CN110321853A - Distribution cable external force damage prevention system based on video intelligent detection - Google Patents
Distribution cable external force damage prevention system based on video intelligent detection Download PDFInfo
- Publication number
- CN110321853A CN110321853A CN201910604299.4A CN201910604299A CN110321853A CN 110321853 A CN110321853 A CN 110321853A CN 201910604299 A CN201910604299 A CN 201910604299A CN 110321853 A CN110321853 A CN 110321853A
- Authority
- CN
- China
- Prior art keywords
- target
- bounding box
- detection
- prediction
- width
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
- G06V20/41—Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items
-
- G—PHYSICS
- G08—SIGNALLING
- G08B—SIGNALLING OR CALLING SYSTEMS; ORDER TELEGRAPHS; ALARM SYSTEMS
- G08B13/00—Burglar, theft or intruder alarms
- G08B13/18—Actuation by interference with heat, light, or radiation of shorter wavelength; Actuation by intruding sources of heat, light, or radiation of shorter wavelength
- G08B13/189—Actuation by interference with heat, light, or radiation of shorter wavelength; Actuation by intruding sources of heat, light, or radiation of shorter wavelength using passive radiation detection systems
- G08B13/194—Actuation by interference with heat, light, or radiation of shorter wavelength; Actuation by intruding sources of heat, light, or radiation of shorter wavelength using passive radiation detection systems using image scanning and comparing systems
- G08B13/196—Actuation by interference with heat, light, or radiation of shorter wavelength; Actuation by intruding sources of heat, light, or radiation of shorter wavelength using passive radiation detection systems using image scanning and comparing systems using television cameras
- G08B13/19602—Image analysis to detect motion of the intruder, e.g. by frame subtraction
-
- G—PHYSICS
- G08—SIGNALLING
- G08B—SIGNALLING OR CALLING SYSTEMS; ORDER TELEGRAPHS; ALARM SYSTEMS
- G08B13/00—Burglar, theft or intruder alarms
- G08B13/18—Actuation by interference with heat, light, or radiation of shorter wavelength; Actuation by intruding sources of heat, light, or radiation of shorter wavelength
- G08B13/189—Actuation by interference with heat, light, or radiation of shorter wavelength; Actuation by intruding sources of heat, light, or radiation of shorter wavelength using passive radiation detection systems
- G08B13/194—Actuation by interference with heat, light, or radiation of shorter wavelength; Actuation by intruding sources of heat, light, or radiation of shorter wavelength using passive radiation detection systems using image scanning and comparing systems
- G08B13/196—Actuation by interference with heat, light, or radiation of shorter wavelength; Actuation by intruding sources of heat, light, or radiation of shorter wavelength using passive radiation detection systems using image scanning and comparing systems using television cameras
- G08B13/19617—Surveillance camera constructional details
-
- G—PHYSICS
- G08—SIGNALLING
- G08B—SIGNALLING OR CALLING SYSTEMS; ORDER TELEGRAPHS; ALARM SYSTEMS
- G08B13/00—Burglar, theft or intruder alarms
- G08B13/18—Actuation by interference with heat, light, or radiation of shorter wavelength; Actuation by intruding sources of heat, light, or radiation of shorter wavelength
- G08B13/189—Actuation by interference with heat, light, or radiation of shorter wavelength; Actuation by intruding sources of heat, light, or radiation of shorter wavelength using passive radiation detection systems
- G08B13/194—Actuation by interference with heat, light, or radiation of shorter wavelength; Actuation by intruding sources of heat, light, or radiation of shorter wavelength using passive radiation detection systems using image scanning and comparing systems
- G08B13/196—Actuation by interference with heat, light, or radiation of shorter wavelength; Actuation by intruding sources of heat, light, or radiation of shorter wavelength using passive radiation detection systems using image scanning and comparing systems using television cameras
- G08B13/19654—Details concerning communication with a camera
- G08B13/1966—Wireless systems, other than telephone systems, used to communicate with a camera
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/07—Target detection
Abstract
The invention discloses a kind of distribution cable external force damage prevention systems based on video intelligent detection, and distribution type fiber-optic is used to realize the sensing Trigger Function of cable laying whole process external force damage prevention with cable laying;Later, using the trigger signal, standby camera is waken up, carries out the scene crawl of presetting bit monitoring position;The scene picture is uploaded to target detection server, and identification obtains the behavioral agent of intrusion event on the server, judges that behavior endangers.The device can be round-the-clock unattended, all standing cable laying route, is triggered by vibration and realizes standby wakeup, has lower power consumption, is suitble to the fields scenes such as photovoltaic power supply.Particularly, which utilizes computer vision and the relevant target detection technique of image procossing, is capable of the classification of automatic identification invasion object, provides a kind of effective monitoring means for the detection of smart grid.
Description
Technical field
The present invention relates to the monitoring and protecting fields of computer vision technique and underground cable, are specifically examined based on video intelligent
The distribution cable external force damage prevention system of survey.
Background technique
There are mainly two types of fiber cable laying modes: one is aerial layings, and one is use buried form;It is buried laying by
In not needing the materials such as electric pole, porcelain insulator, cross-arm, bracing wire, investment can be greatlyd save, and in view of laying work condition,
Environmental quality etc. has become a kind of trend of the following optical cable installation.The especially buried armour of city 10KV ring network power supply system
Dress electric power optical cable has the characteristics that the magnitude of value is larger, optical cable ditch is routed, it is big to be widely distributed in remote area etc. three relatively, leads to 10KV
City ring network power supply system is destroyed again and again by external force or burglar's attack.
The antitheft means of optical cable mainly include covering optical cable trench cover, artificial right place inspection etc., for more remote one
The place of point but rarely has patrol officer, results in that a large amount of criminals are organic to be taken advantage of.Rising in recent years is carried out using sensor
Optical cable is antitheft, realizes whether route completely detects by the sensor installed together with optical cable, can preferably prevent from stealing and go
For generation, but due to detection specific aim it is poor, production when passing through for stormy weather, excavator operative scenario, traction engine
The no separating capacity of raw vibration, and vibrate optical cable and laid by the way of parallel optical cable, radiation scope is limited, whole to miss
Report rate is higher.
Existing Video Supervision Technique carries out long-term monitoring primarily directed to the fixed point of key area, can be used for finding
Theft early period, breakoff phenomenon, and corresponding early warning and treatment measures are provided, but there are three aspects: monitoring information amount
Contradiction greatly between limited monitoring personnel;Data storage capacity is big, analysis is difficult;Passive monitoring, real-time response is poor, subsequent ability
Processing.
The patent of Patent No. CN206115620U discloses a kind of tamper-proof monitoring device of optical cable, including intrusion alarm
Device, video server and linkage alarm device;The intrusion alarm device includes central alarm control host, vibrating sensing light
Cable, optical ca ble protection net, it is online that the vibrating sensing optical cable is laid on optical ca ble protection in a manner of " S " type, the vibrating sensing optical cable
Upper several fiber optic intrusion alarms of setting, the fiber optic intrusion alarm are connect with central alarm control host signal;It is described
Video server includes video server, high-definition camera, and the high-definition camera is connect with video server, on site
The acquisition and transmission of video image;The linkage alarm device is connect with central alarm control host, video server.This reality
Be with novel beneficial effect: radiation scope increases, and can carry out intelligent recognition and analysis to intrusion behavior, in combination with
Video monitoring system links, and avoids due to reporting or failing to report caused loss by mistake.
But the patent is not directed to video, photo, which is done, further to be handled, i.e. the parts such as Target detection and identification, is not had
Rationally solved;Therefore the present invention in view of the deficiencies of the prior art, now provides a solution.
Summary of the invention
The purpose of the present invention is to provide the distribution cable external force damage prevention systems detected based on video intelligent.
