rectangle IR_mThe rectangular ring of composition is adopted Sample annulus, is divided into four strips mosaic region S_t,S_b,S_l,S_r；Wherein, S_t,S_b,S_l,S_rRespectively sky, rail, left side is protected Hurdle and guardrail region, right side；

By described four strips mosaic region S of every two field picture_t,S_b,S_l,S_rBy image volume around, be corrected to rule square Shape strips S_t',S_b',S_l',S_r'；

By 2 × N number of corrected described shape of rectangular ribbon S_l',S_r', split is carried out respectively according to respective sample plane, raw Become the left and right panorama sketch of railway scene.

Determine the step of described end point particularly as follows:

End point coordinate in image coordinate system is (x₀,y₀)^T；The analytic expression of i-th line section is x+k_iy+b_i=0, k_i For the slope of i-th line section, b_iFor intercept；Weight w_iIt is the length of i-th line section；N is total row of pixel in guard rail panorama sketch Number；

$x_0=\frac{g_1{m}_{22}-g_2{m}_{12}}{m_{11}{m}_{22}-m_{12}^2},y_0=\frac{g_2{m}_{11}-g_1{m}_{12}}{m_{11}{m}_{22}-m_{12}^2};$

Wherein,

$m_{11}=\underset{i=1}{\overset{N}{\Σ}}\frac{w_i}{1+k_i^2},m_{12}=\underset{i=1}{\overset{N}{\Σ}}\frac{w_i{k}_i}{1+k_i^2},m_{22}=\underset{i=1}{\overset{N}{\Σ}}\frac{w_i{k}_i^2}{1+k_i^2},$

$g_1=-\underset{i=1}{\overset{N}{\Σ}}\frac{w_i{b}_i}{1+k_i^2},g_2=-\underset{i=1}{\overset{N}{\Σ}}\frac{w_i{k}_i{b}_i}{1+k_i^2}.$

Described step 23 includes:

For every string j of guard rail panorama sketch, calculate gray average M (1, j), standard deviation V (1, j) and gradient mean value G (1,j)；J=1 ..., N；N is total columns of pixel in guard rail panorama sketch；

Use F (j) maximum entropy, calculate segmentation threshold (ε^*,ξ^*,η^*)；F (j) is the perpendicular of extraction in described guard rail panorama sketch The feature distribution of Nogata upwards each column pixel；

According to described segmentation threshold (ε^*,ξ^*,η^*), F (j) is carried out binarization segmentation, guard rail region is set to 1, background Region is set to 0；

Wherein,；V (1, j) it is the gray average of each column pixel in guard rail panorama sketch；

M (1, j) it is the gray variance of each column pixel in guard rail panorama sketch；G (1, j) it is each column picture in guard rail panorama sketch The gradient mean value of element.

$M(1,j)=\frac{1}{b_h-t_h}\underset{i=t_h}{\overset{b_h}{\Σ}}p(i,j);$

$V(1,j)=\frac{1}{b_h-t_h}\underset{i=t_h}{\overset{b_h}{\Σ}}{\left(p\right(i,j)-M(1,j\left)\right)}^2;$

$G(1,j)=\frac{1}{t_h-b_h}\underset{i=b_h}{\overset{t_h}{\Σ}}\sqrt{{\left(p\right(i+1,j)-p(i,j\left)\right)}^2+{\left(p\right(i,j+1)-p(i,j\left)\right)}^2};$

Wherein, (i j) is (i, gray value j), the t of pixel in guard rail panorama sketch to p_h,b_hIt is respectively the top of guard rail With bottom coordinate position in guard rail panorama sketch.

Another embodiment of the present invention is below described.

The present invention with detection high-speed railway closed operation environment closure as object, it is provided that one utilizes propulsion to regard The Panoramagram montage of frequency, and by panorama sketch feature detection, it is achieved the dynamic detection of high-speed railway guard rail defect.

The technical scheme is that: the propulsion video of record high-speed railway ambient condition is sampled by tunnel Splicing is converted to Railway Environment panorama sketch；Panorama sketch is carried out feature extraction；According to the feature row obtained after feature extraction Journey encodes, and provides railway protective hurdle damage condition and judges and assessment.It is to say, video image to be transformed to a kind of static panorama Bitmap-format, i.e. generates propulsion video panorama under camera optical axis with direction of motion uniform condition；Several based on detection region What structural generation propulsion video virtual sampling channel model；Mean-Variance based on guard rail panorama sketch-gradient (MVG) The segmentation of stereogram maximum entropy threshold and guardrail based on run-length encoding detection.

Specifically comprise the steps of

Step 1, positions propulsion video, the step of virtual rectangle sampling channel based on rail track priori；

Step 2, segmentation propulsion frame of video be sky, the step of left and right four parts of images；

Step 3, the step that Panoramagram montage based on propulsion frame of video calculates；

Step 4, the step to guard rail panorama sketch feature calculation；

Step 5, guardrail detecting step based on run-length encoding.

