CN110321853B - Distributed cable external-damage-prevention system based on video intelligent detection - Google Patents

Distributed cable external-damage-prevention system based on video intelligent detection Download PDF

Info

Publication number
CN110321853B
CN110321853B CN201910604299.4A CN201910604299A CN110321853B CN 110321853 B CN110321853 B CN 110321853B CN 201910604299 A CN201910604299 A CN 201910604299A CN 110321853 B CN110321853 B CN 110321853B
Authority
CN
China
Prior art keywords
target
bounding box
loss
detection
prediction
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.)
Active
Application number
CN201910604299.4A
Other languages
Chinese (zh)
Other versions
CN110321853A (en
Inventor
倪晓璐
周铭权
孟庆铭
董琪
周杰
赵志刚
裘明松
俞挺
孙海华
胡文宇
金会会
赵志杭
郭能俊
罗利峰
杨洋
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hangzhou Juqi Information Technology Co ltd
Original Assignee
Hangzhou Juqi Information Technology Co ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Hangzhou Juqi Information Technology Co ltd filed Critical Hangzhou Juqi Information Technology Co ltd
Priority to CN201910604299.4A priority Critical patent/CN110321853B/en
Publication of CN110321853A publication Critical patent/CN110321853A/en
Application granted granted Critical
Publication of CN110321853B publication Critical patent/CN110321853B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

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 distributed cable anti-external-damage system based on video intelligent detection, which adopts distributed optical fibers along with cable laying to realize a sensing triggering function of cable laying full-range anti-external-damage; then, awakening the standby camera by adopting the trigger signal, and capturing a scene at a preset position monitoring position; and uploading the scene picture to a target detection server, identifying and acquiring a behavior subject of the intrusion event on the server, and judging behavior hazard. The device can be in an all-weather unattended state, and the line is laid to the full coverage cable, triggers through the vibration and realizes the standby awakening, possesses lower consumption, is fit for open-air scenes such as photovoltaic power supply. Particularly, the device utilizes computer vision and image processing related target detection technology, can automatically identify the type of the invading object, and provides an effective monitoring means for the detection of the smart grid.

Description

Distributed cable external-damage-prevention system based on video intelligent detection
Technical Field
The invention relates to the field of computer vision technology and monitoring protection of buried cables, in particular to a distributed cable external-damage-prevention system based on video intelligent detection.
Background
The optical cable laying mode mainly has two kinds: one is overhead laying, and the other is in a buried mode; the buried cable is laid without materials such as an electric pole, a porcelain insulator, a cross arm, a stay wire and the like, so that investment can be greatly saved, and the buried cable becomes a trend of future optical cable erection in consideration of laying engineering conditions, environmental characteristics and the like. Especially, the buried armored power cable of the urban 10KV looped network power supply system has the characteristics of large value, cable channel wiring, wide distribution in relatively remote areas and the like, so that the 10KV urban looped network power supply system is frequently damaged by external force or attacked by thieves.
The optical cable anti-theft means mainly comprises an optical cable trench cover plate covering, manual timing and fixed-point inspection and the like, and inspection personnel are available in remote places, so that a large number of lawbreakers can be provided with the method. The optical cable is prevented from being stolen by using the sensor which is raised in recent years, whether a circuit is completely detected or not is realized by the sensor which is installed together with the optical cable, the stealing behavior can be well prevented, but the detection pertinence is poor, the capability of distinguishing vibration generated when stormy weather, excavator working conditions and heavy duty vehicles pass through is not provided, the vibration optical cable is laid in a parallel optical cable mode, the radiation range is limited, and the whole false alarm rate is higher.
The existing video monitoring technology mainly aims at the fixed point of a key area to carry out long-term monitoring, can be used for discovering early theft and damage phenomena and providing corresponding early warning and processing measures, but has three problems: the contradiction between the large amount of monitoring information and limited monitoring personnel; the data storage capacity is large and the analysis is difficult; and passive monitoring is adopted, the real-time response is poor, and the processing can be carried out after the fact.
Patent No. CN206115620U discloses an optical cable damage prevention monitoring device, which comprises an intrusion alarm device, a video server and a linkage alarm device; the intrusion alarm device comprises a central alarm control host, a vibration sensing optical cable and an optical cable protective net, wherein the vibration sensing optical cable is laid on the optical cable protective net in an S-shaped mode, a plurality of optical fiber intrusion alarms are arranged on the vibration sensing optical cable, and the optical fiber intrusion alarms are in signal connection with the central alarm control host; the video server comprises a video server and a high-definition camera, and the high-definition camera is connected with the video server and is used for collecting and transmitting a field video image; and the linkage alarm device is connected with the central alarm control host and the video server. The beneficial effects of the utility model reside in that: the radiation range is enlarged, the intrusion behavior can be intelligently identified and analyzed, and meanwhile, the linkage is carried out by combining a video monitoring system, so that the loss caused by misinformation or missing report is avoided.
However, the patent does not aim at video and photo further processing, namely, parts such as target detection and recognition and the like are not reasonably solved; the present invention therefore addresses the deficiencies of the prior art by now providing a solution.
Disclosure of Invention
The invention aims to provide a distributed cable anti-external-damage system based on video intelligent detection.