The purpose of the present invention can be achieved through the following technical solutions:
Based on the distribution cable external force damage prevention system of video intelligent detection, including defence area optical fiber, optical fiber host, interlink warning
Relay, ball machine camera, server and alarm;
Wherein, the defence area optical fiber is made of sensor fibre and reference optical fiber, builds light according to Michelson Interference Principle
Learn circuit, and on optical fiber host realize coherent light waves transmitting and interfere light wave reception;
The optical fiber host includes laser light source and photodetector, for realizing the detection of interference light intensity, and according to letter
It is number strong and weak to generate alarm signal;
The interlink warning relay uses dry contact normally open, and external 5V DC power supply is generating outer broken triggering letter
It is connected after number, gives ball machine camera+5V level, ball machine camera is waken up from dormant state;
The ball machine camera is built-in with 4G module, for carrying out shooting photo to surveillance area and obtaining live video
Information, and server host is sent for the photo of shooting and the video information got by 4G wireless network;
The server host is used to receive the photo of ball machine camera shooting, and specified folder is arrived in storage;
The server host is used for after detecting the more new photo in file, starting analysis;It is regarded using computer
Feel target detection technique relevant with image procossing, identifies outer broken event, and mark includes the picture of outer broken event, determines invasion
The type of object;And the mark picture generated, it is shown on server host;
The server host is also built-in with AI image recognition program software, and the server host is also used to drive control
Ball machine camera, actively wakes up camera, obtains the video information and shooting photo at scene, and carries out to video information and picture
Analysis;
The alarm continues sounding alarm by buzzer after confirmation has outer broken event to occur;
Wherein, the server host is to the processing step of video information and picture, specifically:
Step 1: first pre-processing video information or picture, adjusts suitable dimension of picture;
Step 2: and then be passed to convolutional neural networks and obtain numerous candidate results, finally use non-maxima suppression
Algorithm obtains final result;For AI image recognition program software by the bounding box of target detection, confidence and condition classification are general
Rate is unified into an independent neural network, and neural network extracts the feature of entire image to predict the position of each bounding box
Parameter and its said target classification;
Wherein, the method for training deep neural network building Model of Target Recognition, specific step is as follows for this method:
Step (1): as size being 448 × 448 by the resolution adjustment for the samples pictures that ImageNet racing data integrates,
And adjusted picture is divided into 7 × 7 grids;Choose predicted boundary frame of the bounding box in samples pictures as target detection
Template, tagged to the grid cell of samples pictures, wherein label contains the center of target, width, highly and
Classification;
Step (2): building YOLO target detection model;Detailed process are as follows:
S1: using preceding 20 convolutional layers, and 4 maximum pond layers carry out pre-training;
S2: classification task is moved into target detection;
S3: 4 convolutional layers of addition and 2 full articulamentums on the basis of original 20 convolutional layers and 4 maximum pond layers,
Start random initializtion network parameter, constitutes YOLO target detection model to get original neural network model has been arrived;Convolution
Layer is responsible for extracting detection clarification of objective;Preceding 20 convolutional layers, 4 maximum pond layers, and 4 last convolutional layers and 2
Full articulamentum constitutes a complete sorter network;
S4: the full articulamentum predicted boundary frame coordinate of the last layer and class probability;YOLO algorithm by the width of image and
Height has obtained w, h come the width and height for normalizing bounding box, and converts centre bit for the center position coordinates of bounding box
The offset relative to corresponding network cells position is set, i.e., same normalization obtains x, y;
Predict the bounding box of the target, samples pictures will export the side of target by neural network model, that is, YOLO model
Boundary's frame, the centre coordinate including target relative to network unit bounding box, width, height;And between bounding box and template
Friendship and ratio confidence level, a possibility that as whether bounding box of response prediction includes the target and accuracy;
Wherein, it hands over and compares is defined as: the ratio of the intersection of prediction block and template coverage area and their unions, prediction block are
The detection block of model output, the detection block that template as manually marks;If prediction block and template are completely coincident, hand over and ratio is
1;
Confidence level is defined asIf target is not in the grid, confidence level 0, if
Within a grid, it hands over and ratio is exactly the value of confidence level;
Each grid cell predicted condition class probability Pr (Classi| Object), and by condition class probability and individually
The confidence level of bounding box prediction is multiplied:
Wherein, Pr (Object) indicates that some grid cell contains the probability of detection target, Pr (Classi| Object) table
Show the grid cell for the condition class probability of detection target;Pr(Classi) indicate that some grid cell contains certain classification
Detection target probability;The friendship of both prediction blocks and actual detection block of expression model output coverage area
And compare;
The obtained value that is multiplied provides the specific confidence of class of each bounding box, is demonstrated by such and appears in bounding box
In probability and evaluation bounding box include the target fine or not degree;
S5: the activation primitive using linear activation primitive as the full articulamentum of the last layer, other convolutional layers, Chi Hua
Layer, full articulamentum activation primitive be that linear activation primitive is corrected in leakage;
Step (3): amendment classification and positioning mistake;
Be size by the resolution adjustment of samples pictures it is 448 × 448, and adjusted picture is divided into 7 × 7 grids
The YOLO target detection model that input step (2) obtains, i.e., original neural network model have obtained original nerve net later
The LOSS value of network, the appropriate network weight weight values for changing random initializtion, obtains a new neural network model;And by multiple
Training iteration carrys out trim network weighted value, and learnt in the process using LOSS change curve, that is, model training of neural network
The variation of error under the propulsion of the change curve of error between parameter and standard parameter, observation training and fine tuning, when error exists
When one steady state value oscillates about, show that the error of model cannot further reduce;Show to classify at this time and positions mistake
Substantially it corrects successfully;
Meanwhile with the increase of frequency of training, observes and the overall performance index of computation model, overall performance index include
It hands over and than IOU, accurate rate Precision, recall rate Recall and accurate precision AP;When overall performance index is gradually increasing, and
And final stabilization fluctuates near a higher value, and accurate rate, the average accurate precision of recall rate and model prediction reaches
0.9 or more;Indicate the deep neural network model needed;
Step (4): interlink warning relay triggering ball machine camera is obtained to the effective video or photograph of intelligent checking system
The resolution adjustment of piece is that size is 448 × 448, and adjusted picture is divided into after 7 × 7 grids input step (3)
The deep neural network model arrived carries out target detection, obtains numerous candidate results, is finally obtained using non-maxima suppression algorithm
Final result out exports the location parameter of the bounding box of target detection, generic and positioning result.
Further, the defence area optical fiber is distributed single mode optical fiber.
Further, the optical fiber host generates alarm signal using Michelson Interference Principle.
Further, the trained deep neural network building Model of Target Recognition the step of (1) in corresponding center
The specific computational algorithm of position, width and height is as follows:
The absolute altitude and absolute width and the absolute coordinate of template center of template determining first;
Secondly then have, the height of template is the ratio of absolute altitude and resolution sizes, and similarly the width of template is exhausted
To the ratio of width and resolution sizes;The width and height of template can be obtained accordingly;
The ratio of difference and grid cell width of the abscissa of center between absolute abscissa and grid cell width,
Similarly, the ratio of difference and grid cell height of the ordinate of center between absolute ordinate and grid cell height;According to
The abscissa and ordinate of the center of template can be obtained in this.
Further, in the step S5 of step (2) the building YOLO target detection model, the linear activation letter of leakage amendment
Number are as follows:
In formula, x is upper one layer of output, and by the effect of activation primitive, output maps f (x) as next layer of input.
Further, the accurate rate Precision is specially to predict the probability of class of being really positive in the sample of class that is positive;
Recall rate Recall is specially to be predicted to be positive in all positive classes the probability of class;
Accurate precision AP is specially the accurate precision of model prediction.
Further, server host in the processing step two of video information and picture, calculate by the non-maxima suppression
Method refers to search local maximum, inhibits non-maximum element;When non-maxima suppression algorithm one target is produced it is multiple
When candidate frame, by institute it is framed sorted from large to small according to score, for each frame, if it is not suppressed, just will
All frames of the IOU greater than thresh with it are set as inhibiting, and finally return to no repressed frame.
Further, in step (3) amendment classification and positioning mistake, LOSS is calculated according to the loss function of YOLO algorithm
Value, specific LOSS value are as follows:
In formula,Indicate whether target appears in grid cell i,Indicate j-th of side in grid cell i
Boundary's frame fallout predictor is to the prediction " responsible ";I.e. if there are target, the values of j-th of bounding box fallout predictor in grid cell i
It is effective to the prediction, thenOtherwise If in grid cell i not
There are targets, then
S2Indicate that the number of meshes divided, B indicate bounding box number;(x, y) is prediction target relative to network unit side
The position of boundary's frame,It is physical location of the target relative to network unit bounding box;(w, h) be predicted boundary frame width and
Highly,It is the actual width of bounding box and height;C indicates the score of confidence level,Indicate the bounding box and standard of prediction
The cross section of frame;pi(c) classification of prediction is indicated,Indicate actual classification;
λcoordFor the loss weight of the bounding box confidence level prediction comprising target, λnoobjFor the bounding box not comprising target
The loss weight of confidence level prediction;λ is used in this modelcoord=5, λnoobj=0.5.