Wherein, step 1 specifically includes:

One group of annulus sequence { S is extracted from video sequence₁,S₂,S₃..., S_N, and meet adjacent overall view ring belt to thing Reason space sampling between neither overlap the most continuously every, i.e. " fully sampled ".Splice after annulus sequence is carried out Geometric corrections Generate panoramic picture together.

Based on rail track priori, position propulsion video tunnel, be with the object structure of concerns depth layer The virtual rectangle sampling channel of " fully sampled ".The structure of virtual rectangle sampling channel is based on following objective approximation:

(1) rail facility scene includes: the guard rail of both sides or sound barrier, the rail on ground and be suspended on connecing of top Touching net, similar facility is all considered as being located approximately at same plane, and the distance of plane separation camera is known；

(2) each plane of rail facility scene is the distance nearest depth layer of camera, the panorama sketch of generation does not exist because of Lack sampling and the rail facility loss of learning that causes.

Step 2 specifically includes:

(1) end point is determined

Utilize LSD (line segment detector) Line Segment Detection Algorithm that the video image of video sequence is examined Looking into, all line segments detected are divided into two groups, first group is and trunnion axis angle line segment in the range of [-60 °, 60 °], uses Least Square Method is used to be parallel to the end point that the line segment of rail is formed；Second group be with vertical axes angle [-5 °, 5 °] in the range of line segment, be used for estimating the position of line bar, in order to divide railway scene.

If the coordinate that end point is in image coordinate system is (x₀,y₀)^T, it and i-th line section x+k_iy+b_i=0 (k_iFor tiltedly Rate, b_iFor intercept) distance be

$d_i(x_0,y_0)=\frac{\left|x_0+k_i{y}_0+b_i\right|}{\sqrt{1+k_i^2}}---\left(1\right)$

WhereinAccording to method of least square, minimize point (x₀,y₀)^TWith all line segments spacing From quadratic sum, i.e.

$\underset{x_0{y}_0}{\min }\left\{\underset{i=1}{\overset{N}{\Σ}}{w}_i{d}_i^2(x_0,y_0)\right\}---\left(2\right)$

Wherein, weight w_iIt is the length of i-th line section.

Formula (2) is sought x respectively₀And y₀Derivative, and to make it be zero, obtains following system of linear equations,

$\left\{\begin{array}{c}{m}_{11}{x}_0+m_{12}{y}_0=g_1\\ m_{12}{x}_0+m_{22}{y}_0=g_2\end{array}\right.---\left(3\right)$

Here

N is total columns of pixel in guardrail panorama sketch；

The solution of formula (3) is,

$x_0=\frac{g_1{m}_{22}-g_2{m}_{12}}{m_{11}{m}_{22}-m_{12}^2},y_0=\frac{g_2{m}_{11}-g_1{m}_{12}}{m_{11}{m}_{22}-m_{12}^2}---\left(4\right)$

(x₀,y₀) it is exactly the optimal estimation of end point.In order to improve the accuracy of estimating vanishing point, use cross-iteration Method rejects the line segment of " mistake ".

(2) divide sky, left and right four parts.

The virtual rectangle sampling channel rectangular structure being made up of rail, contact net and guard rail or sound barrier has solid Fixed geometry.According to the shape of the virtual rectangle sampling channel of scene, determine that (Fig. 3 is high-speed iron in selected sampling ring position The scene sample ring on road).

Connection end point Q is to rectangular four summits of sampling ring respectively, by four parts of these four wire segmentings just Be sky, left and right four parts.

Step 3. Panoramagram montage based on propulsion frame of video calculates.

As shown in Figure 4, as follows based on vanishing Point Detection Method and camera motion frame registration process:

(1) the railway scene that will pay close attention to, i.e. guardrail, contact net, sound barrier, the position at rail place constitute a virtual square Shape sampling channel, and carry out " fully sampled " at virtual rectangle sampling channel.

(2) given one section of frame number is the sequence of video images { I of N₁,I₂..., I_N-1, I_N, in every two field picture, choose splicing Region, i.e. determines outer rectangular OR_m:A_mB_mC_mD_mAnd inner rectangular IR_m:A′_mB′_mC′_mD′_m。

(3) scene structure determined according to end point, preassigns outer rectangular OR_m:A_mB_mC_mD_mPosition, and at each frame Middle location is the most identical.Four tops respectively through image left and right sides electric pole and bottom is drawn through end point Q Ray, these four rays by scene cut be sky, rail, left side guardrail and four parts of right side guardrail (the most above sky, Ground, left and right four parts).Four summit A of outer rectangular_m、B_m、C_m、D_mShould respectively fall on four rays, can be true for this The panorama sketch content non-overlapping copies of upper and lower, left and right four scene that keeping normal life activities becomes, it is to avoid mistake is sampled, and also simplify virtual field simultaneously The process that scape is drawn.Furthermore, owing to the direction of motion of video camera is parallel with rail, therefore in image in the expansion of pixel light stream The heart (Focus of Expansion, FOE) overlaps with in image two end points Q that rail is formed a long way off.So outside square The moving direction on four summits of shape, i.e. outer rectangular OR_mThe zoom direction of frame just uniquely determines through end point Q over time, It is denoted as:

$\begin{array}{c}{\stackrel{\→}{n}}_m^A:A_m(x_a,y_a)-Q_m(x_q,y_q)\\ {\stackrel{\→}{n}}_m^B:A_m(x_b,y_b)-Q_m(x_q,y_q)\\ {\stackrel{\→}{n}}_m^C:A_m(x_c,y_c)-Q_m(x_q,y_q)\\ {\stackrel{\→}{n}}_m^D:A_m(x_d,y_d)-Q_m(x_q,y_q)\end{array}---\left(5\right)$

(4) setting train speed as V, the frame speed of video camera is R, and between the most adjacent two frames, the distance of spatial sampling is V/R.If The inner parameter of camera is it is known that then by corresponding geometrical calculation, can try to achieve the translational speed of pixel i.e. image speed in image Degree v.

(5) m frame peripheral rectangle OR is obtained_m:A_mB_mC_mD_mZoom directionAnd image After speed v, the coordinate position corresponding in m-1 frame of this outer rectangular can be byDirectly try to achieve.

Outer rectangular OR in (6) m-1 frames_m-1With inner rectangular IR_m-1Constitute splicing regions S of rectangular ring_m, will S_mZoom direction along outer rectangular summitIt is divided into 4 pieces of trapezoidal strips S_t、S_b、S_l、S_r, respectively sky Sky, rail, left side guardrail and guardrail region, right side.

(7) utilize image volume around (Image Wrapping), the strips S to structure_t、S_b、S_l、S_rCarry out geometric transformation, from And irregular trapezoid-shaped strips is mapped as the shape of rectangular ribbon S of rule_t'、S_b'、S_l'、S_r'.Every two field picture is repeated above mistake Journey, obtains four groups of band sequences, and these sequences being stitched together the most successively has just obtained panorama sketch.The present invention relates to left and right The detection of guard rail, only uses shape of rectangular ribbon S for this_l'、S_r' splice the left and right panorama sketch of railway operation environment obtained, i.e. protect Hurdle panorama sketch.

Step 4. guard rail panorama sketch feature calculation.

(i j) is (i, gray value j), the t of pixel in guard rail panorama sketch to make p_h,b_hBe respectively guard rail top and Bottom coordinate position in panorama sketch, then the gray average of each column pixel, gray variance and gradient mean value are denoted as respectively

$M(1,j)=\frac{1}{b_h-t_h}\underset{i=t_h}{\overset{b_h}{\Σ}}p(i,j)---\left(6\right)$

$V(1,j)=\frac{1}{b_h-t_h}\underset{i=t_h}{\overset{b_h}{\Σ}}{\left(p\right(i,j)-M(1,j\left)\right)}^2---\left(7\right)$

$G(1,j)=\frac{1}{t_h-b_h}\underset{i=b_h}{\overset{t_h}{\Σ}}\sqrt{{\left(p\right(i+1,j)-p(i,j\left)\right)}^2+{\left(p\right(i,j+1)-p(i,j\left)\right)}^2}---\left(8\right)$

After carrying out gray scale and Gradient Features extraction along guard rail panorama sketch vertical direction, obtain

{ F (j) }=(x, y) }=(M (1, j), V (1, j), G (1, j)) } (9)

Here 1≤j≤N, N are total columns of pixel in guard rail panorama sketch, and F (j) represents the three-dimensional feature of each column pixel Distribution.

Gray-scale statistical characteristics at guard rail meets: M (1, j) < ε, V (1, j) < ξ, G (1, j) < η.Here ε, ξ and η are for using Foreground area in divided shielding hurdle i.e. positions the segmentation threshold of guard rail position.

On the vertical direction extracted from guard rail panorama sketch, feature distribution F (j) of each column pixel is as shown in Figure 5.

If given any one threshold value (ε, ξ, η), feature distribution F (j) of guardrail guard rail is split, then three-dimensional The spatial distribution of feature F (j) is divided into 8 characteristic sub-areas (as shown in Figure 6) by (ε, ξ, η), is denoted as:

$\begin{array}{c}{R}_1=\{0<x\&le;\mathrm{\&epsiv;},0<y\&le;\mathrm{\&xi;},0<z\&le;\mathrm{\&eta;}\};\\ R_2=\{\mathrm{\&epsiv;}<x\&le;255,0<y\&le;\mathrm{\&xi;},0<z\&le;\mathrm{\&eta;}\};\\ R_3=\{\mathrm{\&epsiv;}<x\&le;255,\mathrm{\&xi;}<y\&le;255,0<z\&le;\mathrm{\&eta;}\};\\ R_4=\{0<x\&le;\mathrm{\&epsiv;},\mathrm{\&xi;}<y\&le;255,0<z\&le;\mathrm{\&eta;}\};\\ R_5=\{0<x\&le;\mathrm{\&epsiv;},\mathrm{\&xi;}<y\&le;255,\mathrm{\&eta;}<z\&le;255\};\\ R_6=\{0<x\&le;\mathrm{\&epsiv;},0<y\&le;\mathrm{\&xi;},\mathrm{\&eta;}<z\&le;255\};\\ R_7=\{\mathrm{\&epsiv;}<x\&le;255,0<y\&le;\mathrm{\&xi;},\mathrm{\&eta;}<z\&le;255\};\\ R_8=\{\mathrm{\&epsiv;}<x\&le;255,\mathrm{\&xi;}<y\&le;255,\mathrm{\&eta;}<z\&le;255\}.\end{array}---\left(10\right)$

Each guard rail in the vertical direction has the brightness value of approximation and uniform Luminance Distribution, and background area is then The scene that some brightness are randomly distributed.The feature which results in guard rail position is mostly focused on R₁Region, and the spy of background Levy and be mostly focused on R₈Region.Although other 6 regions also contains guardrail and the background information of part, in order to simplify calculating Process, ignores test result indicate that of these secondary feature regions and positioning result there is no big impact.

In F (j), any one characteristic vector is positioned at guardrail region R₁Or background area R₈Probability be respectively P_F(ε,ξ, η) and P_B(ε,ξ,η)；

$P_F(\mathrm{\&epsiv;},\mathrm{\&xi;},\mathrm{\&eta;})=\underset{x=0}{\overset{\mathrm{\&epsiv;}}{\&Sigma;}}\underset{y=0}{\overset{\mathrm{\&xi;}}{\&Sigma;}}\underset{z=0}{\overset{\mathrm{\&eta;}}{\&Sigma;}}{p}_{xyz}---\left(11\right)$

$P_B(\mathrm{\&epsiv;},\mathrm{\&xi;},\mathrm{\&eta;})=\underset{x=\mathrm{\&epsiv;}+1}{\overset{255}{\&Sigma;}}\underset{y=\mathrm{\&xi;}+1}{\overset{255}{\&Sigma;}}\underset{z=\mathrm{\&eta;}+1}{\overset{255}{\&Sigma;}}{p}_{xyz}---\left(12\right)$

Here, p_xyzIt is positioned at the probability in prospect or background for the pixel in image.

And meet

P_F(ε,ξ,η)+P_B(ε,ξ,η)≈1 (13)

According to the definition of entropy, guardrail region is respectively with the three-dimensional entropy of background area

$H_F(\mathrm{\&epsiv;},\mathrm{\&xi;},\mathrm{\&eta;})=-\underset{x=0}{\overset{\mathrm{\&epsiv;}}{\&Sigma;}}\underset{y=0}{\overset{\mathrm{\&xi;}}{\&Sigma;}}\underset{z=0}{\overset{\mathrm{\&eta;}}{\&Sigma;}}\frac{p_{xyz}}{P_F(\mathrm{\&epsiv;},\mathrm{\&xi;},\mathrm{\&eta;})}ln\left(\frac{p_{xyz}}{P_F(\mathrm{\&epsiv;},\mathrm{\&xi;},\mathrm{\&eta;})}\right)---\left(14\right)$

$H_B(\mathrm{\&epsiv;},\mathrm{\&xi;},\mathrm{\&eta;})=-\underset{x=\mathrm{\&epsiv;}+1}{\overset{255}{\&Sigma;}}\underset{y=\mathrm{\&xi;}+1}{\overset{255}{\&Sigma;}}\underset{z=\mathrm{\&eta;}+1}{\overset{255}{\&Sigma;}}\frac{p_{xyz}}{P_B(\mathrm{\&epsiv;},\mathrm{\&xi;},\mathrm{\&eta;})}ln\left(\frac{p_{xyz}}{P_B(\mathrm{\&epsiv;},\mathrm{\&xi;},\mathrm{\&eta;})}\right)---\left(15\right)$

The general three entropy of F (j) is

H (ε, ξ, η)=H_F(ε,ξ,η)+H_B(ε,ξ,η) (16)

Principle is split, it is possible to make above formula H (ε, ξ, η) obtain the tlv triple (ε of maximum by maximum entropy^*,ξ^*,η^*) it is exactly institute The optimal segmenting threshold asked, i.e.

$({\mathrm{\&epsiv;}}^{*},{\mathrm{\&xi;}}^{*},{\mathrm{\&eta;}}^{*})=\underset{(\mathrm{\&epsiv;},\mathrm{\&xi;},\mathrm{\&eta;})}{\mathrm{argmax}}\left\{H\right(\mathrm{\&epsiv;},\mathrm{\&xi;},\mathrm{\&eta;}\left)\right\}---\left(17\right)$

According to threshold value (ε^*,ξ^*,η^*) F (j) is carried out binarization segmentation, guard rail region is set to 1, background area is set to 0, i.e.