The purpose of the invention can be realized by the following technical scheme:
the distributed cable external-damage-prevention system based on video intelligent detection comprises a defense area optical fiber, an optical fiber host, a linkage alarm relay, a dome camera, a server and an alarm;
the defense area optical fiber consists of a sensing optical fiber and a reference optical fiber, an optical loop is built according to the Michelson interference principle, and the emission of coherent light waves and the receiving of interference light waves are realized on an optical fiber host;
the optical fiber host comprises a laser light source and a photoelectric detector, is used for realizing the detection of interference light intensity and generates an alarm signal according to the strength of the signal;
the linkage alarm relay is connected with a 5V direct current power supply in an external connection mode in a dry contact normally open mode, is connected after an external break trigger signal is generated, and gives +5V level to the camera of the ball machine to awaken the camera of the ball machine from a dormant state;
the camera of the dome camera is internally provided with a 4G module which is used for shooting pictures and acquiring field video information of a monitored area and sending the shot pictures and the acquired video information to the server host through a 4G wireless network;
the server host is used for receiving the pictures shot by the camera of the dome camera and storing the pictures into a designated folder;
the server host is used for starting analysis after detecting the updated photos in the folder; identifying an external damage event by using a target detection technology related to computer vision and image processing, identifying a picture containing the external damage event, and determining the type of an invading object; the generated identification picture is displayed on the server host;
the server host is also internally provided with AI image recognition program software and is also used for driving a camera of the ball control machine, actively awakening the camera, acquiring on-site video information and shot pictures and analyzing the video information and the pictures;
the alarm is continuously sounded by a buzzer to give an alarm after the occurrence of an external damage event is confirmed;
the processing steps of the server host computer on the video information and the pictures are specifically as follows:
the method comprises the following steps: firstly, preprocessing video information or pictures and adjusting the size of the pictures to be proper;
step two: then the signal is transmitted into a convolutional neural network to obtain a plurality of candidate results, and finally a non-maximum suppression algorithm is used to obtain a final result; the AI image recognition program software unifies the bounding boxes of the target detection, the confidence score and the conditional category probability into an independent neural network, and the neural network extracts the characteristics of the whole image to predict the position parameters of each bounding box and the target category to which the bounding box belongs;
the method for training the deep neural network to construct the target recognition model comprises the following specific steps:
step (1): adjusting the resolution of the sample pictures on the ImageNet competition dataset to 448 x 448, and dividing the adjusted pictures into 7 x 7 grids; selecting a boundary frame in a sample picture as a standard frame of a prediction boundary frame for target detection, and marking a label on a grid unit of the sample picture, wherein the label contains the central position, the width, the height and the category of a target;
step (2): constructing a YOLO target detection model; the specific process is as follows:
s1: pre-training with the first 20 convolutional layers and 4 maximum pooling layers;
s2: migrating the classification task to target detection;
s3: adding 4 convolutional layers and 2 full-connection layers on the basis of the original 20 convolutional layers and 4 maximum pooling layers, starting to initialize network parameters randomly, forming a YOLO target detection model, and obtaining an original neural network model; the convolutional layer is responsible for extracting the characteristics of the detection target; the first 20 convolutional layers, the 4 largest pooling layers, and the last 4 convolutional layers and 2 fully-connected layers form a complete classification network;
s4: the last layer of full connection layer predicts the coordinate and the class probability of the bounding box; the YOLO algorithm normalizes the width and height of the bounding box by the width and height of the image to obtain w, h, and converts the center position coordinates of the bounding box into the offset of the center position relative to the corresponding network grid unit position, namely, the x, y is obtained by the same normalization;
predicting a boundary frame of the target, wherein the boundary frame of the target is output by the sample picture through a neural network model, namely a YOLO model, and comprises a central coordinate, a width and a height of the target relative to a boundary frame of a network unit; and the confidence of the intersection ratio between the bounding box and the standard box, namely the probability and the accuracy of whether the bounding box of the reaction prediction contains the target or not;
wherein the cross-over ratio is defined as: the intersection of the coverage ranges of the prediction frame and the standard frame and the ratio of the union of the coverage ranges of the prediction frame and the standard frame are determined, wherein the prediction frame is a detection frame output by the model, and the standard frame is a detection frame marked manually; if the prediction frame and the standard frame are completely overlapped, the intersection ratio is 1;
confidence is defined as
Figure GDA0002960095550000041
If the target is not in the grid, the confidence coefficient is 0, and if the target is in the grid, the intersection ratio is the value of the confidence coefficient;
predicting conditional Class probability Pr (Class) per grid celliI Object) and multiplies the conditional class probability by the confidence of the single bounding box prediction:
Figure GDA0002960095550000042
wherein, Pr (object) represents the probability that a certain grid cell contains the detection target, and Pr (Class)iI Object) represents a conditional category probability of the lattice cell for the detection target; pr (Class)i) Representing the probability that a certain grid cell contains a certain class of detection target;
Figure GDA0002960095550000051
the intersection ratio of the coverage ranges of a prediction box and an actual detection box which represent the output of the model;
providing a class-specific confidence score of each bounding box by the multiplied value, expressing the probability of the class appearing in the bounding box and evaluating the quality degree of the bounding box containing the target;
s5: adopting a linear activation function as an activation function of the last full-connection layer, wherein the activation functions of other convolution layers, pooling layers and full-connection layers are leakage correction linear activation functions;
and (3): correcting classification and positioning errors;
adjusting the resolution of the sample picture to 448 × 448, dividing the adjusted picture into 7 × 7 grids, inputting the 7 × 7 grids into the YOLO target detection model obtained in the step (2), namely the original neural network model, obtaining the LOSS value of the original neural network, and appropriately changing the network weight value initialized randomly to obtain a new neural network model; fine-tuning the network weight value through multiple training iterations, observing the change of the error under the promotion of training and fine tuning by adopting an LOSS change curve of a neural network, namely the change curve of the error between the learned parameter and the standard parameter in the model training process, and indicating that the error of the model cannot be further reduced when the error vibrates around a constant value; this indicates that the classification and positioning errors have been substantially corrected successfully;
meanwhile, observing and calculating the overall performance indexes of the model along with the increase of the training times, wherein the overall performance indexes comprise an intersection ratio IOU, Precision, Recall rate Recall and accurate Precision AP; when the overall performance index gradually rises and finally is stabilized to fluctuate near a higher value, and the average accuracy of the precision, the recall rate and the model prediction reaches more than 0.9; the required deep neural network model is obtained through representation;
and (4): and (3) adjusting the resolution of an effective video or picture of the intelligent detection system obtained by the camera of the linked alarm relay triggering ball machine to 448 multiplied by 448, dividing the adjusted picture into 7 multiplied by 7 grids, inputting the 7 multiplied by 7 grids into the deep neural network model obtained in the step (3) for target detection to obtain a plurality of candidate results, finally obtaining a final result by using a non-maximum inhibition algorithm, and outputting the position parameters, the category and the positioning result of the boundary frame of the target detection.
