CN1581209A - Converting station circular/ellipsoid/square instrument monitoring method based object profile - Google Patents

Abstract

The present ivnention relates to a circular/elliptic/sequence meter monitoring method based on conout profile of body in transformer substation, belonging to the field of transformer substation meter monitoring technology. It is characterized by that its total control terminal computer has a remoter meter detection/fetch module capable of implementing circular/elliptic/square search and detection algorithm based on contour profile of body, and said monitoring method includes the following steps: utilizing wireless local area network to control patrol inspection robot in the transformer substation, according to predefined patrol inspection line searching the space position of the meter locked by subprogram in preset search space range, then using visible light camera head to obtain the geometrical image of the motor, then transferring said image into total control terminal, makin data processing so as to attain the goal of implementing remote monitoring and control.

Description

Based on circle/ellipse in the transformer station of contour of object/square instrument monitoring method

Technical field

The invention belongs to transformer station's instrument monitoring technical field based on Computer Image Processing.

Background technology

Divide according to application point, the present invention mainly solves the problem that reads of far distance instrument.Address this problem relevant patent and exist, but its method mainly is by special electricity/light path being set on instrument or directly by manually reading, being delivered to far-end by electricity/optical cable or wireless telecommunications then.For example:

Publication number in 1993 is " remote Meter reading " of GPT Ltd. (England) of 1073008, it is connected instrumentation tap to be measured place by a pair of optical fiber, arrive sensor by reflected light path and be delivered to far-end then, " radio-frequency (RF) repeater of instrument automatic Car Plate Reading System " (RF REPEATER FORAUTOMATIC METER READING SYSTEM) that to also have US publication in 1998 be CA2292695 similarly; Publication number in 2004 is " based on the far distance instrument reading system and the method for mobile communication " of LG Electronics Inc. (Korea S) of 1477540, it mainly is an instrument reading system of finishing transmission/collection user's metrical information by wireless network, the detection of really not carrying out instrument with read " utilizing public broadcasting channel to carry out the automatic instrument reading system " (Automatic meterreading system employing common broadcast command channel) that to also have US publication in 2004 be US6684245 similarly.Though the former can solve the problem that far distance instrument reads, need transform instrument, for this hazardous environment of transformer station and the situation that requires various device to run without interruption continuously, any transformation all is difficult to carry out relatively.

The present invention adopts artificial intelligence technology, utilizes the visible image capturing head, does not need environment and equipment are done any transformation, does not also bring any pollution.Identification is analyzed and detected to means by image recognition to current images acquired, near the instrument of the appointment FEEDBACK CONTROL camera finds, and read; By WLAN, the meter diagram that reads the result and collect is looked like to pass back remote console and deposits database at last.

Technologically speaking, the present invention is primarily aimed at circle/square instrument and detects, and only reads the situation (return results is the pointer angle) of having only a pointer in the dial plate.The present invention has adopted the circle/ellipse/rectangle detection method based on contour of object, does not find similar detection method at present, belongs to achievement in research voluntarily.Other method that detects about circle/ellipse has: the Hough conversion, determine the descriptive geometry parameter according to unique point in the parameter space ballot; Various improvement algorithms on Hough conversion basis, the cutting parameter space detects, quick Hough conversion (N.Guil, E.L.Zapata, Lower Order Circle and Ellipse HoughTransform, J.Pattern Recognition vol.30, no.10, pp.1729-1744, October (1997)), random Hough transformation (M.A.Fischler, R.C.Bolles.Random Sample Consensus:A Paradigm for Model Fitting with Applications to ImageAnalysis and Automated Cartography.Comm.of the ACM, Vol.24, pp.381-395,1981) etc.; Carry out the match of straight-line segment, camber line from unique point, iterate until find target (Euijin KIM, Miki Haseyama, Fast and Robust EllipseExtraction from Complicated Images, ICITA2002); Also have in addition and directly utilize circle and oval symmetry, detect (Qin-Zhong Ye according to various histograms, A preprocessing method for Hough Transform to Detect circles, Proc.IEEEConf.Computer Vision andPattern Recognition, pp.651-653 (1986)) etc.Because the regularity of Hough conversion, its calculated amount is exponential form along with the parameter space growth to be increased, be difficult to satisfy the requirement that detects real-time for ellipse, and circular instrument is because the difference of observation visual angle only relies on the detection of circle not meet the demands.

The characteristics of the present invention on technical method are, utilize this characteristic of continuity of contour of object in the real scene, the detection of entire image is decomposed into the detection of several successive profile subregion, simplified the Hough conversion effectively, thereby obviously reduced the complicacy of algorithm, satisfied the requirement of real-time detection; On the basis that reads in the instrument area detection result of pointer, adopt the method for angular histogram to carry out straight-line detection in the dial plate zone and finish, have good robustness.

Summary of the invention

The object of the present invention is to provide a kind of method based on computer vision, do under the prerequisite of any change in the equipment instrument that does not need to transformer station this hazardous environment and operation continuously, utilize artificial intelligence technology, employing pattern-recognition means realize the automatic detection of far distance instrument truly and read.

Crusing robot (seeing accompanying drawing 8) is advanced automatically according to predetermined route, according to the mission requirements of patrolling and examining (for now, for example, whether checkout equipment sound normal, instrument reads, temperature detection), patrol and examine at this corresponding task anchor point can be set on the route, crusing robot arrives the function sub-modules that predetermined anchor point position halts and calls far distance instrument and detect/read, wait for that this submodule carries out its appointed task after the order that the task of receiving is finished/finished, dolly moves on.