Further, the loss function is lost by coordinate, and confidence level loss, classification loses three parts composition;
Wherein, coordinate loss includes that position and width are high, specifically:
I.e. to S2The position of the B bounding box prediction of the cell of a division, the loss of width and height are calculated;Cause
It is greater than the width of wisp predicted boundary frame and the error of height for the width of big object predicted boundary frame and the error of height,
So carrying out out radical sign to width and height in loss function, solves the micro-locality deviation of big bounding box;
Confidence level loss specifically:
I.e. to S2The loss of the confidence score of the B bounding box prediction of the cell of a division;The former is to containing
The loss of the confidence score of each bounding box prediction of the cell of target is detected, the latter indicates to without containing detection target
The loss of the confidence score of each bounding box prediction of cell;
Classification loss specifically:
Indicate to S2The classification for the detection target that the cell of a division is predicted is lost.
Beneficial effects of the present invention:
The present invention is a distribution cable external force damage prevention system, uses distribution type fiber-optic with cable laying to realize cable
Lay the sensing Trigger Function of whole external force damage prevention;Later, using the trigger signal, standby camera is waken up, carries out presetting bit prison
The scene crawl that location is set;Followed by, which is uploaded to target detection server, and identification obtains on the server
The behavioral agent of intrusion event judges that behavior endangers.The device can be round-the-clock unattended, all standing cable laying route, leads to
It crosses vibration triggering and realizes standby wakeup, have lower power consumption, be suitble to the fields scenes such as photovoltaic power supply.Particularly, which utilizes
Computer vision and the relevant target detection technique of image procossing are capable of the classification of automatic identification invasion object, are smart grid
Detection a kind of effective monitoring means are provided.
Detailed description of the invention
In order to facilitate the understanding of those skilled in the art, the present invention will be further described below with reference to the drawings.
Fig. 1 is the system block diagram of external force damage prevention system of the present invention;
Fig. 2 is the samples pictures processing schematic on the ImageNet racing data collection in the present invention;
Fig. 3 is the schematic diagram of original neural network model;
Fig. 4 is the flow chart of training deep neural network building Model of Target Recognition;
Fig. 5 is dependency structure schematic diagram of the present invention.
Specific embodiment
As shown in Figs. 1-5, a kind of distribution cable external force damage prevention system based on video intelligent detection, including defence area optical fiber,
Optical fiber host, interlink warning relay, ball machine camera, server and alarm;
Wherein, the defence area optical fiber is made of sensor fibre and reference optical fiber, builds light according to Michelson Interference Principle
Learn circuit, and on optical fiber host realize coherent light waves transmitting and interfere light wave reception;
The optical fiber host includes laser light source and photodetector, for realizing the detection of interference light intensity, and according to letter
It is number strong and weak to generate alarm signal;
The interlink warning relay uses dry contact normally open, and external 5V DC power supply is generating outer broken triggering letter
It is connected after number, gives ball machine camera+5V level, ball machine camera is waken up from dormant state;
The ball machine camera is built-in with 4G module, for carrying out shooting photo to surveillance area and obtaining live video
Information, and server host is sent for the photo of shooting and the video information got by 4G wireless network;
The server host is used to receive the photo of ball machine camera shooting, and specified folder is arrived in storage;
The server host is used for after detecting the more new photo in file, starting analysis;It is regarded using computer
Feel target detection technique relevant with image procossing, identifies outer broken event, and mark includes the picture of outer broken event, determines invasion
The type of object;And the mark picture generated, it is shown on server host;
The server host is also built-in with AI image recognition program software, and the server host is also used to drive control
Ball machine camera, actively wakes up camera, obtains the video information and shooting photo at scene, and carries out to video information and picture
Analysis;
The alarm continues sounding alarm by buzzer after confirmation has outer broken event to occur;
Wherein, the server host is to the processing step of video information and picture, specifically:
Step 1: first pre-processing video information or picture, adjusts suitable dimension of picture;
Step 2: and then be passed to convolutional neural networks and obtain numerous candidate results, finally use non-maxima suppression
Algorithm obtains final result;For AI image recognition program software by the bounding box of target detection, confidence and condition classification are general
Rate is unified into an independent neural network, and neural network extracts the feature of entire image to predict the position of each bounding box
Parameter and its said target classification;
Wherein, the method for training deep neural network building Model of Target Recognition, specific step is as follows for this method:
Step (1): as size being 448 × 448 by the resolution adjustment for the samples pictures that ImageNet racing data integrates,
And adjusted picture is divided into 7 × 7 grids;Choose predicted boundary frame of the bounding box in samples pictures as target detection
Template, tagged to the grid cell of samples pictures, wherein label contains the center of target, width, highly and
Classification;
Step (2): building YOLO target detection model;Detailed process are as follows:
S1: using preceding 20 convolutional layers, and 4 maximum pond layers carry out pre-training;
S2: classification task is moved into target detection;
S3: 4 convolutional layers of addition and 2 full articulamentums on the basis of original 20 convolutional layers and 4 maximum pond layers,
Start random initializtion network parameter, constitutes YOLO target detection model to get original neural network model has been arrived;Convolution
Layer is responsible for extracting detection clarification of objective;Preceding 20 convolutional layers, 4 maximum pond layers, and 4 last convolutional layers and 2
Full articulamentum constitutes a complete sorter network;
S4: the full articulamentum predicted boundary frame coordinate of the last layer and class probability;YOLO algorithm by the width of image and
Height has obtained w, h come the width and height for normalizing bounding box, and converts centre bit for the center position coordinates of bounding box
The offset relative to corresponding network cells position is set, i.e., same normalization obtains x, y;
Predict the bounding box of the target, samples pictures will export the side of target by neural network model, that is, YOLO model
Boundary's frame, the centre coordinate including target relative to network unit bounding box, width, height;And between bounding box and template
Friendship and ratio confidence level, a possibility that as whether bounding box of response prediction includes the target and accuracy;
Wherein, it hands over and compares is defined as: the ratio of the intersection of prediction block and template coverage area and their unions, prediction block are
The detection block of model output, the detection block that template as manually marks;If prediction block and template are completely coincident, hand over and ratio is
1;
Confidence level is defined asIf target is not in the grid, confidence level 0, if
Within a grid, it hands over and ratio is exactly the value of confidence level;
Each grid cell predicted condition class probability Pr (Classi| Object), and by condition class probability and individually
The confidence level of bounding box prediction is multiplied:
Wherein, Pr (Object) indicates that some grid cell contains the probability of detection target, Pr (Classi| Object) table
Show the grid cell for the condition class probability of detection target;Pr(ClasSi) indicate that some grid cell contains certain classification
Detection target probability;The friendship of both prediction blocks and actual detection block of expression model output coverage area
And compare;
The obtained value that is multiplied provides the specific confidence of class of each bounding box, is demonstrated by such and appears in bounding box
In probability and evaluation bounding box include the target fine or not degree;
S5: the activation primitive using linear activation primitive as the full articulamentum of the last layer, other convolutional layers, Chi Hua
Layer, full articulamentum activation primitive be that linear activation primitive is corrected in leakage;
Step (3): amendment classification and positioning mistake;
Be size by the resolution adjustment of samples pictures it is 448 × 448, and adjusted picture is divided into 7 × 7 grids
The YOLO target detection model that input step (2) obtains, i.e., original neural network model have obtained original nerve net later
The LOSS value of network, the appropriate network weight weight values for changing random initializtion, obtains a new neural network model;And by multiple
Training iteration carrys out trim network weighted value, and learnt in the process using LOSS change curve, that is, model training of neural network
The variation of error under the propulsion of the change curve of error between parameter and standard parameter, observation training and fine tuning, when error exists
When one steady state value oscillates about, show that the error of model cannot further reduce;Show to classify at this time and positions mistake
Substantially it corrects successfully;
Meanwhile with the increase of frequency of training, observes and the overall performance index of computation model, overall performance index include
It hands over and than IOU, accurate rate Precision, recall rate Recall and accurate precision AP;When overall performance index is gradually increasing, and
And final stabilization fluctuates near a higher value, and accurate rate, the average accurate precision of recall rate and model prediction reaches
0.9 or more;Indicate the deep neural network model needed;
Step (4): interlink warning relay triggering ball machine camera is obtained to the effective video or photograph of intelligent checking system
The resolution adjustment of piece is that size is 448 × 448, and adjusted picture is divided into after 7 × 7 grids input step (3)
The deep neural network model arrived carries out target detection, obtains numerous candidate results, is finally obtained using non-maxima suppression algorithm
Final result out exports the location parameter of the bounding box of target detection, generic and positioning result.
Wherein, the defence area optical fiber is distributed single mode optical fiber.
Wherein, the optical fiber host generates alarm signal using Michelson Interference Principle.