Step 5. guardrail based on run-length encoding detects

According to maximizing entropy segmentation principle, the guardrail in railway scene and background positions can be positioned, represent at F (j)=1 The position of guard rail, represents the position of background at F (j)=0.As it is shown in figure 5, white portion (value is 1) represents guard rail, and black Territory, zone (value is 0) represents background.The two-value code of a succession of 0 and 1 composition obtained according to threshold segmentation method is carried out stroke Coding.

If there is defect in guardrail, pixel wide D of the background area between the most adjacent two guard rails_curIt is naturally larger than protection Normal pixel spacing d between hurdle, if therefore D_cur> k d (k is that experience regulates parameter), then can be determined that this region exists and lack Damage.

Another embodiment of the present invention is below described.

Embodiment 1: a kind of new Panoramagram montage method based on propulsion video virtual rectangle sampling channel model.

The acquisition of railway panorama sketch comprises three below step: the collection of forward video, structure splicing regions, band are spelled Connect.

Shown in the following algorithm of algorithm 1 of virtual rectangle sampling channel model generation railway scene panorama sketch.

As shown in Figure 6, the panorama sketch knot that the low-quality video (720 × 576) gathered for (150km/h) under high-speed condition generates Really.The method that the present invention proposes can generate gratifying panoramic picture.As letter do not lost by guard rail nearby and electric pole Breath and distortion are less.Electric pole at a distance significantly stretches distortion owing to " over-sampling " there occurs, but this is not in actually detected The part being concerned about.

Embodiment 2: guardrail based on Panoramagram montage detects.

As it is shown in figure 9, the propulsion video image gathered for comprehensive detection train.Train keeps the most at the uniform velocity, shooting The acquisition frame rate of machine is 25 frames/second.Fig. 9 (a) is the railway scene of guard rail N/D, and Fig. 9 (b) is the rail yard that there is defect Scape.

Part railway panorama sketch (left side) generated is as shown in Figure 10 (a), and Figure 10 (b) is the portion extracted from Figure 10 (a) Divide guard rail panorama sketch (i ∈ [405,490], j ∈ [2500,3000], wherein, i is the row-coordinate of pixel, and j is row coordinate).

As follows according to the guard rail guard rail location algorithm maximizing entropy segmentation principle realization:

Figure 11 be the stereogram that proposes of the present invention maximize entropy split-run and one-dimensional grey level histogram split-run and The comparing result of two dimensional gray-histogram of gradients split-run.Figure 11 (a) is for divide according only to gray scale one dimensional histograms maximum entropy threshold That cut as a result, it is possible to observe that many background areas are all guard rail by erroneous segmentation, this will cause detection to be failed to report.Figure 11 (b) For split according to Gray Level-Gradient (GLGM) two-dimensional histogram maximum entropy threshold as a result, it is possible to observe row coordinate 2770 position The guard rail at place is background by erroneous segmentation, and this will cause detection wrong report.Figure 11 (c) is the three-dimensional Nogata proposed based on the present invention Figure segmentation result, by compared with the guardrail physical location shown in Figure 10 (b), it can be seen that the knot of stereogram segmentation Fruit is the most accurate.

After extracting the position of guardrail, two-value code is carried out run-length encoding, to reach the lightweight storage of guardrail information With access purpose.Algorithm 3 is used quickly to identify the damage location of guard rail from guardrail panorama sketch run-length encoding afterwards.Such as Figure 11 Shown in (c), there is the region of obvious guardrail defect at two in the pixel coverage of [2500,3500], lay respectively at interval [3000,3138] and [3290,3462] place.

Experiment video be avi form, time be about 1 hour, data volume is 1.35GB, and sum of all pixels is (1024 × 768)/frame × 85500 frames.Utilizing panoramic mosaic to generate the image of the jpg form that data volume is 84MB, sum of all pixels is 267500 × 600 (sum of all pixels of the sum of all pixels × vertical direction of horizontal direction).By the foreground and background region of guardrail in panorama sketch only with single The numeral of byte represents, finally gives 8917 position digital codings (37 3 41 3 33 2 42 3 ...), and data volume is only 17.4KB.Therefore, the data volume that present invention achieves railway guardrail is extracted from the multi-stage compression of GB level MB level KB level, Overcome because the video data volume is big, informationm storage and retrieval difficulty and the application bottleneck of detection algorithm that causes.

The present invention gives the algorithm of defect location based on panorama sketch run-length encoding detection guard rail guard rail.

Owing to the traveling stroke of train is the longest, the illumination of whole guard rail panorama sketch because of by weather, environment, towards etc. factor Impact there will be change, therefore, it is necessary to guard rail panorama sketch is divided into K subgraph, and it is suitable to calculate one for every width subgraph Local threshold (ε_i,ξ_i,η_i), i=1 here, 2,3 ..., K.