The present invention in crusing robot total system design the position and see accompanying drawing 1 alternately.The invention is characterized in that it may further comprise the steps successively:

(1) in database, deposit far distance instrument detection/reading submodule in, wherein preestablish following parameters and formula:

Crusing robot carries out instrument at needs and detects to make an inspection tour on the line with the equipment that reads and can be clear that the anchor point position of equipment instrument, the inceptive direction of this position camera The Cloud Terrace and the presetting bit of camera lens;

The information of this anchor point place instrument---the center of dial plate, image, dial plate size, instrument record scale reduction formula

The dial plate background is the black/white dichromatism;

The scope of crusing robot search volume: road plane is stopped error ± 15cm, car body declination error ± 30 of crusing robot °;

The threshold value T1=50 that contour feature is counted,

Match factor threshold value T2=0.75,0＜T2＜1 wherein,

(2) overhead control end computer control crusing robot arrives instrument and detects and the anchor point that reads, and reads corresponding this anchor point information, calls and starts above-mentioned far distance instrument detection/reading submodule;

(3) computing machine arrives assigned address according to above-mentioned anchor point information by WLAN (wireless local area network) control camera and camera lens based on ICP/IP protocol;

(4) computing machine receives the image sequence that crusing robot is beamed back by WLAN (wireless local area network), according to this anchor point meter information, in the search volume of regulation, call following search subroutine, control The Cloud Terrace and camera lens carry out the dial plate searching and detecting, and this search subroutine may further comprise the steps successively:

(4.1) set following numerical value:

#define SEARCHSTEPX horizontal direction search step number,

#define SEARCHSTEPY vertical direction search step number,

#define SEARCHSTEPZ depth direction search step number,

#define STEPX horizontal direction step-size in search,

#define STEPY vertical direction step-size in search,

#define STEPZ depth direction step-size in search,

The order of turning left of #define CMD_LEFT The Cloud Terrace,

The order of turning right of #define CMD_RIGHT The Cloud Terrace,

#define CMD_UP The Cloud Terrace upwards changes order,

#define CMD_DOWN The Cloud Terrace changes order downwards,

#define CMD_ZOOMIN cam lens draws order forward,

#define CMD_ZOOMOUT cam lens pulls back order,

The dial plate size that #define GOODSIZE is suitable is represented with pixel number,

#define IMAGEWIDTH picture traverse,

#define IMAGEHEIGHT picture altitude,

The range size of #define CENTERWIDTH picture centre,

The state variable of the current working direction of mark camera lens:

IsForward for very, represents that current camera lens furthers forward, otherwise for zooming out, pre-sets isForward=TRUE;

IsRight for very, represents the deflection to the right forward of current camera lens, otherwise for deflection left, pre-sets isRight=TRUE;

Write down current lens location P1=P2=(0,0,0), wherein, P1 (x, y, z) mark camera lens current location, P2 (x, y, z) searching position that set the goal last time of mark camera lens;

(4.2) detect present image:

If detect panel board, just carry out according to the following steps:

(4.2.1) preserve lens location P2;

(4.2.2) if the position of dial plate center and picture centre is inconsistent, computing machine earlier the adjustment camera lens about, upper-lower position, detect present image more again;

(4.2.2) as if the position consistency of dial plate center and picture centre, computing machine judges just whether the dial plate size is suitable;

(4.2.3) if off size suitable, computing machine is adjusted the front and back position of camera lens earlier, detects present image more again;

(4.2.4) if size is suitable, computing machine is lock onto target just, and program is returned reading, sends the finish command; If detect less than panel board, the computer control camera lens is return last position of preserving, and then carries out according to the following steps:

(4.2.5) computing machine furthers camera lens earlier, detects present image more again; If still detect less than, again camera lens is zoomed out, detect present image again, if still detect less than, just carry out next step;

After (4.2.6) computing machine is moved camera lens to proximal most position, respectively the The Cloud Terrace position is moved to the rightlyest and the most left again, detect present image again, if still detect, just carry out next procedure less than present image;

(4.2.7) computing machine moves to the position, highest and lowest to the The Cloud Terrace position respectively again, detects present image again, if the search failure sends the finish command;

(4.2.8) computing machine zooms out camera lens again, and repeating step (4.2.5) is to (4.2.8);

(5) behind the target lock-on, computing machine makes it obtain instrument to be detected suitable image size and size according to the testing result controls lens, and the reading pointer angle also is converted into the dial plate reading, beams back by WLAN (wireless local area network); For the situation that can not read, by WLAN (wireless local area network) suitable image sequence is beamed back demonstration, by manually reading concurrent crusing robot " task is finished " signal of giving; If the target search failure also sends the detection failure information by WLAN (wireless local area network) to computing machine and issues crusing robot " task is finished " signal simultaneously;

(6) after computing machine receives the video image that camera lens gathers, just start detection program, handle video image according to the following steps:

(6.1) pre-service, it may further comprise the steps successively:

(6.1.1) video image of input is converted into gray level image by following formula:

Each gray values of pixel points

$I=\frac{R+B+G}{3},$

Wherein, R, G, B are respectively the brightness value of red, green, blue three looks;

(6.1.2) according to the following steps gray level image is carried out medium filtering:

Earlier (x y), puts new brightness value to it as this by the brightness value of 8 points around each pixel of COMPUTER CALCULATION and the mean value I of this original brightness value

$I(x,y)=\frac{\underset{i=-1}{\overset{1}{\mathrm{\Σ}}}\underset{j=-1}{\overset{1}{\mathrm{\Σ}}}I(x+i,y+j)}{9},$

With known method grey level histogram is done equalization then and handle, the concrete principle and the algorithm of histogram equalization are known;

(6.1.3) the Flame Image Process function library of usefulness open source code, be that OPENCV carries out the Canny rim detection to above-mentioned gray level image, obtain the bianry image of same size, wherein, the pixel that contains marginal information is white, other zone is a black, and these white pixels are called the edge feature point;

(6.1.4) close (expanding earlier, again corrosion) operating function with the morphology in the OPENCV function library again, above-mentioned bianry image is handled, obtain the bianry image that the continuous as far as possible available outline line of same object edge image is represented;

(6.2) carrying out geometric figure according to the following steps detects:

(6.2.1) all outline lines of above-mentioned entire image are encoded by the chain code form, obtain a profile chained list;

Be every the outline line cutting that surpasses T1 of counting that comprises independent subimage (6.2.2), remove remaining outline line;

(6.2.3) utilize the RANSAC method on each contour images, i.e. stochastic sampling ballot figure fitting process carries out the match of assignment graph:

(6.2.3.1) match of standard round:

Central coordinate of circle (x0, y0):

$x_0=\frac{d_{y1}{d}_{y2}{d}_{y3}-(x_1^2{d}_{y2}+x_2^2{d}_{y3}+x_3^2{d}_{y1})}{2(x_1{d}_{y3}+x_2{d}_{y1}+x_3{d}_{y2})},$