Wherein, the trained deep neural network building Model of Target Recognition the step of (1) in corresponding center,
Width and the specific computational algorithm of height are as follows:
The absolute altitude and absolute width and the absolute coordinate of template center of template determining first;
Secondly then have, the height of template is the ratio of absolute altitude and resolution sizes, and similarly the width of template is exhausted
To the ratio of width and resolution sizes;The width and height of template can be obtained accordingly;
The ratio of difference and grid cell width of the abscissa of center between absolute abscissa and grid cell width,
Similarly, the ratio of difference and grid cell height of the ordinate of center between absolute ordinate and grid cell height;According to
The abscissa and ordinate of the center of template can be obtained in this.
Wherein, in the step S5 of step (2) the building YOLO target detection model, linear activation primitive is corrected in leakage
Are as follows:
In formula, x is upper one layer of output, and by the effect of activation primitive, output maps f (x) as next layer of input.
Wherein, the accurate rate Precision is specially to predict the probability of class of being really positive in the sample of class that is positive;
Recall rate Recall is specially to be predicted to be positive in all positive classes the probability of class;
Accurate precision AP is specially the accurate precision of model prediction.
Wherein, in the processing step two of video information and picture, the non-maxima suppression algorithm is server host
Refer to search local maximum, inhibits non-maximum element;When non-maxima suppression algorithm produces multiple candidates to a target
When frame, by institute it is framed sorted from large to small according to score, for each frame, if it is not suppressed, will just own
It is set as inhibiting with its frame of the IOU greater than thresh, finally returns to no repressed frame.
Wherein, in step (3) amendment classification and positioning mistake, LOSS value, tool are calculated according to the loss function of YOLO algorithm
The LOSS value of body are as follows:
In formula,Indicate whether target appears in grid cell i,Indicate j-th of side in grid cell i
Boundary's frame fallout predictor is to the prediction " responsible ";I.e. if there are target, the values of j-th of bounding box fallout predictor in grid cell i
It is effective to the prediction, thenOtherwise If do not deposited in grid cell i
In target, then
S2Indicate that the number of meshes divided, B indicate bounding box number;(x, y) is prediction target relative to network unit side
The position of boundary's frame,It is physical location of the target relative to network unit bounding box;(w, h) be predicted boundary frame width and
Highly,It is the actual width of bounding box and height;C indicates the score of confidence level,Indicate the bounding box and standard of prediction
The cross section of frame;pi(c) classification of prediction is indicated,Indicate actual classification;
λcoordFor the loss weight of the bounding box confidence level prediction comprising target, λnoobjFor the bounding box not comprising target
The loss weight of confidence level prediction;λ is used in this modelcoord=5, λnoobj=0.5.
Wherein, the loss function is lost by coordinate, and confidence level loss, classification loses three parts composition;
Wherein, coordinate loss includes that position and width are high, specifically:
I.e. to S2The position of the B bounding box prediction of the cell of a division, the loss of width and height are calculated;Cause
It is greater than the width of wisp predicted boundary frame and the error of height for the width of big object predicted boundary frame and the error of height,
So carrying out out radical sign to width and height in loss function, solves the micro-locality deviation of big bounding box;
Confidence level loss specifically:
I.e. to S2The loss of the confidence score of the B bounding box prediction of the cell of a division;The former is to containing
The loss of the confidence score of each bounding box prediction of the cell of target is detected, the latter indicates to without containing detection target
The loss of the confidence score of each bounding box prediction of cell;
Classification loss specifically:
Indicate to S2The classification for the detection target that the cell of a division is predicted is lost.
The present invention in the specific implementation process, including following part:
1) defence area optical fiber: the defence area optical fiber is made of sensor fibre and reference optical fiber, is taken according to Michelson Interference Principle
Build optical circuit, and on optical fiber host realize coherent light waves transmitting and interfere light wave reception;
2) optical fiber host: including the main components such as laser light source and photodetector, for realizing the detection of interference light intensity,
And alarm signal is generated according to signal strength or weakness;
3) interlink warning relay: using dry contact normally open, and external 5V DC power supply is generating outer broken trigger signal
After connect, give ball machine camera+5V level, by ball machine camera from dormant state wake up, into intellectual analysis mode;
4) ball machine camera: holder, presetting bit, upload function of taking pictures;The ball machine camera is built-in with 4G module, shooting
Photo will be sent to server host by 4G wireless network;
5) server host;Server receives the photo of ball machine camera shooting, and specified folder is arrived in storage;AI identifies journey
After sequence detects the more new photo in file, starting analysis;Utilize computer vision and the relevant target detection of image procossing
Technology identifies outer broken event, and identifies the picture comprising outer broken event, determines the type of invasion object;In general, host server
Be placed on fixed-site, as electric power overhaul department master control room in.
6) user terminal software platform: software platform is installed on a host server, and AI image recognition program is in the server
After upper operation, the mark picture of generation will be shown in the platform;
Software platform also provides the management function of ball machine camera, can taking human as actively wake up camera, check scene
Video, photographic analysis;
7) after confirmation has outer broken event to occur, sounding alarm alarm: is continued by buzzer.
Wherein, the defence area optical fiber is distribution, single mode optical fiber;
Wherein, the optical fiber host generates alarm signal using Michelson Interference Principle;
Wherein, the ball machine camera be can be by cradle head control, and presetting bit can be set, 4G communication module can be passed through
Access, and can will be taken a picture, server host is uploaded to by 4G communication module;
Wherein, the server host can have certain calculating analysis ability, " AI needed for mountable this system
Image recognition program software ";
Wherein, " the AI image recognition program software " of the server host installation, the processing to video information and picture
Step, specifically:
Step 1: first pre-processing video information or picture, adjusts suitable dimension of picture;
Step 2: and then be passed to convolutional neural networks and obtain numerous candidate results, finally use non-maxima suppression
Algorithm obtains final result, non-maxima suppression algorithm, that is, local optimum method;AI image recognition program software examines target
The bounding box of survey, confidence and condition class probability are unified into an independent neural network, and neural network is extracted whole
The feature of width image predicts the location parameter and its said target classification of each bounding box;Namely picture input is trained
Model in, the model export picture in each target location parameter and its said target classification.It is briefly exactly the network
Disposably all targets in picture all can be predicted.Therefore the AI image recognition software based on convolutional neural networks
While guaranteeing processing speed, higher average accurate precision still can satisfy;By in a model to 16000 batches
Target detection training, the accurate rate of model prediction, recall rate, average accurate precision all reached 0.9 or more.
Wherein, non-maxima suppression algorithm is the prior art;Non-maxima suppression algorithm refers to search local maximum, suppression
Make non-maximum element;When algorithm produces multiple candidate frames to a target, institute framed is arranged from big to small according to score
All frames of the IOU greater than thresh with it, if it is not suppressed, just are set as inhibiting by sequence for each frame,
Finally return to no repressed frame.
Wherein, training deep neural network building Model of Target Recognition method, this method specifically includes the following steps:
Step (1): as shown in Fig. 2, the resolution adjustment for the samples pictures that ImageNet racing data is integrated is size
It is 448 × 448, and adjusted picture is divided into 7 × 7 grids;The bounding box in samples pictures is chosen as target detection
Predicted boundary frame template, and tagged by grid cell of the artificial mode to samples pictures, wherein label contains
There are the center of target, width, height and classification;The specific computational algorithm of center, width and height is as follows:
The absolute altitude and absolute width and the absolute coordinate of template center of template determining first;
Secondly then have, the height of template is the ratio of absolute altitude and resolution sizes, and similarly the width of template is exhausted
To the ratio of width and resolution sizes;Difference and net of the abscissa of center between absolute abscissa and grid cell width
The ratio of lattice cell width, similarly, difference and grid of the ordinate of center between absolute ordinate and grid cell height
The ratio of cell height.
Step (2): building YOLO target detection model;Detailed process are as follows:
S1: using preceding 20 convolutional layers, and 4 maximum pond layers carry out pre-training;
S2: classification task is moved into target detection;
S3: adding 4 convolutional layers and 2 full articulamentums on the basis of original 20 convolutional layers, starts random initializtion
Network parameter constitutes original neural network model;Convolutional layer is responsible for extracting detection clarification of objective.As shown in figure 3,
Conv.Layer indicates that convolutional layer, Maxpool Layer indicate that maximum pond layer, Conn.Layer indicate full articulamentum;Preceding 20
A convolutional layer, 4 maximum pond layers, and 4 last convolutional layers and 2 full articulamentums constitute a complete classification net
Network.