Further, the defense area optical fiber is a distributed single-mode optical fiber.
Further, the optical fiber host generates an alarm signal by adopting the Michelson interference principle.
Further, the specific calculation algorithm of the corresponding center position, width and height in the step (1) of training the deep neural network to construct the target recognition model is as follows:
firstly, determining the absolute height and the absolute width of a standard frame and the absolute coordinate of the central position of the standard frame;
secondly, the height of the standard frame is the ratio of the absolute height to the resolution, and the width of the standard frame is the ratio of the absolute width to the resolution in the same way; the width and the height of the standard frame can be obtained according to the standard frame;
the abscissa of the central position is the ratio of the difference between the absolute abscissa and the grid cell width to the grid cell width, and similarly, the ordinate of the central position is the ratio of the difference between the absolute ordinate and the grid cell height to the grid cell height; from this, the abscissa and ordinate of the center position of the standard frame can be obtained.
Further, in the step S5 of constructing the YOLO target detection model in the step (2), the leakage correction linear activation function is:
Figure GDA0002960095550000061
in the formula, q is the output of the previous layer, and the output mapping f (q) is used as the input of the next layer through the action of the activation function.
Further, the Precision is specifically the probability of the true positive class in the sample predicted as the positive class;
the Recall rate Recall is the probability of being predicted as the positive class in all the positive classes;
the accuracy AP is specifically the accuracy of the model prediction.
Furthermore, in the second step of processing the video information and the picture by the server host, the non-maximum suppression algorithm is to search a local maximum and suppress non-maximum elements; when the non-maximum suppression algorithm generates multiple candidate boxes for a target, all boxes are sorted from large to small according to score, and for each box, if the box is not suppressed, all boxes with IOU larger than thresh are set as suppression, and finally, the non-suppressed boxes are returned.
Further, in the step (3) of correcting the classification and positioning errors, a LOSS value is calculated according to a LOSS function of the YOLO algorithm, and the specific LOSS value is as follows:
Figure GDA0002960095550000071
in the formula (I), the compound is shown in the specification,
Figure GDA0002960095550000072
indicating whether the target is present in grid cell i,
Figure GDA0002960095550000073
the jth bounding box predictor represented in grid cell i is "responsible" for the prediction; that is, if there is an object in grid cell i, the value of the jth bounding box predictor is valid for that prediction
Figure GDA0002960095550000074
Otherwise
Figure GDA0002960095550000075
Figure GDA0002960095550000076
If there is no target in grid cell i, then
Figure GDA0002960095550000077
S2B represents the number of divided grid cells, and B represents the number of bounding boxes; (x, y) is the position of the predicted target relative to the network element bounding box,
Figure GDA0002960095550000078
is the actual position of the target relative to the network element bounding box; (w, h) is the width and height of the predicted bounding box,
Figure GDA0002960095550000079
is the actual width and height of the bounding box; c represents the score of the confidence level and,
Figure GDA00029600955500000710
representing the intersection of the predicted bounding box and the standard box; p is a radical ofi(c) The category of the prediction is represented by,
Figure GDA00029600955500000711
representing the actual category;
λcoordloss weight, λ, for bounding box confidence prediction involving objectsnoobjA loss weight predicted for the bounding box confidence that does not contain the target; in this model λ is usedcoord＝5，λnoobj＝0.5。
Further, the loss function consists of three parts, namely coordinate loss, confidence coefficient loss and category loss;
the coordinate loss comprises a position and a width and a height, and specifically comprises the following steps:
Figure GDA0002960095550000081
namely to S2Calculating the loss of the predicted positions, widths and heights of B bounding boxes of the divided cells; because the error of the width and the height of the large object prediction boundary box is larger than the error of the width and the height of the small object prediction boundary box, the root sign is carried out on the width and the height in the loss function, and the tiny position deviation of the large boundary box is solved;
the confidence loss is specifically:
Figure GDA0002960095550000082
namely to S2Loss of confidence scores for B bounding box predictions for each divided cell; the former represents the loss of confidence score predicted for each bounding box of a cell containing a detection target, and the latter represents the loss of confidence score predicted for each bounding box of a cell not containing a detection target;
the category loss is specifically:
Figure GDA0002960095550000083
i.e. represent pairs S2The class loss of the detection target predicted by each divided cell.