$y_0=\frac{d_{x1}{d}_{x2}{d}_{x3}-(y_1^2{d}_{x2}+y_2^2{d}_{x3}+y_3^2{d}_{x1})}{2(y_1{d}_{x3}+y_2{d}_{x1}+y_3{d}_{x2})};$

Radius $R=\sqrt{{(x_0-x_1)}^2+{(y_0-y_1)}^2};$

P wherein _i(x _i, y _i) (i=1,2,3) be three unique points on the outline line,

d _xi＝(x _i-x _(i+1)％3)，

d _yi＝(y _i-y _(i+1)％3)，

Subscript " %3 " is represented the remainder divided by 3;

(6.2.3.2) general oval match:

Oval have following expression formula as general quafric curve:

a ₀x ²+2a ₁xy+a ₂y ²+2a ₃x+2a ₄y＝1，

Coefficient has wherein passed through normalized, and the parameter on the right is changed to 1, does not consider the situation through initial point;

Five unique point P on the contouring line _i(x _i, y _i) (i=1 ..., 5), the structure system of linear equations:

A＝P ^-1B，

Wherein,

A＝[a ₀?a ₁?a ₂?a ₃?a ₄] ^T，B＝[1?1?1?1?1] ^T，

$P=\left[\begin{array}{ccccc}{x_1}^2& 2x_1{y}_1& {y_1}^2& 2x_1& 2y_1\\ {x_2}^2& 2x_2{y}_2& {y_2}^2& 2x_2& {2y}_2\\ {x_3}^2& 2x_3{y}_3& {y_3}^2& 2c_3& 2y_3\\ {x_4}^2& 2x_4{y}_4& {y_4}^2& 2x_4& 2y_4\\ {x_5}^2& 2x_5{y}_5& {y_5}^2& 2x_5& 2y_5\end{array}\right],$

Mark I ₁=a ₀+ a ₂,

$I_2=\left[\begin{array}{ccc}{a}_0& a_1& a_3\\ a_1& a_2& a_4\\ a_3& a_4& -1\end{array}\right],$

According to following conditions restriction, i.e. the elliptic parameter condition that should satisfy, can judge:

I ₁And I ₂Should satisfy:

I ₁I ₂＜0， $I_3=\left[\begin{array}{cc}{a}_0& a_1\\ a_1& a_2\end{array}\right]=a_0{a}_2-a_1^2>0;$

According to the conic section characteristic, the quafric curve general parameters is converted into the description form at elliptical center commonly used, axle, inclination angle, conversion formula is as follows:

Major axis- $a=\sqrt{-\frac{I_3}{{\mathrm{\λ}}_1{I}_2}},$ Major axis- $b=\sqrt{-\frac{I_3}{{\mathrm{\λ}}_2{I}_2}},$

λ wherein ₁And λ ₂Be matrix I ₂Eigenwert;

The major axis drift angle- $\mathrm{\θ}=\frac{1}{2}\arctan \frac{2a_1}{a_0-a_2};$

Central point

(6.2.3.3) match of rectangle:

Existing rectangle fitting algorithm based on profile obtains in the employing OPENCV function library, directly exports four fixed points of detected rectangle coordinate figure;

(6.2.4) testing result aftertreatment:

(6.2.4.1) determining of the marginal information unique point of match: whether it exists the unique point that is fit to determine by each angle on the angle of calculating 360 degree;

(6.2.4.2) judge with the method for grey level histogram whether the dial plate background color that detects is consistent with the known background color, presses following steps and carries out:

(6.2.4.2.1) create 128 ballot box VoteBox[128], ballot box is emptied zero setting;

(6.2.4.2.2) for the brightness value I (0＜=I＜=255) of each pixel in testing result zone, corresponding I/2 ballot box is with this ballot box note one number, VoteBox[I/2] ++;

(6.2.4.2.3) 128 ballot boxes of traversal, record who gets the most votes ballot box sequence number is MaxBox;

If (6.2.4.2.4) MaxBox＜32, then background is a black; If MaxBox＞96, then background is a white; Otherwise background color mistake;

(6.2.4.2.5) whether the background color of Jian Ceing is consistent with the known background color, if consistent, detects below continuing; Otherwise give up;

(6.2.4.3) pointer detects:

The angle of the unique point of statistics in the dial plate and the dial plate line of centres and image X-axis formation is made angular histogram, and according to the method for step (6.2.4.2), the angle that peak value occurs is the maximum ballot box of number of votes obtained just, is the angle at pointer place.

The present invention adopts the circle/ellipse/rectangle detection algorithm based on contour of object, effectively the geometric figure detection problem with complexity is divided into the plurality of sub problem, new detection algorithm can improve detection speed greatly under the prerequisite that does not reduce the detection accuracy, thereby satisfy the requirement of system real time, realistic demands of applications.

The concrete detection algorithm time performance contrast table of the present invention is as follows:

	Time (ms)
		Pre-service	????10
Profile extracts	????＜5
		The RANSAC match	????50

Table 1. each several part average detected

Method	The Hough circle detects fast	The present invention's circle detects	Hough is oval fast detects	The present invention is oval to be detected
					Test picture 1	??63(ms)	31(ms)	????Failed	??230(ms)
Test picture 2	??63(ms)	31(ms)	????Failed	??94(ms)

The contrast of table 2. time performance

Description of drawings:

Fig. 1. the invention module is positioned at the structural representation (dark color is labeled as the part that the present invention relates to) of total system

Fig. 2. the search synoptic diagram: the space is divided into the search dot matrix of 3*3, and initial position is positioned at the position No. 0, for avoiding the camera repeat search, the space each point searched in turn by above-mentioned numbering, the arrow signal direction of advancing and reconnoitering, part is not drawn; Invisible part does not mark sequence number yet

Fig. 3. image detection is picture group as a result

(3.a) test picture 1 circle testing result

(3.b) test picture 1 oval testing result

(3.c) test picture 2 circle testing results

(3.d) test picture 2 oval testing results

Fig. 4. chain representation synoptic diagram

(4.a) method for expressing of chain code; ' * ' mark flex point

(4.b) chain code direction dictates mark

Fig. 5. the search strategy process flow diagram

Fig. 6. pointer extracting is figure as a result, the result of solid white line for detecting

Fig. 7. the image pre-service is picture group as a result

(7.a) gray level image is through the medium filtering denoising

(7.b) histogram equalization adapts to light and changes

(7.c) the Canny operator detects and obtains edge image

(7.d) through the edge image after the closed operation

Fig. 8. crusing robot is at the on-the-spot outside drawing of transformer station

Fig. 9. the detection algorithm process flow diagram

Embodiment

1. system hardware structure

Total system is by carrying out communication by the WLAN (wireless local area network) based on the TCP/IP procotol between base station (overhead control end computing machine) and the movement station (crusing robot computing machine), thereby realizes long-range control and detection.

