The content of the invention
1st, technical problem to be solved：
Existing electric power facility external force damage prevention uses the methods of manual inspection and camera supervised video recording, and its major defect isNeed to expend larger human resources, and can not accomplish that real-time response is alarmed for broken event outside electric power facility.Regarded intelligentlyFrequency analysis algorithm can not judge the specific objective behavior that may be damaged to transmission line of electricity in terms of target analysis, it is impossible to meshMark behavior type is accurately judged.
2nd, technical scheme：
In order to solve problem above, the invention provides a kind of transmission line of electricity video external force damage prevention of Behavior-based control analytical technologyMethod, comprise the following steps：Step 1: IMAQ, IMAQ pre-processes to the original image of camera acquisition,Scaling including image, contrast enhancing and noise reduction process；Step 2: motion detection, motion detection passes through in background modelingIn more new strategy in increase Rule of judgment, to have target occur region carry out locking background process；Step 3: block is examinedSurvey, detected by block by the block structure body array representation of the motion block in foreground image；Step 4: target following,During the tracking of target, progress in the generation selection foreground blocks extracted in the current frame to the target occurred in former frameMatch somebody with somebody, choose that suitable foreground blocks are matched, the process constantly matched determine present frame some foreground blocks correspond to it is previousSome moving target in frame, is then constantly updated to clarification of objective parameter；Step 5: goal filtering, target mistakeFilter is filtered according to the size of target, color, shape, movement velocity, direction and movement locus；Step 6: target identification,Target identification is to carry out Classification and Identification to target；Step 7: goal behavior is analyzed, it is determined that after target type, targetedlyGoal behavior analysis is carried out, described goal behavior is divided into general objectives behavior and the class of specific objective behavior two, the general meshMark behavior refers to that the behavior type that moving target all has, including target are swarmed into, left, repeating to swarm into, hover and stop firstOnly situation, the specific objective behavior are the goal behavior of certain types of target；Step 8: external force damage prevention is alarmed, when it is determined that meshAfter marking type and goal behavior, whether broken event outside electric power facility is occurred according to rule judgment set in advance, is being judged reallyIn the case of the fixed outer broken event of generation, intelligent analysis system will be pushed alarming short message, picture and alarm by wireless network and be regardedFrequently, avoid causing serious electric power facility to damage accident.
Further, on the basis of the motion detection, target following, target identification, goal behavior analysis is completed, energyElectric power facility scene modeling is enough carried out, model of place includes static layer, wherein three parts of dynamic layer and statistics layer, static layerFor representing the fixed object in electric power facility scene, dynamic layer is used to represent the moving target in scene, and statistics layer is by dividingThe target frequency of occurrence in each region in scene is analysed, goal activities hot spot region figure is calculated.
3rd, beneficial effect：
The present invention is directed to the transmission line of electricity external force damage prevention demand of power system, proposes a kind of the anti-outer of Behavior-based control analytical technologyBroken intelligent analysis method, can not only realize the detection, tracking and identification of moving target, additionally it is possible to which goal behavior type is carried outAccurately judge.
Embodiment
The present invention is described in detail below.
Step 1: IMAQ
IMAQ pre-processes to the original image of camera acquisition, including the scaling of image, contrast enhancing withAnd the processing such as noise reduction.It is high-resolution situation such as 720P, 1080P for original image, original image can be contracted toRelatively low resolution ratio, to reduce the amount of calculation of subsequent motion detection algorithm, ensure the real-time processing to moving object detection tracking.
Step 2: motion detection
The motion detection is the situation of change by analyzing each pixel in video image, the picture that will wherein change greatlyVegetarian refreshments is judged as moving pixel, changes less pixel and be judged as background pixel point, motion detection employs a kind of improvementVIBE algorithms, by increasing Rule of judgment in the more new strategy in background modeling, to have target occur region lockDetermine background process.
It is that basic motion detection algorithm is only used in order to realize the detection of static target using improved VIBE algorithmsMoving target is detected, can be faded away when target switchs to static from motion state.