The invention has the beneficial effects that:
the invention is a distributed cable anti-external-damage system, which adopts distributed optical fiber along with cable laying to realize the sensing triggering function of cable laying full-range anti-external-damage; then, awakening the standby camera by adopting the trigger signal, and capturing a scene at a preset position monitoring position; and then uploading the scene picture to a target detection server, identifying and acquiring a behavior subject of the intrusion event on the server, and judging behavior damage. The device can be in an all-weather unattended state, and the line is laid to the full coverage cable, triggers through the vibration and realizes the standby awakening, possesses lower consumption, is fit for open-air scenes such as photovoltaic power supply. Particularly, the device utilizes computer vision and image processing related target detection technology, can automatically identify the type of the invading object, and provides an effective monitoring means for the detection of the smart grid.
Drawings
In order to facilitate understanding for those skilled in the art, the present invention will be further described with reference to the accompanying drawings.
FIG. 1 is a system block diagram of an anti-vandalism system of the present invention;
FIG. 2 is a schematic diagram of sample picture processing on an ImageNet race dataset according to the present invention;
FIG. 3 is a schematic diagram of an original neural network model;
FIG. 4 is a flow chart of training a deep neural network to build a target recognition model;
FIG. 5 is a schematic diagram of the related structure of the present invention.
Detailed Description
As shown in fig. 1-5, a distributed cable anti-external-damage system based on video intelligent detection comprises a defense area optical fiber, an optical fiber host, a linkage alarm relay, a dome camera, a server and an alarm;
the defense area optical fiber consists of a sensing optical fiber and a reference optical fiber, an optical loop is built according to the Michelson interference principle, and the emission of coherent light waves and the receiving of interference light waves are realized on an optical fiber host;
the optical fiber host comprises a laser light source and a photoelectric detector, is used for realizing the detection of interference light intensity and generates an alarm signal according to the strength of the signal;
the linkage alarm relay is connected with a 5V direct current power supply in an external connection mode in a dry contact normally open mode, is connected after an external break trigger signal is generated, and gives +5V level to the camera of the ball machine to awaken the camera of the ball machine from a dormant state;
the camera of the dome camera is internally provided with a 4G module which is used for shooting pictures and acquiring field video information of a monitored area and sending the shot pictures and the acquired video information to the server host through a 4G wireless network;
the server host is used for receiving the pictures shot by the camera of the dome camera and storing the pictures into a designated folder;
the server host is used for starting analysis after detecting the updated photos in the folder; identifying an external damage event by using a target detection technology related to computer vision and image processing, identifying a picture containing the external damage event, and determining the type of an invading object; the generated identification picture is displayed on the server host;
the server host is also internally provided with AI image recognition program software and is also used for driving a camera of the ball control machine, actively awakening the camera, acquiring on-site video information and shot pictures and analyzing the video information and the pictures;
the alarm is continuously sounded by a buzzer to give an alarm after the occurrence of an external damage event is confirmed;
the processing steps of the server host computer on the video information and the pictures are specifically as follows:
the method comprises the following steps: firstly, preprocessing video information or pictures and adjusting the size of the pictures to be proper;
step two: then the signal is transmitted into a convolutional neural network to obtain a plurality of candidate results, and finally a non-maximum suppression algorithm is used to obtain a final result; the AI image recognition program software unifies the bounding boxes of the target detection, the confidence score and the conditional category probability into an independent neural network, and the neural network extracts the characteristics of the whole image to predict the position parameters of each bounding box and the target category to which the bounding box belongs;
the method for training the deep neural network to construct the target recognition model comprises the following specific steps:
step (1): adjusting the resolution of the sample pictures on the ImageNet competition dataset to 448 x 448, and dividing the adjusted pictures into 7 x 7 grids; selecting a boundary frame in a sample picture as a standard frame of a prediction boundary frame for target detection, and marking a label on a grid unit of the sample picture, wherein the label contains the central position, the width, the height and the category of a target;
step (2): constructing a YOLO target detection model; the specific process is as follows:
s1: pre-training with the first 20 convolutional layers and 4 maximum pooling layers;
s2: migrating the classification task to target detection;
s3: adding 4 convolutional layers and 2 full-connection layers on the basis of the original 20 convolutional layers and 4 maximum pooling layers, starting to initialize network parameters randomly, forming a YOLO target detection model, and obtaining an original neural network model; the convolutional layer is responsible for extracting the characteristics of the detection target; the first 20 convolutional layers, the 4 largest pooling layers, and the last 4 convolutional layers and 2 fully-connected layers form a complete classification network;
s4: the last layer of full connection layer predicts the coordinate and the class probability of the bounding box; the YOLO algorithm normalizes the width and height of the bounding box by the width and height of the image to obtain w, h, and converts the center position coordinates of the bounding box into the offset of the center position relative to the corresponding network grid unit position, namely, the x, y is obtained by the same normalization;