2. search utility carries out according to the following steps successively:

At first set:

#define SEARCHSTEPX horizontal direction search step number,

#define SEARCHSTEPY vertical direction search step number,

#define SEARCHSTEPZ depth direction search step number,

#define STEPX horizontal direction step-size in search,

#define STEPY vertical direction step-size in search,

#define STEPZ depth direction step-size in search,

The order of turning left of #define CMD_LEFT The Cloud Terrace,

The order of turning right of #define CMD_RIGHT The Cloud Terrace,

#define CMD_UP The Cloud Terrace upwards changes order,

#define CMD_DOWN The Cloud Terrace changes order downwards,

#define CMD_ZOOMIN cam lens draws order forward,

#define CMD_ZOOMOUT cam lens pulls back order,

The dial plate size that #define GOODSIZE is suitable is represented with pixel number,

#define IMAGEWIDTH picture traverse,

#define IMAGEHEIGHT picture altitude,

The range size of #define CENTERWIDTH picture centre.

Because the anchor point that the very difficult sometimes desirable arrival of the error of the error of GPS location, crusing robot planning and the uncertainty of vehicle body direction, crusing robot is set and the deflection of setting.Consider in this sense, realize that the search in the certain limit seems very necessary.This is another core content that the present invention is only second to detection algorithm.

The stop precision of inspection car at the anchor point place is ± 15cm at present, and the car body direction when inspection car is stopped simultaneously can not be set also and can not feed back.This is just explanation also, though detect the presetting bit that the anchor point place has set camera and The Cloud Terrace, the sightless situation of instrument still can occur, so just requires instrument monitoring that certain function of search is partly arranged, and remedies the not enough situation of precision of stopping.According to test result, determine the limits of error that to tolerate: road plane anchor point error ± 15cm, car body declination error ± 30 °.

Search strategy itself is fairly simple: with the above-mentioned deviation range of inspection car (road plane anchor point error ± 15cm, car body declination error ± 30 °) as whole search volume, difference discretize (the concrete parameter of discretize is determined by preceding faceted search step number) on three directions, promptly determine some equally distributed check points, obtain the search dot matrix of a three-dimensional like this, detect at these points by the camera order by control program and get final product.The strategy (seeing accompanying drawing 2) of representing whole search with a very simple cube structure figure.The diagram search strategy can once travel through each search point efficiently and can not duplicate the situation of search.

In addition, after the searching and detecting lock onto target, The Cloud Terrace is adjusted automatically according to testing result, and no longer progressively searches for according to search strategy.But, because transformer station's background more complicated also needs the influence of considering that the flase drop survey causes.The present invention is directed to above-mentioned various situation and carried out effective processing---the spatial point P1 of record locking target, if tracking target track rejection after a period of time, the control The Cloud Terrace comes back to this and presses search strategy and continue search, concrete steps are described below:

A. camera initially arrives preset position.Use two record camera positions trivector P1 (x, y, z), P2 (x, y, z), the former current position of mark camera wherein, the searching position during latter's mark camera lock onto target last time; And quantity of state isForward, the isRight of the current working direction of mark camera, when the former is true, represent that then the current camera lens of camera furthers forward, otherwise camera zooms out; When the latter is true, then represent the current deflection to the right of camera, otherwise camera deflection left.The above-mentioned variable of initialization, P1=P2=camera initial position; IsForward=isRight=TRUE.(program preestablishes search depth and step-length)

B. detect present image and control search:

1) if do not detect target,

If P1!=P2, camera return to P1 position, P2=P1.

According to deep search

If camera lens is furthering towards the place ahead (isForward==TRUE)

If camera lens arrived recently (P1.z＞=SEARCHSTEPZ),

Search for according to horizontal direction

If The Cloud Terrace moves right (isRight==TRUE)

That if The Cloud Terrace has arrived is the rightest (P1.x＞=SEARCHSTEPX)

Then move a step CMD_UP search on the The Cloud Terrace

If search for complete space (P1.y＞=SEARCHSTEPY) jump to C to go on foot;

The CMD_RIGHT search otherwise The Cloud Terrace moves to left

If The Cloud Terrace is moved to the left (isRight==FALSE)

If The Cloud Terrace has arrived the most left (P1.x＜=0)

Then move a step CMD_UP search on the The Cloud Terrace

If search for complete space (P1.y＞=SEARCHSTEPY) jump to C to go on foot;

The CMD_LEFT search otherwise The Cloud Terrace moves to left

The CMD_ZOOMIN otherwise camera lens furthers (as shown in Figure 2)

Otherwise, if camera lens zooms out (isForward==FALSE) towards the rear

If camera lens has arrived farthest (P1.z＜=0),

Search for according to horizontal direction

If The Cloud Terrace moves right (isRight==TRUE)

That if The Cloud Terrace has arrived is the rightest (P1.x＞=SEARCHSTEPX)

Then move a step CMD_UP search on the The Cloud Terrace

If search for complete space (P1.y＞=SEARCHSTEPY) jump to C to go on foot;

The CMD_RIGHT search otherwise The Cloud Terrace moves to left

If The Cloud Terrace is moved to the left (isRight==FALSE)

If The Cloud Terrace has arrived the most left (P1.x＜=0)

Then move a step CMD_UP search on the The Cloud Terrace

If search for complete space (P1.y＞=SEARCHSTEPY) jump to C to go on foot;

The CMD_LEFT search otherwise The Cloud Terrace moves to left

Otherwise camera lens zooms out CMD_ZOOMOUT (as shown in Figure 2)

Record current location P1 returns the B step and continues to carry out;

2) if the target of detecting:

Testing result is returned two main features---and geometric center Center (x, y), target size Size (for circle, is radius; Ellipse then is a major axis; Rectangle then is wide).Document image center point coordinate ORIGIN (IMAGEWIDTH/2, IMAGEHEIGHT/2).Record current search position is at P2.