The input of motion detection block is colored or gray level image, is exported as the binary image of foreground moving object.
Step 3: block detects
The specific method of block detection is：All connected domains in foreground image are analyzed, then by each connected domainPosition, size parameter block array representation.
Step 4: target following
The target following uses the track algorithm based on Kalman filtering.Moving target is general in video monitoring sceneAt least have three kinds of states：Target enters monitoring scene for the first time, target is moved and is traced in the scene, target is from sceneExit., it is necessary to the generation selection prospect extracted in the current frame to the target occurred in former frame during the tracking of targetMatched in block, choose that suitable foreground blocks are matched, some prospect of present frame is determined in the process constantly matchedBlock corresponds to some moving target in former frame, and then constantly clarification of objective parameter is updated, it is achieved thereby that frameThe tracking of target between frame.
Step 5: goal filtering
Goal filtering is filtered according to the size of target, color, shape, movement velocity, direction and movement locus.It is rightIn target size characteristic, it is necessary to calculate projection size of the target in original image according to camera parameters.If the mesh of trackingDimensioning is less than minimum size threshold, or more than full-size threshold value, is then regarded as noise and is filtered out.In outdoor sceneIn, can be by interference filterings such as pedestrian and bicycles by minimum size threshold., can be by mesh for the color characteristic of targetLogo image is changed to hsv color space and handled.In the HSV images of target, according to panel tone set in advance, saturationDegree and luminance threshold scope, statistics meet the pixel number of condition.If pixel number is less than the percentage parameter of setting,It is regarded as noise and is filtered out.
Step 6: target identification
Target identification is to carry out Classification and Identification to target.Described target identification includes single frames target identification and multiple frame cumulationIdentification, the single frames target identification are to carry out Classification and Identification to the target in single frame video image, and the multiple frame cumulation is identified asOn the basis of single frames target identification, Classification and Identification is carried out to target by the recognition result of multiple frame cumulation.
Described single frames target identification carries out Classification and Identification using deep learning method to the target in single frame video image.This method is needed to gather substantial amounts of sample data, and then deep neural network is trained by sample data, can be realizedHigher recognition accuracy.According to the difference of application scenarios, it is necessary to gather corresponding target identification sample respectively.For indoor fieldScape, Classification and Identification mainly is carried out to personnel and toy.For the Large Construction of outdoor scene, mainly road pavement travelingVehicle is such as crane, excavator and other types vehicle carries out Classification and Identification such as car, truck, Bus Carriage.UseSelf-built vehicle classification Sample Storehouse is trained to single frames target identification module, and this Sample Storehouse includes the classes such as crane, car, truckType vehicle is mainly used in the Classification and Identification of crane and other types vehicle in the sample image of different angle.
On the basis of single frames target identification, Classification and Identification is carried out to target by the recognition result of multiple frame cumulation, so as toRealize higher recognition accuracy.
Described multiple frame cumulation identifies that in the case of target is divided into k types the calculation formula of multiple frame cumulation identification is：
N is target type sum in formula,It is the probability of kth class for target,For target kth class migration index,Number is identified for multiple frame cumulation of the target in kth class,For target number summation is identified in all types of accumulations.
Pass through changeThe rate of false alarm and rate of failing to report of target identification can be adjusted.Such as pass through increase, mesh can be reducedThe rate of failing to report of kth class is marked on, but the rate of false alarm of target can be increased simultaneously.Being recorded a video in actual scene, by way of sectional drawing to moreFrame accumulation identification module is tested, and the time of 4 tests is the morning 7:00 to afternoon 17:00, count the vehicle fleet passed byFor 9748, wherein crane wrong report number is 233, and average rate of false alarm is 2.5%.
Step 7: goal behavior is analyzed
Described goal behavior is divided into general objectives behavior and the class of specific objective behavior two, and the general objectives behavior refers toThe behavior type that moving target all has, including target are swarmed into, left, repeating to swarm into, hover and stop situation first, describedSpecific objective behavior is the goal behavior of certain types of target.