S4: the full articulamentum predicted boundary frame coordinate of the last layer and class probability;YOLO algorithm by the width of image and
It highly come the width and height that normalize bounding box, and is the offset of corresponding network position by the coordinate transformation of bounding box.
The bounding box for predicting the target, including the centre coordinate, width and height relative to network unit bounding box, and
A possibility that confidence level of friendship and ratio between bounding box and template, i.e., whether the bounding box of response prediction includes the target with
And accuracy;Samples pictures by neural network model, that is, YOLO model will export the bounding box of target include target relative to
The centre coordinate of network unit bounding box, width, height etc.;.
Wherein, it hands over and compares is defined as: the ratio of the intersection of prediction block and template coverage area and their unions, prediction block are
The detection block of model output, the detection block that template as manually marks;
If prediction block and template are completely coincident, hand over and than being 1;Confidence level is defined as
If target is not in the grid, confidence level 0, if within a grid, handing over and ratio being exactly the value of confidence level.
Each grid cell predicted condition class probability Pr (Classi| Object), and by condition class probability and individually
The confidence level of bounding box prediction is multiplied:
In formula, Pr (Object) indicates that some grid cell contains the probability of detection target, Pr (Classi| Object) table
Show the grid cell for the condition class probability of detection target.Pr(Classi) indicate that some grid cell contains certain classification
Detection target probability.The friendship of both prediction blocks and actual detection block of expression model output coverage area
The ratio between collection and union;
The obtained value that is multiplied provides the specific confidence of class of each bounding box, is demonstrated by such and appears in bounding box
In probability and evaluation bounding box include the target fine or not degree;
S5: using activation primitive of the linear activation primitive as the full articulamentum of the last layer, other convolutional layers, pond layer,
The activation primitive of full articulamentum is that linear activation primitive is corrected in leakage.
Wherein, linear activation primitive is corrected in leakage:
In formula, x is upper one layer of output, and by the effect of activation primitive, output maps f (x) as next layer of input.
YOLO target detection model is exactly by 20 convolutional layers mentioned above, 4 maximum pond layers and last 4
The network structure of a convolutional layer and 2 full articulamentum compositions is that one kind has existed and widely used target detection model;
It is wherein connected between layers by activation primitive, to form the mapping relations of whole network structure input and output.But
The network weight weight values for being each layer are random initializtion at the beginning, that is, have obtained so-called original neural network mould
Type, but be inaccurate, so needing to be iterated trained corrective networks weighted value later finally obtains last neural network
Model.
Step (3): amendment classification and positioning mistake;
Be size by the resolution adjustment of samples pictures it is 448 × 448, and adjusted picture is divided into 7 × 7 grids
The YOLO target detection model that input step (2) obtains, i.e., original neural network model have obtained original nerve net later
The LOSS value of network, the appropriate network weight weight values for changing random initializtion, obtains a new neural network model.And by multiple
Training iteration carrys out trim network weighted value, and learnt in the process using LOSS change curve, that is, model training of neural network
The variation of error under the propulsion of the change curve of error between parameter and standard parameter, observation training and fine tuning, when error exists
When one steady state value oscillates about, show that the error of model cannot further reduce, model at this time has been provided with very strong table
Sign ability can satisfy required requirement.Then show that classifying and position mistake has corrected success.
At the same time, with the increase of frequency of training, the simultaneously friendship of computation model is observed and than IOU, accurate rate
The performance indicators such as Precision, recall rate Recall, accurate precision AP;
When each performance indicator is gradually increasing, and it is finally stable fluctuated near a higher value, and major part
Energy index reaches 0.9 or more;The deep neural network model needed;
Wherein, accurate rate Precision is that prediction is positive the probability of class of being really positive in the sample of class;Recall rate Recall
For the probability for the class that is predicted to be positive in all positive classes;Accurate precision AP is the accurate precision of model prediction;
Wherein, the loss function of YOLO algorithm calculates LOSS value method particularly includes:
In formula,Indicate whether target appears in grid cell i,Indicate j-th of side in grid cell i
Boundary's frame fallout predictor is to the prediction " responsible ";I.e. if there are target, the values of j-th of bounding box fallout predictor in grid cell i
It is effective to the prediction, thenOtherwiseIf be not present in grid cell i
Target, then
S2Indicate that the number of meshes divided, B indicate bounding box number;(x, y) is prediction target relative to network unit side
The position of boundary's frame,It is physical location of the target relative to network unit bounding box;(w, h) be predicted boundary frame width and
Highly,It is the actual width of bounding box and height;C indicates the score of confidence level,Indicate the bounding box and standard of prediction
The cross section of frame;pi(c) classification of prediction is indicated,Indicate actual classification;
λcoordFor the loss weight of the bounding box confidence level prediction comprising target, λnoobjFor the bounding box not comprising target
The loss weight of confidence level prediction.λ is used in this modelcoord=5, λnoobj=0.5.
Wherein, the loss function is lost by coordinate, and confidence level loss, classification loses three parts composition;
Coordinate loss includes that position and width are high, specifically:
I.e. to S2The position of the B bounding box prediction of the cell of a division, the loss of width and height are calculated;Cause
It is greater than the width of wisp predicted boundary frame and the error of height for the width of big object predicted boundary frame and the error of height,
So carrying out out radical sign to width and height in loss function, solves the micro-locality deviation of big bounding box;
Confidence level loss specifically:
I.e. to S2The loss of the confidence score of the B bounding box prediction of the cell of a division;The former is to containing
The loss of the confidence score of each bounding box prediction of the cell of target is detected, the latter indicates to without containing detection target
The loss of the confidence score of each bounding box prediction of cell;
Classification loss specifically:
Indicate to S2The classification for the detection target that the cell of a division is predicted is lost.
Step (4): interlink warning relay triggering ball machine camera is obtained to the effective video or photograph of intelligent checking system
The resolution adjustment of piece is that size is 448 × 448, and input step three obtains after adjusted picture is divided into 7 × 7 grids
Deep neural network model carry out target detection, obtain numerous candidate results, finally obtained using non-maxima suppression algorithm
Final result exports the location parameter of the bounding box of target detection, generic and positioning result.
A kind of distribution cable external force damage prevention system based on video intelligent monitoring, is a distribution cable external force damage prevention system
System, uses distribution type fiber-optic to realize the sensing Trigger Function of cable laying whole process external force damage prevention with cable laying;Later, it uses
The trigger signal wakes up standby camera, carries out the scene crawl of presetting bit monitoring position;It followed by, will be in the scene picture
Target detection server is reached, and identification obtains the behavioral agent of intrusion event on the server, judges that behavior endangers.The device
Can be round-the-clock unattended, all standing cable laying route is triggered by vibration and realizes standby wakeup, has lower power consumption, is fitted
Close the fields scenes such as photovoltaic power supply.Particularly, which utilizes computer vision and the relevant target detection technique of image procossing,
It is capable of the classification of automatic identification invasion object, provides a kind of effective monitoring means for the detection of smart grid.
Above content is only to structure of the invention example and explanation, affiliated those skilled in the art couple
Described specific embodiment does various modifications or additions or is substituted in a similar manner, without departing from invention
Structure or beyond the scope defined by this claim, is within the scope of protection of the invention.