If target take back (Center.x＜ORIGIN.x-CENTERWIDTH/2),

The control The Cloud Terrace CMD_LEFT that moves to left;

Otherwise, (Center.x＞ORIGIN.x+CENTERWIDTH/2) if target takes over

The control The Cloud Terrace CMD_RIGHT that moves to right;

Otherwise, if target (Center.y＜ORIGIN.y-CENTERWIDTH/2) on the upper side

Move CMD_UP on the control The Cloud Terrace;

Otherwise, if target (Center.y＞ORIGIN.y+CENTERWIDTH/2) on the lower side

The control The Cloud Terrace moves down CMD_DOWN;

Otherwise, if target size (Size＜BETTERSIZE) less than normal

The controls lens CMD_ZOOMIN that furthers;

Otherwise, if target size (Size＞BETTERSIZE+20) bigger than normal

Controls lens zooms out CMD_ZOOMOUT;

Otherwise target reaches the optimum position, jumps to the C step;

Returning the B step continues to carry out;

C. search finishes, and sends out the task termination signal, withdraws from.

The process flow diagram of search strategy is seen accompanying drawing 5.

3. trace routine comprises following steps successively:

Input: the video image of camera collection;

Output: have instrument? if exist, and output meter dish position Center (x, y), size Size, pointer angular readings (Pointer).

The first step, the image pre-service:

The images acquired camera obtains true color image, contains R, G, and three passages of B transfer true color image to be detected to gray level image according to detecting needs, and are as follows to the concrete reduction formula of each pixel:

$I=\frac{R+B+G}{3}--\left(3\right)$

Carry out medium filtering for gray level image, remove a part because the assorted point of the noise that image acquisition is brought; Concrete mode is that mean value by calculating the original brightness value of the brightness value of 8 points around each point and this point is as the new brightness value of this point.Computing formula is as follows:

$I(x,y)=\frac{\underset{i=-1}{\overset{1}{\mathrm{\Σ}}}\underset{j=-1}{\overset{1}{\mathrm{\Σ}}}I(x+i,y+j)}{9}--\left(2\right)$

Wherein, I (x, y) remarked pixel point (x, brightness value y).

In order to adapt to different illumination conditions, also must carry out histogram equalization and handle, thereby obtain the pre-service gray level image gray level image.Output result after the processing remains a width of cloth gray level image.

Then, utilize classical Canny rim detection (J.Canny.A Computational Approach to Edge Detection, IEEE, Trans.on Pattern Analysis and Machine Intelligence, 8 (6), pp.679-698 (1986)) marginal information of operator extraction gray level image obtains edge image.The present invention utilizes OPENCV processing built-in function to carry out the Canny rim detection, the gray level image of input, and through the bianry image of the same size of output after handling, the pixel that contains marginal information is a white, other zone is a black.The pixel that these are white is called the edge feature point.Follow-up match detects and is primarily aimed at these unique points.

Because background or The noise, the edge of same object may be split into discontinuous two sections or several sections, can utilize morphology to close (expand, corrode) computing for this reason, makes that same object edge image is continuous as much as possible.Here still use the OPENCV function library to handle, closed operation is input as the bianry image after Canny detects, and output still is bianry image, but unique point changes later on through closed operation.

The concrete result images of above-mentioned pretreated each step is seen accompanying drawing 7.

Second step, geometric detection:

The chain representation is a kind ofly to be communicated with the method for expressing of single pixel wide based on 8-, and accompanying drawing 4 has provided an example.From starting point, each step is all pointed to the next unique point of current point.The direction character marking position relation between them, see accompanying drawing 4 (b).Like this, accompanying drawing 4 (a) can be designated as 44646000002424.In order further to reduce unique point, and do not lose necessary profile information, record flex point wherein only in expression, mark in the accompanying drawing 4 (a) ' point of * '.Like this, for circle and oval the detection, can filter out the influence of long straight line rapidly by this method for expressing.

The outline line that entire image is all is encoded by this mode of chain code, obtains a profile chained list.

Below briefly introduce RANSAC (Random Sample Consensus, stochastic sampling ballot) figure approximating method.It is a kind of good fitting algorithm, and concrete steps are as follows:

1) in a sample set that contains m data element, selects n sample data randomly;

2) by n sample data drawing for estimate shape parameter, calculate then in certain scope, what have individually can be designated as K by match in all m data;

3) if K is enough big, then accept this fitting result, and utilize least square to recomputate this graphic parameter by all data on can match; Otherwise re-execute 1)-3)

4) allow number of times L if repeat number of times greater than maximum, then failure;

5) if possible contain a plurality of figures, then 3) in do not withdraw from after finding fitting result, but from m data element, deduct, in remaining sample, continue to carry out 1 by the data element of match)-3).

The RANSAC method is actually a kind of Hough conversion of simplification.

According to top explanation, the chain representation is actually with a continuous chained list and has marked a continuous edge contour.For pre-service back edge image, utilize the chain code coding, obtain a plurality of different outline lines, every the outline line cutting that surpasses given threshold value of counting that will comprise is independent subimage, removes remaining outline line; Then, on each contour images, utilize the RANSAC method to carry out the match of assignment graph (circle/ellipse/rectangle), remove the fitting result that fitting degree is lower than given threshold value;

This part has made full use of the contour of object continuity Characteristics, has solved the deficiency of existing method on detection speed well.

The 3rd step, circle and oval match:

The circle of a standard can quantic be designated as:

(x-x ₀) ²+(y-y ₀) ²＝R ²????(3)

It contains three parameters, be respectively the center of circle (x0, y0) and radius R.If three parameters above having determined have so just been determined a circle.