Detected by scene repeatedly artificial visual, determine defect at the guardrail physical presence 38 in this section of circuit, be denoted as GT =38.Testing result contrasts based on different threshold segmentation methods as shown in table 1, are contrasted by the real conditions with guardrail defect Finding, the method that the present invention proposes correctly has detected defect at 36 (TP=36)；There is at 3 normal guard rail position by mistake Inspection be defect (FP=3)；The breakage missed at two is had not to be detected (FN=2).Used here as Detection accuracy and recalling Rate verifies the effectiveness of detection method, and the threshold segmentation method proposed based on the present invention in testing result contrasts achieves The accuracy rate of 92.3% and the recall rate of 94.7%, be better than the testing result of other partitioning algorithm.

The testing result contrast of the different threshold segmentation method of table 1

In order to verify the suitability of the maximum entropy threshold dividing method combining multiple features MVG stereogram further, I Again the cement guard rail of the another kind of guard rail of high-speed railway white is carried out same experiment.The cement protection of white Hurdle is different from the irony guardrail of the color and luster dimness shown in Figure 10 (b), and the most generally, the brightness in guard rail region is on the contrary Higher than background area.As shown in Figure 12 (d), for the segmentation result of the guard rail panorama sketch of high-speed railway, by with figure Guard rail panorama sketch in 12 (a) compares, it appeared that the partitioning algorithm that the present invention proposes is either for metal protection hurdle Or cement guard rail, all achieves gratifying segmentation result.

Figure 13 is the flow chart of steps of high-speed railway guard rail defect method for quick.

Wherein, a kind of high-speed railway guard rail detection method based on propulsion video panorama, comprise the following steps:

According to rail track priori, behind propulsion video virtual rectangle sampling channel position, location, propulsion is regarded Frequently frame be divided into sky, left and right four parts, and respectively four parts are carried out Panoramagram montage.It is to say, with panorama map grid Video data is carried out lossless information extraction by formula, has obtained lightweight detection image information, has not only dropped relative to video data Low storage and access expense, and video is converted into a kind of shape being more suitable for artificial inspection or computer analyzing and processing Formula.

Panoramagram montage based on propulsion frame of video calculates, and relies on geometry priori and the motion of camera of scene Velocity information, quickly obtains the position of sampling ring, it is achieved the quick alignment between stitching image, generates left and right guard rail panorama Figure, its overall display effect is substantially better than the L-K optical flow method of characteristic matching.It is to say, utilize, propulsion video is virtual to be adopted Sample channel pattern, overcomes the static fuzzy of propulsion video panorama arrowband splicing, and utilizes the elder generation of region geometry structure Test, build the panoramic mosaic algorithm of quick propulsion video.

The maximum entropy threshold dividing method that left and right guard rail panorama sketch uses MVG stereogram is split, segmentation Go out the foreground and background region of guardrail in panorama sketch, it is achieved being automatically positioned of guard rail position.To extracting guard rail position two Value code carries out run-length encoding, it is achieved that lightweight storage and the access of guardrail information, according to algorithm 3 from the stroke of guardrail panorama sketch Coding quickly identifies the position on damage protection hurdle.

It is to say, the automated detection method of guardrail defect based on guardrail panorama sketch (Fence panorama), profit The position of vertical guard rail in guardrail is automatically extracted with maximum entropy threshold dividing method based on MVG stereogram so that it is with Background image separates；And the run-length encoding compression of the two-value code after guardrail panorama sketch is separated, the coded format of compression contains Whole positional informationes of guardrail, corresponding decoding algorithm recovers the position of guardrail from coding, and achieves the disappearance inspection of guardrail Survey.

The method have the advantages that

(1) a kind of new Panoramagram montage method that the present invention proposes, the method is a kind of empty based on propulsion video Intend the Panoramagram montage method of sampling channel model, filled up camera optical axis and regarded with the propulsion under direction of motion uniform condition Frequently the blank of panorama picture formation；Can be widely applied to the automatic Pilot of the vehicle such as track traffic, highway, safety inspection, The driving recording of high compression ratio, and the detection of rail in high speed railway glacing flatness, foreign body intrusion detection, the inspection of Along Railway landslide Survey.Owing to the sound barrier used in some location is for lowering surrounding enviroment noise jamming, there is the merit of guard rail simultaneously Can, but bigger difference is physically being had with guard rail, therefore can also be different according to the design of different types of guard rail Damage condition detection algorithm.

(2) detection method of the automatic visual of the panorama sketch based on propulsion video of the present invention, regarding magnanimity The panorama sketch form of frequency lightweight according to this carries out lossless information extraction, not only reduces storage and the visit of video data form Ask expense, and video be converted into a kind of being more suitable for and manually inspect and the form of computer analyzing and processing, solve by Propulsion video is to the undistorted rapid translating of panorama sketch.Save greatly it addition, store video information with the form of panorama sketch The memory space of amount, replaces propulsion video image as detection object using the form of panorama sketch, can realize high-speed railway envelope The quick detection in closed loop border, meets actual demand.