predicting a boundary frame of the target, wherein the boundary frame of the target is output by the sample picture through a neural network model, namely a YOLO model, and comprises a central coordinate, a width and a height of the target relative to a boundary frame of a network unit; and the confidence of the intersection ratio between the bounding box and the standard box, namely the probability and the accuracy of whether the bounding box of the reaction prediction contains the target or not;
wherein the cross-over ratio is defined as: the intersection of the coverage ranges of the prediction frame and the standard frame and the ratio of the union of the coverage ranges of the prediction frame and the standard frame are determined, wherein the prediction frame is a detection frame output by the model, and the standard frame is a detection frame marked manually; if the prediction frame and the standard frame are completely overlapped, the intersection ratio is 1;
confidence is defined as
Figure GDA0002960095550000111
If the target is not in the grid, the confidence coefficient is 0, and if the target is in the grid, the intersection ratio is the value of the confidence coefficient;
predicting conditional Class probability Pr (Class) per grid celliI Object) and multiplies the conditional class probability by the confidence of the single bounding box prediction:
Figure GDA0002960095550000112
wherein, Pr (object) represents the probability that a certain grid cell contains the detection target, and Pr (Class)iI Object) represents a conditional category probability of the lattice cell for the detection target; pr (Class)i) Representing the probability that a certain grid cell contains a certain class of detection target;
Figure GDA0002960095550000113
the intersection ratio of the coverage ranges of a prediction box and an actual detection box which represent the output of the model;
providing a class-specific confidence score of each bounding box by the multiplied value, expressing the probability of the class appearing in the bounding box and evaluating the quality degree of the bounding box containing the target;
s5: adopting a linear activation function as an activation function of the last full-connection layer, wherein the activation functions of other convolution layers, pooling layers and full-connection layers are leakage correction linear activation functions;
and (3): correcting classification and positioning errors;
adjusting the resolution of the sample picture to 448 × 448, dividing the adjusted picture into 7 × 7 grids, inputting the 7 × 7 grids into the YOLO target detection model obtained in the step (2), namely the original neural network model, obtaining the LOSS value of the original neural network, and appropriately changing the network weight value initialized randomly to obtain a new neural network model; fine-tuning the network weight value through multiple training iterations, observing the change of the error under the promotion of training and fine tuning by adopting an LOSS change curve of a neural network, namely the change curve of the error between the learned parameter and the standard parameter in the model training process, and indicating that the error of the model cannot be further reduced when the error vibrates around a constant value; this indicates that the classification and positioning errors have been substantially corrected successfully;
meanwhile, observing and calculating the overall performance indexes of the model along with the increase of the training times, wherein the overall performance indexes comprise an intersection ratio IOU, Precision, Recall rate Recall and accurate Precision AP; when the overall performance index gradually rises and finally is stabilized to fluctuate near a higher value, and the average accuracy of the precision, the recall rate and the model prediction reaches more than 0.9; the required deep neural network model is obtained through representation;
and (4): and (3) adjusting the resolution of an effective video or picture of the intelligent detection system obtained by the camera of the linked alarm relay triggering ball machine to 448 multiplied by 448, dividing the adjusted picture into 7 multiplied by 7 grids, inputting the 7 multiplied by 7 grids into the deep neural network model obtained in the step (3) for target detection to obtain a plurality of candidate results, finally obtaining a final result by using a non-maximum inhibition algorithm, and outputting the position parameters, the category and the positioning result of the boundary frame of the target detection.
The defense area optical fiber is a distributed single-mode optical fiber.
The optical fiber host generates an alarm signal by adopting the Michelson interference principle.
The specific calculation algorithm of the corresponding center position, width and height in the step (1) of training the deep neural network to construct the target recognition model is as follows:
firstly, determining the absolute height and the absolute width of a standard frame and the absolute coordinate of the central position of the standard frame;
secondly, the height of the standard frame is the ratio of the absolute height to the resolution, and the width of the standard frame is the ratio of the absolute width to the resolution in the same way; the width and the height of the standard frame can be obtained according to the standard frame;
the abscissa of the central position is the ratio of the difference between the absolute abscissa and the grid cell width to the grid cell width, and similarly, the ordinate of the central position is the ratio of the difference between the absolute ordinate and the grid cell height to the grid cell height; from this, the abscissa and ordinate of the center position of the standard frame can be obtained.
In step S5 of constructing the YOLO target detection model in step (2), the leakage correction linear activation function is:
Figure GDA0002960095550000131
in the formula, q is the output of the previous layer, and a mapping f (q) is output as the input of the next layer under the action of an activation function;
the Precision is specifically the probability of truly being the positive type in the samples predicted to be the positive type;
the Recall rate Recall is the probability of being predicted as the positive class in all the positive classes;
the accuracy AP is specifically the accuracy of the model prediction.
In the second step of processing the video information and the pictures by the server host, the non-maximum suppression algorithm is to search a local maximum and suppress non-maximum elements; when the non-maximum suppression algorithm generates multiple candidate boxes for a target, all boxes are sorted from large to small according to score, and for each box, if the box is not suppressed, all boxes with IOU larger than thresh are set as suppression, and finally, the non-suppressed boxes are returned.