Claims (9)
1. the distribution cable external force damage prevention system based on video intelligent detection, which is characterized in that including defence area optical fiber, optical fiber master
Machine, interlink warning relay, ball machine camera, server and alarm;
Wherein, the defence area optical fiber is made of sensor fibre and reference optical fiber, is built optics according to Michelson Interference Principle and is returned
Road, and on optical fiber host realize coherent light waves transmitting and interfere light wave reception;
The optical fiber host includes laser light source and photodetector, for realizing the detection of interference light intensity, and it is strong according to signal
Weak generation alarm signal;
The interlink warning relay uses dry contact normally open, external 5V DC power supply, after generating outer broken trigger signal
It connects, gives ball machine camera+5V level, ball machine camera is waken up from dormant state;
The ball machine camera is built-in with 4G module, for carrying out shooting photo to surveillance area and obtaining live video letter
Breath, and server host is sent for the photo of shooting and the video information got by 4G wireless network;
The server host is used to receive the photo of ball machine camera shooting, and specified folder is arrived in storage;
The server host is used for after detecting the more new photo in file, starting analysis;Using computer vision and
The relevant target detection technique of image procossing identifies outer broken event, and identifies the picture comprising outer broken event, determines invasion object
Type;And the mark picture generated, it is shown on server host;
The server host is also built-in with AI image recognition program software, and the server host is also used to drive control ball machine
Camera actively wakes up camera, obtains the video information and shooting photo at scene, and divides video information and picture
Analysis;
The alarm continues sounding alarm by buzzer after confirmation has outer broken event to occur;
Wherein, the server host is to the processing step of video information and picture, specifically:
Step 1: first pre-processing video information or picture, adjusts suitable dimension of picture;
Step 2: and then be passed to convolutional neural networks and obtain numerous candidate results, finally use non-maxima suppression algorithm
Obtain final result;AI image recognition program software unites the bounding box of target detection, confidence and condition class probability
In one to one independent neural network, neural network extracts the feature of entire image to predict the location parameter of each bounding box
And its said target classification;
Wherein, the method for training deep neural network building Model of Target Recognition, specific step is as follows for this method:
Step (1): as size being 448 × 448 by the resolution adjustment for the samples pictures that ImageNet racing data integrates, and will
Adjusted picture is divided into 7 × 7 grids;Choose samples pictures in bounding box as target detection predicted boundary frame mark
Quasi- frame, tagged to the grid cell of samples pictures, wherein label contains the center of target, width, height and classification;
Step (2): building YOLO target detection model;Detailed process are as follows:
S1: using preceding 20 convolutional layers, and 4 maximum pond layers carry out pre-training;
S2: classification task is moved into target detection;
S3: 4 convolutional layers of addition and 2 full articulamentums on the basis of original 20 convolutional layers and 4 maximum pond layers start
Random initializtion network parameter constitutes YOLO target detection model to get original neural network model has been arrived;Convolutional layer is negative
Duty extracts detection clarification of objective;Preceding 20 convolutional layers, 4 maximum pond layers, and 4 last convolutional layers and 2 connect entirely
It connects layer and constitutes a complete sorter network;
S4: the full articulamentum predicted boundary frame coordinate of the last layer and class probability;YOLO algorithm passes through the width and height of image
W, h have been obtained to normalize width and the height of bounding box, and has converted center phase for the center position coordinates of bounding box
For the offset of corresponding network grid position, i.e., same normalization obtains x, y;
Predict the bounding box of the target, samples pictures will export the boundary of target by neural network model, that is, YOLO model
Frame, the centre coordinate including target relative to network unit bounding box, width, height;And between bounding box and template
A possibility that confidence level of friendship and ratio, as whether the bounding box of response prediction includes the target and accuracy;
Wherein, it hands over and compares is defined as: the ratio of the intersection of prediction block and template coverage area and their unions, prediction block, that is, model
The detection block of output, the detection block that template as manually marks;If prediction block and template are completely coincident, hand over and than being 1;
Confidence level is defined asIf target is not in the grid, confidence level 0, if in net
In lattice, hands over and ratio is exactly the value of confidence level;
Each grid cell predicted condition class probability Pr (Classi| Object), and by condition class probability and single bounding box
The confidence level of prediction is multiplied:
Wherein, Pr (Object) indicates that some grid cell contains the probability of detection target, Pr (Classi| Object) it indicates to be somebody's turn to do
Condition class probability of the grid cell for detection target;Pr(Classi) indicate the inspection that some grid cell contains certain classification
Survey the probability of target;Indicate the prediction block of model output and friendship and the ratio of both actual detection blocks coverage area;
The obtained value that is multiplied provides the specific confidence of class of each bounding box, is demonstrated by such and appears in bounding box
Probability and evaluation bounding box include the fine or not degree of the target;
S5: using activation primitive of the linear activation primitive as the full articulamentum of the last layer, other convolutional layers, pond layer, entirely
The activation primitive of articulamentum is that linear activation primitive is corrected in leakage;
Step (3): amendment classification and positioning mistake;
Be size by the resolution adjustment of samples pictures it is 448 × 448, and adjusted picture is divided into after 7 × 7 grids
The YOLO target detection model that input step (2) obtains, i.e., original neural network model have obtained original neural network
LOSS value, the appropriate network weight weight values for changing random initializtion, obtains a new neural network model;And by repeatedly training
Iteration carrys out trim network weighted value, and using the parameter learnt during LOSS change curve, that is, model training of neural network
The variation of error under the propulsion of the change curve of error between standard parameter, observation training and fine tuning, when error is in a perseverance
When definite value oscillates about, show that the error of model cannot further reduce;Show to classify and position at this time wrong basic
It corrects successfully;
Meanwhile with the increase of frequency of training, observes and the overall performance index of computation model, overall performance index include handing over simultaneously
Than IOU, accurate rate Precision, recall rate Recall and accurate precision AP;When overall performance index is gradually increasing, and most
Stabilization fluctuates near a higher value eventually, and accurate rate, the average accurate precision of recall rate and model prediction reach 0.9 with
On;Indicate the deep neural network model needed;
Step (4): interlink warning relay triggering ball machine camera is obtained to the effective video or photo of intelligent checking system
Resolution adjustment is that size is 448 × 448, and adjusted picture is divided into what input step (3) after 7 × 7 grids obtained
Deep neural network model carries out target detection, obtains numerous candidate results, is finally obtained most using non-maxima suppression algorithm
Eventually as a result, the location parameter of the bounding box of output target detection, generic and positioning result.
2. the distribution cable external force damage prevention system according to claim 1 based on video intelligent detection, which is characterized in that institute
Stating defence area optical fiber is distributed single mode optical fiber.
3. the distribution cable external force damage prevention system according to claim 1 based on video intelligent detection, which is characterized in that institute
Optical fiber host is stated, using Michelson Interference Principle, generates alarm signal.
4. the distribution cable external force damage prevention system according to claim 1 based on video intelligent detection, which is characterized in that institute
State trained deep neural network building Model of Target Recognition the step of (1) in corresponding center, the tool of width and height
Body computational algorithm is as follows:
The absolute altitude and absolute width and the absolute coordinate of template center of template determining first;
Secondly then have, the height of template is the ratio of absolute altitude and resolution sizes, and similarly the width of template is absolutely wide
The ratio of degree and resolution sizes;The width and height of template can be obtained accordingly;
The ratio of difference and grid cell width of the abscissa of center between absolute abscissa and grid cell width, together
Reason, the ratio of difference and grid cell height of the ordinate of center between absolute ordinate and grid cell height;Accordingly
The abscissa and ordinate of the center of template can be obtained.
5. the distribution cable external force damage prevention system according to claim 1 based on video intelligent detection, which is characterized in that step
Suddenly in the step S5 of (2) described building YOLO target detection model, linear activation primitive is corrected in leakage are as follows:
In formula, x is upper one layer of output, and by the effect of activation primitive, output maps f (x) as next layer of input.
6. the distribution cable external force damage prevention system according to claim 1 based on video intelligent detection, which is characterized in that institute
State accurate rate Precision be specially predict to be positive class sample in be really positive the probability of class;
Recall rate Recall is specially to be predicted to be positive in all positive classes the probability of class;
Accurate precision AP is specially the accurate precision of model prediction.
7. the distribution cable external force damage prevention system according to claim 1 based on video intelligent detection, which is characterized in that clothes
Device host be engaged in in the processing step two of video information and picture, the non-maxima suppression algorithm refers to search local maximum
Value, inhibits non-maximum element;When non-maxima suppression algorithm produces multiple candidate frames to a target, institute framed is pressed
It sorts from large to small according to score, for each frame, if it is not suppressed, is just greater than all IOU with it
The frame of thresh is set as inhibiting, and finally returns to no repressed frame.
8. the distribution cable external force damage prevention system according to claim 1 based on video intelligent detection, which is characterized in that step
Suddenly in (3) amendment classification and positioning mistake, LOSS value, specific LOSS value are calculated according to the loss function of YOLO algorithm are as follows:
In formula,Indicate whether target appears in grid cell i,Indicate j-th of bounding box in grid cell i
Fallout predictor is to the prediction " responsible ";I.e. if there are targets in grid cell i, the value of j-th of bounding box fallout predictor is to this
Prediction is effective, thenOtherwise If mesh is not present in grid cell i
Mark, then
S2Indicate that the number of meshes divided, B indicate bounding box number;(x, y) is prediction target relative to grid cell boundary frame
Position,It is physical location of the target relative to grid cell boundary frame;(w, h) is the width and height of predicted boundary frame,It is the actual width of bounding box and height;C indicates the score of confidence level,Indicate the bounding box and template of prediction
Cross section;pi(c) classification of prediction is indicated,Indicate actual classification;
λcoordFor the loss weight of the bounding box confidence level prediction comprising target, λnoobjFor the bounding box confidence level not comprising target
The loss weight of prediction;λ is used in this modelcoord=5, λnoobj=0.5.