For the match of circle, in fact be exactly to require to obtain three top parameters by known unique point.Have the condition of unique solution to be not difficult to know according to linear equation, three not the unique point of conllinear can determine a circle.In conjunction with on regard to the introduction of RANSAC method, for the n=3 above the match of circle.Concrete calculation of parameter is according to three unique point P on outline line of simultaneous _i(x _i, y _i) i=1,2, the equation of a circle of 3} can in the hope of, be direct computing formula below:

$x_0=\frac{d_{y1}{d}_{y2}{d}_{y3}-(x_1^2{d}_{y2}+x_2^2{d}_{y3}+x_3^2{d}_{y1})}{2(x_1{d}_{y3}+x_2{d}_{y1}+x_3{d}_{y2})},$

$y_0=\frac{d_{x1}{d}_{x2}{d}_{x3}-(y_1^2{d}_{x2}+y_2^2{d}_{x3}+y_3^2{d}_{x1})}{2(y_1{d}_{x3}+y_2{d}_{x1}+y_3{d}_{x2})},$

$R=\sqrt{{(x_0-x_1)}^2+{(y_0-y_1)}^2};$

Wherein, d _Xi=(x _i-x _{(i+1) %3}), d _Yi=(y _i-y _{(i+1) %3}), %3 represents the remainder divided by 3.

Equally, the oval Algebraic Expression form of standard is as follows:

$\frac{{(x-x_0)}^2}{a^2}+\frac{{(y-y_0)}^2}{b^2}=1--\left(5\right)$

But this expression formula is only described the axis of symmetry ellipse parallel with coordinate axis, and this can not meet the demands far away in system requirements.But the expression of shape ellipse is too loaded down with trivial details arbitrarily, and it is except four top parameters, elliptical center (x0, y0) and major and minor axis a, b also have an inclined angle alpha outward.The expression formula of five parametrical nonlinearities wishes to be difficult to find the solution by the way that top circle is intended platform.The present invention avoids this difficult problem, and it is oval to adopt another quantic to describe, and promptly uses conical section, is expressed as follows:

a ₀x ²+2a ₁xy+a ₂y ²+2a ₃x+2a ₄y＝1????(6)

This expression formula was directly removed the situation of initial point with parameter normalization.It is the same so just can to justify match, by five any 3 not the unique point of conllinear obtain five parameters above the Solving Linear.

But the front was mentioned, and (6) formula is only expressed general quafric curve, so the curve of finding the solution might not be oval.According to following conditions restriction (condition that elliptic parameter should satisfy), we can judge:

Mark I ₁=a ₀+ a ₂,

$I_2=\left[\begin{array}{ccc}{a}_0& a_1& a_3\\ a_1& a_2& a_4\\ a_3& a_4& -1\end{array}\right],$

I then ₁And I ₂Satisfy:

I ₁I ₂＜0，???????????????????(7)

And $I_3=\left[\begin{array}{cc}{a}_0& a_1\\ a_1& a_2\end{array}\right]=a_0{a}_2-a_1^2>0;----\left(8\right)$

Just can obtain the parameter of general algebra form of fitted ellipse like this by above-mentioned two steps, we are converted into it center, the axle of custom, the description form at inclination angle again, and conversion formula is as follows:

$a=\sqrt{-\frac{I_3}{{\mathrm{\λ}}_1{I}_2}},b=\sqrt{-\frac{I_3}{{\mathrm{\λ}}_2{I}_2}}--\left(9\right)$

λ wherein ₁And λ ₂Be matrix I ₂Eigenwert;

$x_0=-\frac{(a_3{\cos }^2\mathrm{\θ}+a_4\cos \mathrm{\θ}\sin \mathrm{\θ})}{a_0+a_1\tan \mathrm{\θ}}+\frac{(a_3{\sin }^2\mathrm{\θ}+a_4\cos \mathrm{\θ}\sin \mathrm{\θ})}{a_2-a_1\tan \mathrm{\θ}},$

$y_0=-\frac{(a_3\cos \mathrm{\θ}\sin \mathrm{\θ}+a_4\mathrm{si}{n}^2\mathrm{\θ})}{a_0+a_1\tan \mathrm{\θ}}+\frac{(-a_3\cos \mathrm{\θ}\sin \mathrm{\θ}+a_4\mathrm{co}{s}^2\mathrm{\θ})}{a_2-a_1\tan \mathrm{\θ}},$

$\mathrm{\θ}=\frac{1}{2}\arctan \frac{2a_1}{a_0-a_2},--\left(10\right)$

About the match of rectangle, existing rectangle fitting algorithm based on profile obtains in the employing OPENCV function library, is directly output as four fixed point arrays of detected rectangle;

The part testing result is referring to accompanying drawing 3;

The 4th step, the testing result aftertreatment:

The introduction of front is known, the fitting effect that the RANSAC method is estimated figure is to investigate by the ratio of unique point match, for example, article one, contain 200 unique points on 1/4 the camber line, the RANSAC approximating method can obtain a detection circle and make that 200 unique points can both be by match, and such fitting result is very perfect.But this situation can not occur in actual detection, for example, therefore the cable of one section arc can not be considered to a possible circular dial plate, whole clearly visible because panel board necessarily requires, so require to detect whole circle or ellipse figure, rather than part camber line.Detection algorithm among the present invention proposes the notion of " the match factor ".The match factor in fact be exactly fitting result under the polar expression situation of figure, the description of the fitting degree on whole radian:

The proposition of the match factor has not only solved the problem that occurs above, and has solved the problem of the many sizes of target, has realized the normalization that fitting effect is estimated, and has strengthened the robustness that system detects.

Whether consider actual application scenarios, algorithm does not need really to be converted into polar coordinate representation, only need calculating 360 degree to go up each angle and exist edge feature point to get final product.Be implemented as follows:

Begin

Count：＝0；

For?I：＝0?to?360?do??/*?for?each?angle?*/

Compute?the?corresponding?point?P?of?I

If?P?is?a?feature?point?in?the?image

Count：＝Count+1；

Count：＝Count/360；

End；

Above-mentioned testing result only is geometric detection, does not take into full account actual application scenario.Consider the monotonicity of dial plate background color, the dial plate of practical scene is divided into two big class---white background and black backgrounds.Can brightness statistics be carried out in each testing result zone make grey level histogram on gray level image under the prerequisite of known background color like this, if the peak value of grey level histogram is consistent with known background color, testing result is accepted, otherwise gives up.