In the step (3) of correcting the classification and positioning errors, calculating a LOSS value according to a LOSS function of a YOLO algorithm, wherein the specific LOSS value is as follows:
Figure GDA0002960095550000141
in the formula (I), the compound is shown in the specification,
Figure GDA0002960095550000142
indicating whether the target is present in grid cell i,
Figure GDA0002960095550000143
the jth bounding box predictor represented in grid cell i is "responsible" for the prediction; that is, if there is an object in grid cell i, the value of the jth bounding box predictor is valid for that prediction
Figure GDA0002960095550000144
Otherwise
Figure GDA0002960095550000145
Figure GDA0002960095550000146
If there is no target in grid cell i, then
Figure GDA0002960095550000147
S2B represents the number of divided grid cells, and B represents the number of bounding boxes; (x, y) is the position of the predicted target relative to the network element bounding box,
Figure GDA0002960095550000148
is the actual position of the target relative to the network element bounding box; (w, h) is the width and height of the predicted bounding box,
Figure GDA0002960095550000149
is the actual width and height of the bounding box; c represents the score of the confidence level and,
Figure GDA00029600955500001410
representing the intersection of the predicted bounding box and the standard box; p is a radical ofi(c) The category of the prediction is represented by,
Figure GDA00029600955500001411
representing the actual category;
λcoordloss weight, λ, for bounding box confidence prediction involving objectsnoobjA loss weight predicted for the bounding box confidence that does not contain the target; in this model λ is usedcoord＝5，λnoobj＝0.5。
The loss function consists of three parts, namely coordinate loss, confidence coefficient loss and category loss;
the coordinate loss comprises a position and a width and a height, and specifically comprises the following steps:
Figure GDA00029600955500001412
namely to S2Calculating the loss of the predicted positions, widths and heights of B bounding boxes of the divided cells; because the error of the width and the height of the large object prediction boundary box is larger than the error of the width and the height of the small object prediction boundary box, the root sign is carried out on the width and the height in the loss function, and the tiny position deviation of the large boundary box is solved;
the confidence loss is specifically:
Figure GDA0002960095550000151
namely to S2Loss of confidence scores for B bounding box predictions for each divided cell; the former represents the loss of confidence score predicted for each bounding box of a cell containing a detection target, and the latter represents the loss of confidence score predicted for each bounding box of a cell not containing a detection target;
the category loss is specifically:
Figure GDA0002960095550000152
i.e. represent pairs S2The class loss of the detection target predicted by each divided cell.
In a specific implementation process, the method comprises the following steps:
1) defense area optical fiber: the defense area optical fiber consists of a sensing optical fiber and a reference optical fiber, an optical loop is built according to the Michelson interference principle, and the emission of coherent light waves and the receiving of interference light waves are realized on an optical fiber host;
2) the optical fiber host: the detector comprises main components such as a laser light source, a photoelectric detector and the like, is used for realizing the detection of interference light intensity and generating an alarm signal according to the intensity of the signal;
3) linkage alarm relay: a dry contact normally open mode is adopted, a 5V direct current power supply is externally connected, the power supply is switched on after an external break trigger signal is generated, a +5V level is given to a camera of the ball machine, the camera of the ball machine is awakened from a dormant state, and an intelligent analysis mode is entered;
4) the camera of the ball machine: the cloud deck is preset and has a photographing and uploading function; the camera of the dome camera is internally provided with a 4G module, and a shot picture is sent to a server host through a 4G wireless network;
5) a server host; the server receives the pictures shot by the camera of the ball machine and stores the pictures in the designated folder; after detecting the updated photos in the folder, the AI identification program starts analysis; identifying an external damage event by using a target detection technology related to computer vision and image processing, identifying a picture containing the external damage event, and determining the type of an invading object; typically, the host server is placed in a fixed location, such as a main control room of the power maintenance department.
6) A user side software platform: the software platform is installed on a host server, and an AI image recognition program runs on the server to generate an identification picture to be displayed on the platform;
the software platform also provides a management function of the dome camera, so that the camera can be actively awakened manually, and a field video and a photographing analysis can be checked;
7) an alarm device: after the occurrence of the external damage event is confirmed, the buzzer continuously sounds to alarm.
Wherein the defense area optical fiber is a distributed single-mode optical fiber;
the optical fiber host generates an alarm signal by adopting the Michelson interference principle;
the camera of the dome camera can be controlled by a holder, can be provided with a preset position, can be accessed by a 4G communication module, and can upload a shot picture to the server host through the 4G communication module;
the server host can have certain calculation and analysis capacity, and can be provided with 'AI image recognition program software' required by the system;
the AI image identification program software installed on the server host computer processes the video information and the pictures, and specifically comprises the following steps:
the method comprises the following steps: firstly, preprocessing video information or pictures and adjusting the size of the pictures to be proper;
step two: then the signal is transmitted into a convolutional neural network to obtain a plurality of candidate results, and finally a non-maximum suppression algorithm is used to obtain a final result, wherein the non-maximum suppression algorithm is a local optimal method; the AI image recognition program software unifies the bounding boxes of the target detection, the confidence score and the conditional category probability into an independent neural network, and the neural network extracts the characteristics of the whole image to predict the position parameters of each bounding box and the target category to which the bounding box belongs; that is, the picture is input into the trained model, and the model outputs the position parameters of each target in the picture and the target class to which the target belongs. In short, the network can predict all targets in the picture at once. Therefore, the AI image recognition software based on the convolutional neural network can still meet the requirement of higher average accuracy while ensuring the processing speed; through training of 16000 batches of target detection in the model, the accuracy rate, recall rate and average accuracy of model prediction are all over 0.9.