(3) video capture of propulsion obtains mode, owing to its visual field is broad, the depth of field as the scene of a kind of movable type Far, space broad covered area, it is widely used in scene record and the monitor task of movable type.The present invention is directed to high ferro envelope Close running environment state-detection, it is proposed that a kind of video based on propulsion, utilize the high-speed iron of computer vision technique The road automatic method for quick of guard rail damage condition, for utilizing high speed comprehensive detection train to complete high-speed railway closed operation Ambient condition detection provides a kind of new efficient means.

(4) present invention propulsion video virtual sampling channel model based on detection region geometry, before overcoming The static fuzzy spliced to sport video arrowband, and utilize the priori of region geometry structure, build quick propulsion video Panoramic mosaic algorithm, with railway protective hurdle defect detection for applied research background, it is proposed that accordingly based on panorama sketch from Dynamicization visual detection algorithm.High-speed railway environment measuring is regular a, job for normality, the most repeatedly obtains ambient video figure Picture.The information of abnormality is obtained by feature detection or scene image comparison.The present invention is directed to high-speed railway environment measuring carry Go out a kind of new splicing technology of panorama drawing based on high-speed forward sport video.

(5) present invention virtual rectangle based on propulsion video sampling channel Panoramagram montage method, achieves first Camera optical axis generates with the propulsion video panorama under direction of motion uniform condition；Use virtual rectangle sampling channel side Formula, overcomes propulsion video panorama static distortion problem；Based on propulsion video virtual rectangle sampling channel complete Scape figure generates, and does not has particular/special requirement, the video of general drive recorder record can be spliced into multiple different use for equipment Panorama sketch on the way.

Claims (8)

1. a Panoramagram montage method based on virtual sampling channel model, it is characterised in that including:

Obtain the propulsion video of record high-speed railway ambient condition；Extracting frame number from described propulsion video is N's Sequence of video images；

According to the railway scene structure determined via end point, every frame video image sets external sampling rectangle OR_m；

According to the speed of train, every frame video image sets internal sample rectangle IR_m；

By by the described external sampling rectangle OR of every frame video image_mWith described internal sample rectangle IR_mThe rectangular ring of composition is adopted Sample annulus, is divided into four strips mosaic region S_t,S_b,S_l,S_r；

By described four strips mosaic region S of every frame video image_t,S_b,S_l,S_rBy image volume around, be corrected to rule square Shape strips S_t',S_b',S_l',S_r'；

By 4 × N number of corrected described shape of rectangular ribbon S_t',S_b',S_l',S_r', split is carried out respectively according to respective sample plane, Generate the panorama sketch of 4 planes of railway scene.

Method the most according to claim 1, it is characterised in that determine the step of described end point particularly as follows:

End point coordinate in image coordinate system is (x₀,y₀)^T；The analytic expression of i-th line section is x+k_iy+b_i=0, wherein, k_i For the slope of i-th line section, b_iFor intercept；Weight w_iIt is the length of i-th line section；N is total row of pixel in guard rail panorama sketch Number；

$x_0=\frac{g_1{m}_{22}-g_2{m}_{12}}{m_{11}{m}_{22}-m_{12}^2},y_0=\frac{g_2{m}_{11}-g_1{m}_{12}}{m_{11}{m}_{22}-m_{12}^2};$

Wherein,

$m_{11}=\underset{i=1}{\overset{N}{\&Sigma;}}\frac{w_i}{1+k_i^2},m_{12}=\underset{i=1}{\overset{N}{\&Sigma;}}\frac{w_i{k}_i}{1+k_i^2},m_{22}=\underset{i=1}{\overset{N}{\&Sigma;}}\frac{w_i{k}_i^2}{1+k_i^2},$

$g_1=-\underset{i=1}{\overset{N}{\&Sigma;}}\frac{w_i{b}_i}{1+k_i^2},g_2=-\underset{i=1}{\overset{N}{\&Sigma;}}\frac{w_i{k}_i{b}_i}{1+k_i^2}.$

3. the defect of the guard rail of a detection high-speed railway based on the arbitrary described Panoramagram montage method of claim 1-2 The method of state, it is characterised in that including:

Step one, obtains the propulsion video of record high-speed railway ambient condition；

Step 2, Panoramagram montage method based on virtual sampling channel model, is railway by described propulsion video conversion The left and right panorama sketch of running environment, as guard rail panorama sketch；

Step 3, according to maximizing entropy segmentation principle, carries out determining to the position of the guardrail in described guard rail panorama sketch and background Position, represents the position of guard rail, represents the position of background at F (j)=0 at F (j)=1, obtain the two-value code of 0 and 1 composition；J is The row sequence number of guard rail panorama sketch；

Step 4, carries out run-length encoding to described two-value code；

Step 5, according to described run-length encoding, calculates pixel wide D of background area between adjacent two guardrails_cur；

Step 6, according to pixel wide D of the background area between described adjacent two guardrails_curBetween described adjacent two guardrails Normal pixel spacing d, it is judged that whether described adjacent two guardrails exist defect, generate judged result；

Step 7, exports described judged result.