9. the distribution cable external force damage prevention system according to claim 8 based on video intelligent detection, which is characterized in that institute
It states loss function to be lost by coordinate, confidence level loss, classification loses three parts composition;
Wherein, coordinate loss includes that position and width are high, specifically:
I.e. to S2The position of the B bounding box prediction of the cell of a division, the loss of width and height are calculated;Because big
The width of object predicted boundary frame and the error of height are greater than the width of wisp predicted boundary frame and the error of height, so
Radical sign is carried out out to width and height in loss function, solves the micro-locality deviation of big bounding box;
Confidence level loss specifically:
I.e. to S2The loss of the confidence score of the B bounding box prediction of the cell of a division;The former to containing detection mesh
The loss of the confidence score of each bounding box prediction of target cell, the latter indicate to the cell without containing detection target
Each bounding box prediction confidence score loss;
Classification loss specifically:
Indicate to S2The classification for the detection target that the cell of a division is predicted is lost.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910604299.4A CN110321853B (en) | 2019-07-05 | 2019-07-05 | Distributed cable external-damage-prevention system based on video intelligent detection |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910604299.4A CN110321853B (en) | 2019-07-05 | 2019-07-05 | Distributed cable external-damage-prevention system based on video intelligent detection |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110321853A true CN110321853A (en) | 2019-10-11 |
CN110321853B CN110321853B (en) | 2021-05-11 |
Family
ID=68122837
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910604299.4A Active CN110321853B (en) | 2019-07-05 | 2019-07-05 | Distributed cable external-damage-prevention system based on video intelligent detection |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110321853B (en) |
Cited By (22)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111047731A (en) * | 2019-12-25 | 2020-04-21 | 科大国创软件股份有限公司 | AR technology-based telecommunication room inspection method and system |
CN111124015A (en) * | 2019-12-26 | 2020-05-08 | 华电山东新能源有限公司莱西分公司 | Intelligent wind power inspection video monitoring method |
CN111223129A (en) * | 2020-01-10 | 2020-06-02 | 深圳中兴网信科技有限公司 | Detection method, detection device, monitoring equipment and computer readable storage medium |
CN111337789A (en) * | 2019-10-23 | 2020-06-26 | 西安科技大学 | Method and system for detecting fault electrical element in high-voltage transmission line |
CN111339879A (en) * | 2020-02-19 | 2020-06-26 | 安徽领云物联科技有限公司 | Single-person entering monitoring method and device for weapon room |
CN111564015A (en) * | 2020-05-20 | 2020-08-21 | 中铁二院工程集团有限责任公司 | Method and device for monitoring perimeter intrusion of rail transit |
CN111723690A (en) * | 2020-06-03 | 2020-09-29 | 北京全路通信信号研究设计院集团有限公司 | Circuit equipment state monitoring method and system |
CN111797758A (en) * | 2020-07-03 | 2020-10-20 | 成都理工大学 | Identification and positioning technology for plastic bottles |
CN111860510A (en) * | 2020-07-29 | 2020-10-30 | 浙江大华技术股份有限公司 | X-ray image target detection method and device |
CN111862506A (en) * | 2020-07-29 | 2020-10-30 | 杭州巨骐信息科技股份有限公司 | Holographic sensing external force damage monitoring method |
CN111862505A (en) * | 2020-07-29 | 2020-10-30 | 杭州巨骐信息科技股份有限公司 | Holographic perception external damage monitoring system |
CN111914737A (en) * | 2020-07-29 | 2020-11-10 | 姚广元 | Return cable district section early warning intelligent host computer with identification function |
CN112001453A (en) * | 2020-08-31 | 2020-11-27 | 北京易华录信息技术股份有限公司 | Method and device for calculating accuracy of video event detection algorithm |
CN112380997A (en) * | 2020-11-16 | 2021-02-19 | 武汉巨合科技有限公司 | Model identification and undercarriage retraction and extension detection method based on deep learning |
CN112598632A (en) * | 2020-12-16 | 2021-04-02 | 北京卫星制造厂有限公司 | Appearance detection method and device for contact element of crimp connector |
CN113063484A (en) * | 2021-03-31 | 2021-07-02 | 中煤科工集团重庆研究院有限公司 | Vibration identification amplification method |
CN113344037A (en) * | 2021-05-18 | 2021-09-03 | 国网江西省电力有限公司电力科学研究院 | Cable insulation layer and sheath parameter measuring method and measuring device |
CN113486746A (en) * | 2021-06-25 | 2021-10-08 | 海南电网有限责任公司三亚供电局 | Power cable external damage prevention method based on biological induction and video monitoring |
CN113537166A (en) * | 2021-09-15 | 2021-10-22 | 北京科技大学 | Alarm method, alarm device and storage medium |
CN113810660A (en) * | 2021-08-25 | 2021-12-17 | 深圳市恺恩科技有限公司 | External damage prevention detection device for power transmission line and tower with same |
CN114360184A (en) * | 2022-01-11 | 2022-04-15 | 绍兴建元电力集团有限公司 | Multi-device linkage cable channel external damage prevention monitoring method and system |
CN116452878A (en) * | 2023-04-20 | 2023-07-18 | 广东工业大学 | Attendance checking method and system based on deep learning algorithm and binocular vision |
Citations (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7881505B2 (en) * | 2006-09-29 | 2011-02-01 | Pittsburgh Pattern Recognition, Inc. | Video retrieval system for human face content |
CN104008621A (en) * | 2014-06-03 | 2014-08-27 | 天津求实飞博科技有限公司 | Defense area type optical fiber disturbance periphery security and protection system and quick invasion disturbance judgment method |
CN107220603A (en) * | 2017-05-18 | 2017-09-29 | 惠龙易通国际物流股份有限公司 | Vehicle checking method and device based on deep learning |
CN206863889U (en) * | 2017-05-09 | 2018-01-09 | 杭州巨骐信息科技股份有限公司 | The anti-external force of power cable destroys intelligent monitor system |
CN108509954A (en) * | 2018-04-23 | 2018-09-07 | 合肥湛达智能科技有限公司 | A kind of more car plate dynamic identifying methods of real-time traffic scene |
CN108573228A (en) * | 2018-04-09 | 2018-09-25 | 杭州华雁云态信息技术有限公司 | A kind of electric line foreign matter intrusion detection method and device |
CN108682098A (en) * | 2018-07-12 | 2018-10-19 | 国网江苏省电力有限公司扬州供电分公司 | A kind of cable external force damage prevention monitoring and alarming system and its working method |
CN108734117A (en) * | 2018-05-09 | 2018-11-02 | 国网浙江省电力有限公司电力科学研究院 | Cable machinery external corrosion failure evaluation method based on YOLO |
CN108805070A (en) * | 2018-06-05 | 2018-11-13 | 合肥湛达智能科技有限公司 | A kind of deep learning pedestrian detection method based on built-in terminal |
CN109948501A (en) * | 2019-03-13 | 2019-06-28 | 东华大学 | The detection method of personnel and safety cap in a kind of monitor video |
-
2019
- 2019-07-05 CN CN201910604299.4A patent/CN110321853B/en active Active
Patent Citations (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7881505B2 (en) * | 2006-09-29 | 2011-02-01 | Pittsburgh Pattern Recognition, Inc. | Video retrieval system for human face content |
CN104008621A (en) * | 2014-06-03 | 2014-08-27 | 天津求实飞博科技有限公司 | Defense area type optical fiber disturbance periphery security and protection system and quick invasion disturbance judgment method |
CN206863889U (en) * | 2017-05-09 | 2018-01-09 | 杭州巨骐信息科技股份有限公司 | The anti-external force of power cable destroys intelligent monitor system |
CN107220603A (en) * | 2017-05-18 | 2017-09-29 | 惠龙易通国际物流股份有限公司 | Vehicle checking method and device based on deep learning |
CN108573228A (en) * | 2018-04-09 | 2018-09-25 | 杭州华雁云态信息技术有限公司 | A kind of electric line foreign matter intrusion detection method and device |
CN108509954A (en) * | 2018-04-23 | 2018-09-07 | 合肥湛达智能科技有限公司 | A kind of more car plate dynamic identifying methods of real-time traffic scene |
CN108734117A (en) * | 2018-05-09 | 2018-11-02 | 国网浙江省电力有限公司电力科学研究院 | Cable machinery external corrosion failure evaluation method based on YOLO |
CN108805070A (en) * | 2018-06-05 | 2018-11-13 | 合肥湛达智能科技有限公司 | A kind of deep learning pedestrian detection method based on built-in terminal |