Wherein, the non-maximum suppression algorithm is the prior art; the non-maximum suppression algorithm is to search local maximum and suppress non-maximum elements; when the algorithm generates multiple candidate boxes for a target, all boxes are sorted by score from large to small, and for each box, if it is not suppressed, all boxes with IOU greater than thresh are set as suppressed, and finally the non-suppressed boxes are returned.
The method for training the deep neural network to construct the target recognition model specifically comprises the following steps:
step (1): as shown in fig. 2, the resolution of the sample pictures on the ImageNet contest dataset is resized to 448 × 448, and the resized pictures are divided into a 7 × 7 grid; selecting a boundary frame in a sample picture as a standard frame of a prediction boundary frame for target detection, and marking a label on a grid unit of the sample picture in a manual mode, wherein the label contains the central position, the width, the height and the category of a target; the specific calculation algorithm for the center position, width and height is as follows:
firstly, determining the absolute height and the absolute width of a standard frame and the absolute coordinate of the central position of the standard frame;
secondly, the height of the standard frame is the ratio of the absolute height to the resolution, and the width of the standard frame is the ratio of the absolute width to the resolution in the same way; the abscissa of the central position is the ratio of the difference between the absolute abscissa and the grid cell width to the grid cell width, and similarly, the ordinate of the central position is the ratio of the difference between the absolute ordinate and the grid cell height to the grid cell height.
Step (2): constructing a YOLO target detection model; the specific process is as follows:
s1: pre-training with the first 20 convolutional layers and 4 maximum pooling layers;
s2: migrating the classification task to target detection;
s3: adding 4 convolutional layers and 2 full-connection layers on the basis of the original 20 convolutional layers, and starting to initialize network parameters randomly to form an original neural network model; the convolutional layer is responsible for extracting the characteristics of the detection target. As shown in fig. 3, conv. The first 20 convolutional layers, 4 max pooling layers, and the last 4 convolutional layers and 2 fully-connected layers constitute a complete classification network.
S4: the last layer of full connection layer predicts the coordinate and the class probability of the bounding box; the YOLO algorithm normalizes the width and height of the bounding box by the width and height of the image and translates the coordinates of the bounding box into an offset of the corresponding network location.
Predicting the bounding box of the target, including the center coordinates, width and height relative to the bounding box of the network element, and the confidence of the intersection ratio between the bounding box and the standard box, i.e., the likelihood and accuracy of whether the predicted bounding box contains the target; the sample picture outputs a boundary frame of the target through a neural network model, namely a YOLO model, wherein the boundary frame comprises the central coordinate, the width, the height and the like of the target relative to a boundary frame of a network unit; .
Wherein the cross-over ratio is defined as: the intersection of the coverage ranges of the prediction frame and the standard frame and the ratio of the union of the coverage ranges of the prediction frame and the standard frame are determined, wherein the prediction frame is a detection frame output by the model, and the standard frame is a detection frame marked manually;
if the prediction frame and the standard frame are completely overlapped, the intersection ratio is 1; confidence is defined as
Figure GDA0002960095550000181
If the target is not in the grid, the confidence is 0, and if the target is in the grid, the intersection ratio is the value of the confidence.
Predicting conditional Class probability Pr (Class) per grid celli| Object) and combines the conditional category probability with a single bounding boxConfidence of prediction multiplied:
Figure GDA0002960095550000182
wherein Pr (object) represents the probability that a certain grid cell contains a detection target, and Pr (Class)iI Object) represents the conditional category probability of the lattice cell for the detection target. Pr (Class)i) Indicating the probability that a certain grid cell contains a certain class of detection target.
Figure GDA0002960095550000191
The ratio of the intersection and union of the coverage ranges of the prediction frame and the actual detection frame output by the model is represented;
providing a class-specific confidence score of each bounding box by the multiplied value, expressing the probability of the class appearing in the bounding box and evaluating the quality degree of the bounding box containing the target;
s5: and adopting a linear activation function as an activation function of the last full-connection layer, wherein the activation functions of other convolution layers, pooling layers and full-connection layers are leakage correction linear activation functions.
Wherein the leakage correction linear activation function:
Figure GDA0002960095550000192
wherein x is the output of the previous layer, and the output map f (x) is used as the input of the next layer through the action of the activation function.
The YOLO target detection model is a network structure composed of the above-mentioned 20 convolutional layers, 4 max pooling layers, and the last 4 convolutional layers and 2 full connection layers, and is an existing and widely used target detection model; the layers are connected through the activation function, so that the mapping relation of the input and the output of the whole network structure is formed. However, the network weight values of each layer are initialized randomly at first, that is, so-called original neural network models are obtained, but the network weight values are not accurate, so that iterative training is needed to modify the network weight values, and finally the final neural network model is obtained.
And (3): correcting classification and positioning errors;
and (3) adjusting the resolution of the sample picture to 448 multiplied by 448, dividing the adjusted picture into 7 multiplied by 7 grids, inputting the obtained YOLO target detection model in the step (2), namely the original neural network model, to obtain the LOSS value of the original neural network, and properly changing the randomly initialized network weight value to obtain a new neural network model. And fine tuning the network weight value through multiple training iterations, observing the change of the error under the promotion of training and fine tuning by adopting an LOSS change curve of the neural network, namely the change curve of the error between the learned parameter and the standard parameter in the model training process, and indicating that the error of the model cannot be further reduced when the error vibrates around a constant value, wherein the model at the moment has strong representation capability and can meet the required requirements. It indicates that the classification and positioning errors have been successfully corrected.