Method the most according to claim 3, it is characterised in that described step 6 includes:

Work as D_curDuring more than k d, the most described judged result is: described adjacent two guardrails exist defect；Wherein, k is experience regulation Parameter.

Method the most according to claim 3, it is characterised in that described step 2 includes:

Extracting frame number from described propulsion video is the sequence of video images of N；

According to the railway scene structure determined via end point, every frame video image sets external sampling rectangle OR_m；

According to the speed of train, every frame video image sets internal sample rectangle IR_m；

By every frame video image by described external sampling rectangle OR_mWith described internal sample rectangle IR_mThe rectangular ring of composition is adopted Sample annulus, is divided into four strips mosaic region S_t,S_b,S_l,S_r；Wherein, S_t,S_b,S_l,S_rRespectively sky, rail, left side is protected Hurdle and guardrail region, right side；

By described four strips mosaic region S of every frame video image_t,S_b,S_l,S_rBy image volume around, be corrected to rule square Shape strips S_t',S_b',S_l',S_r'；

By 2 × N number of corrected described shape of rectangular ribbon S_l',S_r', carry out split according to respective sample plane respectively, generate ferrum The left and right panorama sketch of road scene.

Method the most according to claim 5, it is characterised in that determine the step of described end point particularly as follows:

End point coordinate in image coordinate system is (x₀,y₀)^T；The analytic expression of i-th line section is x+k_iy+b_i=0, k_iIt is i-th The slope of bar line segment, b_iFor intercept；Weight w_iIt is the length of i-th line section；N is total columns of pixel in guard rail panorama sketch；

$x_0=\frac{g_1{m}_{22}-g_2{m}_{12}}{m_{11}{m}_{22}-m_{12}^2},y_0=\frac{g_2{m}_{11}-g_1{m}_{12}}{m_{11}{m}_{22}-m_{12}^2};$

Wherein,

$m_{11}=\underset{i=1}{\overset{N}{\&Sigma;}}\frac{w_i}{1+k_i^2},m_{12}=\underset{i=1}{\overset{N}{\&Sigma;}}\frac{w_i{k}_i}{1+k_i^2},m_{22}=\underset{i=1}{\overset{N}{\&Sigma;}}\frac{w_i{k}_i^2}{1+k_i^2},$

$g_1=-\underset{i=1}{\overset{N}{\&Sigma;}}\frac{w_i{b}_i}{1+k_i^2},g_2=-\underset{i=1}{\overset{N}{\&Sigma;}}\frac{w_i{k}_i{b}_i}{1+k_i^2}.$

Method the most according to claim 1, it is characterised in that described step 3 includes:

For every string j of guard rail panorama sketch, calculate gray average M (1, j), standard deviation V (1, j) and gradient mean value G (1, j)；J=1 ..., N；N is total columns of pixel in guard rail panorama sketch；

Use F (j) maximum entropy, calculate segmentation threshold (ε^*,ξ^*,η^*)；F (j) is the vertical side extracted in described guard rail panorama sketch The upwards feature distribution of each column pixel；

According to described segmentation threshold (ε^*,ξ^*,η^*), F (j) is carried out binarization segmentation, guard rail region is set to 1, background area It is set to 0；

Wherein,；M (1, j) it is the gray average of each column pixel in guard rail panorama sketch；

V (1, j) it is the gray variance of each column pixel in guard rail panorama sketch；G (1, j) it is each column pixel in guard rail panorama sketch Gradient mean value.

Method the most according to claim 7, it is characterised in that

$M(1,j)=\frac{1}{b_h-t_h}\underset{i=t_h}{\overset{b_h}{\mathrm{\&Sigma;}}}p(i,j);$

$V(1,j)=\frac{1}{b_h-t_h}\underset{i=t_h}{\overset{b_h}{\&Sigma;}}{\left(p\right(i,j)-M(1,j\left)\right)}^2;$

$G\left(1,j\right)=\frac{1}{t_h-b_h}\underset{i=b_h}{\overset{t_h}{\mathrm{\&Sigma;}}}\sqrt{{\left(p\left(i+1,j\right)-p\left(i,j\right)\right)}^2+{\left(p\left(i,j+1\right)-p\left(i,j\right)\right)}^2};$

Wherein, (i j) is (i, gray value j), the t of pixel in guard rail panorama sketch to p_h,b_hIt is respectively top and the end of guard rail Portion's coordinate position in guard rail panorama sketch.