Be implemented as follows:

1. create 128 ballot box VoteBox[128], ballot box is emptied zero setting:

2. for the brightness value I of each pixel in testing result zone, corresponding I/2 ballot box is with this ballot box note one number, VoteBox[I/2] ++;

3. travel through 128 ballot boxes, record who gets the most votes ballot box sequence number is MaxBox;

4. if MaxBox＜32, then background is a black; If MaxBox＞96, then background is a white; Otherwise background color mistake;

5. whether the background color of Jian Ceing is consistent with the known background color, if consistent, detects below continuing; Otherwise give up;

In conjunction with application background, make system's practicability more like this;

In the 5th step, pointer detects:

Illustrate that pointer of the present invention reads the situation that dial plate contains unique apparent pointer that is only applicable to.Algorithm hypothesis pointer passes near dial plate center or the center.

Based on above-mentioned testing result, utilize angular histogram to detect the line segment at pointer place in the dial plate zone.

The statistical graph of the angular histogram angle that to be the unique point of statistics in the dial plate become with the image X-axis with the dial plate line of centres.The method of specific implementation is similar to a last trifle grey level histogram.

Because pointer passes through the dial plate center, and bigger with dial plate background difference, the formation feature is counted more, so a peak value can appear in the angle at pointer place on histogram, thus can be by finding out the testing result that who gets the most votes's ballot box obtains pointer; Test result is seen accompanying drawing 6.

The 6th step, testing result feedback/output:

By the above-mentioned step that respectively detects, can obtain final detection result.If detect less than panel board, algorithm returns, and target is not found in the current detection position; Otherwise algorithm confirms further whether the position at dial plate place is positioned at image central authorities and whether size is suitable, and algorithm returns the cradle head control order; Detailed control mode is seen accompanying drawing 4.

The process flow diagram of detection algorithm is seen Fig. 9.

Claims (1)

1. based on circle/ellipse in the transformer station of contour of object/square instrument monitoring method; It is characterized in that it may further comprise the steps successively:

(1) in database, deposit far distance instrument detection/reading submodule in, wherein preestablish following parameters and formula:

Crusing robot carries out instrument at needs and detects to make an inspection tour on the line with the equipment that reads and can be clear that the anchor point position of equipment instrument, the inceptive direction of this position camera The Cloud Terrace and the presetting bit of camera lens;

The information of this anchor point place instrument---the center of dial plate, image, dial plate size, instrument record scale reduction formula

The dial plate background is the black/white dichromatism;

The scope of crusing robot search volume: road plane is stopped error ± 15cm, car body declination error ± 30 of crusing robot °;

The threshold value T1=50 that contour feature is counted,

Match factor threshold value T2=0.75,0＜T2＜1 wherein,

(2) overhead control end computer control crusing robot arrives instrument and detects and the anchor point that reads, and reads corresponding this anchor point information, calls and starts above-mentioned far distance instrument detection/reading submodule;

(3) computing machine arrives assigned address according to above-mentioned anchor point information by WLAN (wireless local area network) control camera and camera lens based on ICP/IP protocol;

(4) computing machine receives the image sequence that crusing robot is beamed back by WLAN (wireless local area network), according to this anchor point meter information, in the search volume of regulation, call following search subroutine, control The Cloud Terrace and camera lens carry out the dial plate searching and detecting, and this search subroutine may further comprise the steps successively:

(4.1) set following numerical value:

#define SEARCHSTEPX horizontal direction search step number,

#define SEARCHSTEPY vertical direction search step number,

#define SEARCHSTEPZ depth direction search step number,

#define STEPX horizontal direction step-size in search,

#define STEPY vertical direction step-size in search,

#define STEPZ depth direction step-size in search,

The order of turning left of #define CMD_LEFT platform,

The order of turning right of #define CMD_RIGHT The Cloud Terrace,

#define CMD_UP The Cloud Terrace upwards changes order,

#define CMD_DOWN The Cloud Terrace changes order downwards,

#define CMD_ZOOMIN cam lens draws order forward,

#define CMD_ZOOMOUT cam lens pulls back order,

The dial plate size that #define GOODSIZE is suitable is represented with pixel number,

#define IMAGEWIDTH picture traverse,

#define IMAGEHEIGHT picture altitude,

The range size of #define CENTERWIDTH picture centre,

The state variable of the current working direction of mark camera lens:

IsForward for very, represents that current camera lens furthers forward, otherwise for zooming out, pre-sets isForward=TRUE;

IsRight for very, represents the deflection to the right forward of current camera lens, otherwise for deflection left, pre-sets isRight=TRUE;

Write down current lens location P1=P2=(0,0,0), wherein, P1 (x, y, z) mark camera lens current location, P2 (x, y, z) searching position that set the goal last time of mark camera lens;

(4.2) detect present image:

If detect panel board, just carry out according to the following steps:

(4.2.1) preserve lens location P2;

(4.2.2) if the position of dial plate center and picture centre is inconsistent, computing machine earlier the adjustment camera lens about, upper-lower position, detect present image more again;

(4.2.2) as if the position consistency of dial plate center and picture centre, computing machine judges just whether the dial plate size is suitable;

(4.2.3) if off size suitable, computing machine is adjusted the front and back position of camera lens earlier, detects present image more again;

(4.2.4) if size is suitable, computing machine is lock onto target just, and program is returned reading, sends the finish command; If detect less than panel board, the computer control camera lens is return last position of preserving, and then carries out according to the following steps:

(4.2.5) computing machine furthers camera lens earlier, detects present image more again; If still detect less than, again camera lens is zoomed out, detect present image again, if still detect less than, just carry out next step;

After (4.2.6) computing machine is moved camera lens to proximal most position, respectively the The Cloud Terrace position is moved to the rightlyest and the most left again, detect present image again, if still detect, just carry out next procedure less than present image;

(4.2.7) computing machine moves to the position, highest and lowest to the The Cloud Terrace position respectively again, detects present image again, if the search failure sends the finish command;

(4.2.8) computing machine zooms out camera lens again, and repeating step (4.2.5) is to (4.2.8);

(5) behind the target lock-on, computing machine makes it obtain instrument to be detected suitable image size and size according to the testing result controls lens, and the reading pointer angle also is converted into the dial plate reading, beams back by WLAN (wireless local area network); For the situation that can not read, by WLAN (wireless local area network) suitable image sequence is beamed back demonstration, by manually reading concurrent crusing robot " task is finished " signal of giving; If the target search failure also sends the detection failure information by WLAN (wireless local area network) to computing machine and issues crusing robot " task is finished " signal simultaneously;