CN108682098A (en) * | 2018-07-12 | 2018-10-19 | 国网江苏省电力有限公司扬州供电分公司 | A kind of cable external force damage prevention monitoring and alarming system and its working method |
CN109948501A (en) * | 2019-03-13 | 2019-06-28 | 东华大学 | The detection method of personnel and safety cap in a kind of monitor video |
Non-Patent Citations (4)
Title |
---|
JOSEPH REDMON ET AL.: "You Only Look Once:Unified, Real-Time Object Detection", 《2016 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
陆斌: "电缆线路防外破预警系统应用", 《中国电力企业管理》 * |
黄肖为 等: "电力电缆线路回流缆防盗报警装置的研制与应用", 《科技创新与应用》 * |
黄肖为 等: "高压电缆线路现场智能移动巡检终端设计", 《科技创新导报》 * |
Cited By (30)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111337789A (en) * | 2019-10-23 | 2020-06-26 | 西安科技大学 | Method and system for detecting fault electrical element in high-voltage transmission line |
CN111047731A (en) * | 2019-12-25 | 2020-04-21 | 科大国创软件股份有限公司 | AR technology-based telecommunication room inspection method and system |
CN111124015A (en) * | 2019-12-26 | 2020-05-08 | 华电山东新能源有限公司莱西分公司 | Intelligent wind power inspection video monitoring method |
CN111223129A (en) * | 2020-01-10 | 2020-06-02 | 深圳中兴网信科技有限公司 | Detection method, detection device, monitoring equipment and computer readable storage medium |
CN111339879A (en) * | 2020-02-19 | 2020-06-26 | 安徽领云物联科技有限公司 | Single-person entering monitoring method and device for weapon room |
CN111339879B (en) * | 2020-02-19 | 2023-06-02 | 安徽领云物联科技有限公司 | Weapon room single person room entering monitoring method and device |
CN111564015A (en) * | 2020-05-20 | 2020-08-21 | 中铁二院工程集团有限责任公司 | Method and device for monitoring perimeter intrusion of rail transit |
CN111564015B (en) * | 2020-05-20 | 2021-08-24 | 中铁二院工程集团有限责任公司 | Method and device for monitoring perimeter intrusion of rail transit |
CN111723690A (en) * | 2020-06-03 | 2020-09-29 | 北京全路通信信号研究设计院集团有限公司 | Circuit equipment state monitoring method and system |
CN111723690B (en) * | 2020-06-03 | 2023-10-20 | 北京全路通信信号研究设计院集团有限公司 | Method and system for monitoring state of circuit equipment |
CN111797758A (en) * | 2020-07-03 | 2020-10-20 | 成都理工大学 | Identification and positioning technology for plastic bottles |
CN111860510A (en) * | 2020-07-29 | 2020-10-30 | 浙江大华技术股份有限公司 | X-ray image target detection method and device |
CN111914737A (en) * | 2020-07-29 | 2020-11-10 | 姚广元 | Return cable district section early warning intelligent host computer with identification function |
CN111860510B (en) * | 2020-07-29 | 2021-06-18 | 浙江大华技术股份有限公司 | X-ray image target detection method and device |
CN111862505A (en) * | 2020-07-29 | 2020-10-30 | 杭州巨骐信息科技股份有限公司 | Holographic perception external damage monitoring system |
CN111862506A (en) * | 2020-07-29 | 2020-10-30 | 杭州巨骐信息科技股份有限公司 | Holographic sensing external force damage monitoring method |
CN112001453A (en) * | 2020-08-31 | 2020-11-27 | 北京易华录信息技术股份有限公司 | Method and device for calculating accuracy of video event detection algorithm |
CN112001453B (en) * | 2020-08-31 | 2024-03-08 | 北京易华录信息技术股份有限公司 | Method and device for calculating accuracy of video event detection algorithm |
CN112380997A (en) * | 2020-11-16 | 2021-02-19 | 武汉巨合科技有限公司 | Model identification and undercarriage retraction and extension detection method based on deep learning |
CN112598632A (en) * | 2020-12-16 | 2021-04-02 | 北京卫星制造厂有限公司 | Appearance detection method and device for contact element of crimp connector |
CN113063484B (en) * | 2021-03-31 | 2022-10-04 | 中煤科工集团重庆研究院有限公司 | Vibration identification amplification method |
CN113063484A (en) * | 2021-03-31 | 2021-07-02 | 中煤科工集团重庆研究院有限公司 | Vibration identification amplification method |
CN113344037A (en) * | 2021-05-18 | 2021-09-03 | 国网江西省电力有限公司电力科学研究院 | Cable insulation layer and sheath parameter measuring method and measuring device |
CN113486746A (en) * | 2021-06-25 | 2021-10-08 | 海南电网有限责任公司三亚供电局 | Power cable external damage prevention method based on biological induction and video monitoring |
CN113810660A (en) * | 2021-08-25 | 2021-12-17 | 深圳市恺恩科技有限公司 | External damage prevention detection device for power transmission line and tower with same |
CN113537166A (en) * | 2021-09-15 | 2021-10-22 | 北京科技大学 | Alarm method, alarm device and storage medium |
CN113537166B (en) * | 2021-09-15 | 2021-12-14 | 北京科技大学 | Alarm method, alarm device and storage medium |
CN114360184A (en) * | 2022-01-11 | 2022-04-15 | 绍兴建元电力集团有限公司 | Multi-device linkage cable channel external damage prevention monitoring method and system |
CN116452878A (en) * | 2023-04-20 | 2023-07-18 | 广东工业大学 | Attendance checking method and system based on deep learning algorithm and binocular vision |
CN116452878B (en) * | 2023-04-20 | 2024-02-02 | 广东工业大学 | Attendance checking method and system based on deep learning algorithm and binocular vision |
Also Published As
Publication number | Publication date |
---|---|
CN110321853B (en) | 2021-05-11 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110321853A (en) | Distribution cable external force damage prevention system based on video intelligent detection | |
CN104581076A (en) | Mountain fire monitoring and recognizing method and device based on 360-degree panoramic infrared fisheye camera | |
CN110133573A (en) | A kind of autonomous low latitude unmanned plane system of defense based on the fusion of multielement bar information | |
CN109255286B (en) | Unmanned aerial vehicle optical rapid detection and identification method based on deep learning network framework | |
CN106335646A (en) | Interference-type anti-UAV (Unmanned Aerial Vehicle) system | |
CN107016690A (en) | The unmanned plane intrusion detection of view-based access control model and identifying system and method | |
CN114419825B (en) | High-speed rail perimeter intrusion monitoring device and method based on millimeter wave radar and camera | |
CN206759621U (en) | A kind of agricultural insect monitoring device based on Internet of Things | |
CN108230302A (en) | A kind of nuclear power plant's low-temperature receiver marine site invasion marine organisms detection and method of disposal | |
CN108500992A (en) | A kind of multi-functional mobile security robot | |
CN111754714A (en) | Security monitoring system and monitoring method thereof | |
Zhang et al. | Transmission line abnormal target detection based on machine learning yolo v3 | |
CN109711348A (en) | Intelligent monitoring method and system based on the long-term real-time architecture against regulations in hollow panel | |
CN116846059A (en) | Edge detection system for power grid inspection and monitoring | |
CN107369291A (en) | The anti-external force damage alarm system and method for high-tension line based on deep learning | |
CN111369760A (en) | Night pedestrian safety early warning device and method based on unmanned aerial vehicle | |
CN115471865A (en) | Operation site digital safety control method, device, equipment and storage medium | |
Zhen et al. | Transmission tower protection system based on Internet of Things in smart grid | |
CN206594305U (en) | A kind of wild animal monitoring device | |
CN208460141U (en) | Forbidden zone intrusion alarm system | |
CN116343528A (en) | Bridge ship collision alarm sensing equipment and safety monitoring application platform comprising same | |
CN115620239A (en) | Point cloud and video combined power transmission line online monitoring method and system | |
CN115035470A (en) | Low, small and slow target identification and positioning method and system based on mixed vision | |
CN115860144A (en) | Machine learning system for anti-electricity-stealing site | |
CN111901217B (en) | Key area land-air integrated warning system based on microvibration perception |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant | ||
PE01 | Entry into force of the registration of the contract for pledge of patent right |
Denomination of invention: Distributed cable protection system based on video intelligent detection Effective date of registration: 20221109 Granted publication date: 20210511 Pledgee: Hangzhou Fuyang Sub branch of China CITIC Bank Co.,Ltd. Pledgor: HANGZHOU JUQI INFORMATION TECHNOLOGY Co.,Ltd. Registration number: Y2022980021239 |
|
PE01 | Entry into force of the registration of the contract for pledge of patent right |