The goal behavior of the general objectives behavior is analyzed：Start to carry out target after target enters monitoring scene withTrack, one or more detection zones are arranged as required in monitoring scene, corresponding target line can be set to each detection zoneFor analytical parameters, to meet actual application demand.General objectives behavior passes through the position of combining target, motion state and fortuneDynamic time parameter, using the method for rule judgment, the method for rule judgment is：It is identified, when target is for the first time from detection zoneWhen overseas portion enters in detection zone, goal behavior type is defined to swarm into first；When target leaves detection zone, mesh is definedMark behavior type is to leave；When being again introduced into after target is left in detection zone, goal behavior type is defined to repeat to swarm into,Start timing after target enters detection zone, threshold speed is setVth1、Vth2For judging the motion state of target, if targetMovement velocity averageVmean<Vth1Target is then judged for inactive state,Vmean > Vth2Target is then judged for motion state,Vth1 <Vmean <Vth2Then judge that target is defined inactive state, time threshold is setTth1、Tth2For judging the behavior type of target.IfThe time that target is kept in motion in detection zoneTM >Tth1, then goal behavior type is judged to hover.If target existsThe time to be remained static in detection zoneTS >Tth2, then goal behavior type is judged to stop.
Described specific objective behavior is that crane raises arm, and when the behavior type for judging the target is stops, then operation is raisedArm detection algorithm determines whether to raise arm behavior, and it is as follows to raise arm detection algorithm：If target is in the position coordinates of halted state(x0,y0), the height and the width of target are respectivelyh0,w0, then raise arm detection first to target stop position (x0,y0) it is upperSquare rectangular area carries out motion detection, the position coordinates in the region for (x0,y0 +h0), the height and the width point of detection zoneIt is nota*h0 ,b*h0, whereina,bBe detection zone height, width relative to the proportionality coefficient of object height, reality can be passed throughStatistics is tested to determine.Raising in arm detection zone after discovery moving object above target, can be to the fortune at the top of the moving objectDynamic rail mark is analyzed.If the starting point of movement locus for (xt0,yt0), trail termination point for (xt1,yt1), path length.If path lengthLMore than the threshold value of settingLth, and the direction of movement locus is underUp, then it is judged as that crane raises arm.
Step 8: external force damage prevention is alarmed
When it is determined that after target type and goal behavior, electric power whether can occurring according to rule judgment set in advance and setApply outer broken event.In the case where judging to determine that outer broken event occurs, intelligent analysis system will be pushed by wireless network and alarmedShort message, picture and alarm video, avoid causing serious electric power facility to damage accident.
Step 9: electric power facility scene modeling
On the basis of motion detection, target following, identification and goal behavior analysis is completed, electric power facility can be carried outScene modeling.Model of place includes three static layer, dynamic layer and statistics layer parts.Wherein static layer is used to represent that electric power is setApply the fixed object in scene, such as steel tower in the instrument cabinet and outdoor scene in indoor scene, building etc..FixtureBody can use human configuration parameter or automatic identification location algorithm really be sized, shape and position.Dynamic layer is then used for tableShow the moving target in scene, such as the Large Construction vehicle in the personnel and toy, and outdoor scene in indoor sceneWith other types vehicle.The parameter of moving target can be determined by target following and Target Recognition Algorithms.Statistics layer is thenBy analyzing the target frequency of occurrence in each region in scene, goal activities hot spot region figure is calculated.
Compared with prior art, beneficial effect caused by electric power facility external force damage prevention intelligent analysis system provided by the inventionIt is：By adjusting systematic parameter, the electric power facility external force damage prevention demand of several scenes can adapt to.Scene indoors, realize personnelIntrusion detection and Activity recognition analysis.In outdoor scene, Large Construction vehicle motion state such as crane, excavator is realizedWith behavior type（Swarm into, hover, stopping, raising arm etc.）Real time identification analysis.It is broken outside electric power facility when occurring in detection zoneDuring event, it can accurately identify and respond rapidly to alarm, alarming short message, picture and alarm video are pushed by wireless network,Avoid causing serious electric power facility to damage accident.