(6) after computing machine receives the video image that camera lens gathers, just start detection program, handle video image according to the following steps:

(6.1) pre-service, it may further comprise the steps successively:

(6.1.1) video image of input is converted into gray level image by following formula:

Each gray values of pixel points

$I=\frac{R+B+G}{3},$

Wherein, R, G, B are respectively the brightness value of red, green, blue three looks;

(6.1.2) according to the following steps gray level image is carried out medium filtering:

Earlier (x y), puts new brightness value to it as this by the brightness value of 8 points around each pixel of COMPUTER CALCULATION and the mean value I of this original brightness value

$I(x,y)=\frac{\underset{i=-1}{\overset{1}{\mathrm{\Σ}}}\underset{j=-1}{\overset{1}{\mathrm{\Σ}}}I(x+i,y+j)}{9},$

With known method grey level histogram is done equalization then and handle, the concrete principle and the algorithm of histogram equalization are known;

(6.1.3) the Flame Image Process function library of usefulness open source code, be that OPENCV carries out the Canny rim detection to above-mentioned gray level image, obtain the bianry image of same size, wherein, the pixel that contains marginal information is white, other zone is a black, and these white pixels are called the edge feature point;

(6.1.4) close (expanding earlier, again corrosion) operating function with the morphology in the OPENCV function library again, above-mentioned bianry image is handled, obtain the bianry image that the continuous as far as possible available outline line of same object edge image is represented;

(6.2) carrying out geometric figure according to the following steps detects:

(6.2.1) all outline lines of above-mentioned entire image are encoded by the chain code form, obtain a profile chained list;

Be every the outline line cutting that surpasses T1 of counting that comprises independent subimage (6.2.2), remove remaining outline line;

(6.2.3) utilize the RANSAC method on each contour images, i.e. stochastic sampling ballot figure fitting process carries out the match of assignment graph:

(6.2.3.1) match of standard round:

Central coordinate of circle (x0, y0):

$x_0=\frac{d_{y1}{d}_{y2}{d}_{y3}-(x_1^2{d}_{y2}+x_2^2{d}_{y3}+x_3^2{d}_{y1})}{2(x_1{d}_{y3}+x_2{d}_{y1}+x_3{d}_{y2})},$

$y_0=\frac{d_{x1}{d}_{x2}{d}_{x3}-(y_1^2{d}_{x2}+y_2^2{d}_{x3}+y_3^2{d}_{x1})}{2(y_1{d}_{x3}+y_2{d}_{x1}+y_3{d}_{x2})};$

Radius $R=\sqrt{{(x_0-x_1)}^2+{(y_0-y_1)}^2};$

P wherein _i(x _i, y _i) (i=1,2,3) be three unique points on the outline line,

d _xi＝(x _i-x _(i+1)％3)，

d _yi＝(y _i-y _(i+1)％3)，

Subscript " %3 " is represented the remainder divided by 3;

(6.2.3.2) general oval match:

Oval have following expression formula as general quafric curve:

a ₀x ²+2a ₁xy+a ₂y ²+2a ₃x+2a ₄y＝1，

Coefficient has wherein passed through normalized, and the parameter on the right is changed to 1, does not consider the situation through initial point;

Five unique point P on the contouring line _i(x _i, y _i) (i=1 ..., 5), structure system of linear equations: A=P ^-1B,

Wherein,

A＝[a ₀?a ₁?a ₂?a ₃?a ₄] ^T，B＝[1?1?1?1?1] ^T，

$P=\left[\begin{array}{ccccc}{x_1}^2& 2x_1{y}_1& {y_1}^2& {2x}_1& {2y}_1\\ {x_2}^2& {2x}_2{y}_2& {y_2}^2& {2x}_2& {2y}_2\\ {x_3}^2& 2x_3{y}_3& {y_3}^2& 2x_3& 2y_3\\ {x_4}^2& 2x_4{y}_4& {y_4}^2& {2x}_4& 2y_4\\ {x_5}^2& {2x}_5{y}_5& {y_5}^2& {2x}_5& {2y}_5\end{array}\right],$

Mark I ₁=a ₀+ a ₂,

$I_2=\left[\begin{array}{ccc}{a}_0& a_1& a_3\\ a_1& a_2& a_4\\ a_3& a_4& -1\end{array}\right],$

According to following conditions restriction, i.e. the elliptic parameter condition that should satisfy, can judge:

I ₁And I ₂Should satisfy:

I ₁I ₂＜0， $I_3=\left[\begin{array}{cc}{a}_0& a_1\\ a_1& a_2\end{array}\right]=a_0{a}_2-a_1^2>0;$

According to the conic section characteristic, the quafric curve general parameters is converted into the description form at elliptical center commonly used, axle, inclination angle, conversion formula is as follows:

λ wherein ₁And λ ₂Be matrix I ₂Eigenwert;

(6.2.3.3) match of rectangle:

Existing rectangle fitting algorithm based on profile obtains in the employing OPENCV function library, directly exports four fixed points of detected rectangle coordinate figure;

(6.2.4) testing result aftertreatment:

(6.2.4.1) determining of the marginal information unique point of match: whether it exists the unique point that is fit to determine by each angle on the angle of calculating 360 degree;

(6.2.4.2) judge with the method for grey level histogram whether the dial plate background color that detects is consistent with the known background color, presses following steps and carries out:

(6.2.4.2.1) create 128 ballot box VoteBox[128], ballot box is emptied zero setting;

(6.2.4.2.2) for the brightness value I (0＜=I＜=255) of each pixel in testing result zone, corresponding I/2 ballot box is with this ballot box note one number, VoteBox[I/2] ++;

(6.2.4.2.3) 128 ballot boxes of traversal, record who gets the most votes ballot box sequence number is MaxBox;

If (6.2.4.2.4) MaxBox＜32, then background is a black; If MaxBox＞96, then background is a white; Otherwise background color mistake;

(6.2.4.2.5) whether the background color of Jian Ceing is consistent with the known background color, if consistent, detects below continuing; Otherwise give up;

(6.2.4.3) pointer detects:

The angle of the unique point of statistics in the dial plate and the dial plate line of centres and image X-axis formation is made angular histogram, and according to the method for step (6.2.4.2), the angle that peak value occurs is the maximum ballot box of number of votes obtained just, is the angle at pointer place.

Images