Meanwhile, with the increase of the training times, observing and calculating performance indexes such as an intersection ratio IOU, Precision, Recall, accurate Precision AP and the like of the model;
when each performance index gradually rises and finally stably fluctuates around a higher value, and most performance indexes reach more than 0.9; obtaining a required deep neural network model;
wherein, the Precision ratio Precision is the probability of truly being the positive type in the samples predicted as the positive type; the Recall rate Recall is the probability of being predicted as the positive class in all the positive classes; the accurate precision AP is the accurate precision of model prediction;
the specific method for calculating the LOSS value by the LOSS function of the YOLO algorithm comprises the following steps:
Figure GDA0002960095550000201
in the formula (I), the compound is shown in the specification,
Figure GDA0002960095550000202
indicating whether the target is present in grid cell i,
Figure GDA0002960095550000203
the jth bounding box predictor represented in grid cell i is "responsible" for the prediction; that is, if there is an object in grid cell i, the value of the jth bounding box predictor is valid for that prediction
Figure GDA0002960095550000204
Otherwise
Figure GDA0002960095550000205
If there is no target in grid cell i, then
Figure GDA0002960095550000206
S2B represents the number of divided grid cells, and B represents the number of bounding boxes; (x, y) is the position of the predicted target relative to the network element bounding box,
Figure GDA0002960095550000207
is the actual position of the target relative to the network element bounding box; (w, h) is the width and height of the predicted bounding box,
Figure GDA0002960095550000208
is the actual width and height of the bounding box; c represents the score of the confidence level and,
Figure GDA0002960095550000209
representing the intersection of the predicted bounding box and the standard box; p is a radical ofi(c) The category of the prediction is represented by,
Figure GDA00029600955500002010
representing the actual category;
λcoordloss weight, λ, for bounding box confidence prediction involving objectsnoobjBounding box confidence for no-object-containingPredicted loss weight. In this model λ is usedcoord＝5，λnoobj＝0.5。
The loss function consists of three parts, namely coordinate loss, confidence coefficient loss and category loss;
the coordinate loss comprises a position and a width and a height, and specifically comprises the following steps:
Figure GDA0002960095550000211
namely to S2Calculating the loss of the predicted positions, widths and heights of B bounding boxes of the divided cells; because the error of the width and the height of the large object prediction boundary box is larger than the error of the width and the height of the small object prediction boundary box, the root sign is carried out on the width and the height in the loss function, and the tiny position deviation of the large boundary box is solved;
the confidence loss is specifically:
Figure GDA0002960095550000212
namely to S2Loss of confidence scores for B bounding box predictions for each divided cell; the former represents the loss of confidence score predicted for each bounding box of a cell containing a detection target, and the latter represents the loss of confidence score predicted for each bounding box of a cell not containing a detection target;
the category loss is specifically:
Figure GDA0002960095550000213
i.e. represent pairs S2The class loss of the detection target predicted by each divided cell.
And (4): and adjusting the resolution of an effective video or picture of the intelligent detection system obtained by the camera of the linked alarm relay triggering ball machine to 448 multiplied by 448, dividing the adjusted picture into 7 multiplied by 7 grids, inputting the 7 multiplied by 7 grids into the deep neural network model obtained in the third step for target detection to obtain a plurality of candidate results, finally obtaining a final result by using a non-maximum suppression algorithm, and outputting the position parameters, the category and the positioning result of the boundary frame of the target detection.
A distributed cable anti-external-damage system based on video intelligent monitoring is a distributed cable anti-external-damage system, and a sensing triggering function of preventing external damage in the whole process of cable laying is realized by adopting distributed optical fibers along with cable laying; then, awakening the standby camera by adopting the trigger signal, and capturing a scene at a preset position monitoring position; and then uploading the scene picture to a target detection server, identifying and acquiring a behavior subject of the intrusion event on the server, and judging behavior damage. The device can be in an all-weather unattended state, and the line is laid to the full coverage cable, triggers through the vibration and realizes the standby awakening, possesses lower consumption, is fit for open-air scenes such as photovoltaic power supply. Particularly, the device utilizes computer vision and image processing related target detection technology, can automatically identify the type of the invading object, and provides an effective monitoring means for the detection of the smart grid.
The foregoing is merely exemplary and illustrative of the present invention and various modifications, additions and substitutions may be made by those skilled in the art to the specific embodiments described without departing from the scope of the invention as defined in the following claims.

Claims (1)

1. The distributed cable external-damage-prevention system based on video intelligent detection is characterized by comprising a defense area optical fiber, an optical fiber host, a linkage alarm relay, a dome camera, a server and an alarm;
the defense area optical fiber consists of a sensing optical fiber and a reference optical fiber, an optical loop is built according to the Michelson interference principle, and the emission of coherent light waves and the receiving of interference light waves are realized on an optical fiber host;
the optical fiber host comprises a laser light source and a photoelectric detector, is used for realizing the detection of interference light intensity and generates an alarm signal according to the strength of the signal;
the linkage alarm relay is connected with a 5V direct current power supply in an external connection mode in a dry contact normally open mode, is connected after an external break trigger signal is generated, and gives +5V level to the camera of the ball machine to awaken the camera of the ball machine from a dormant state;
the camera of the dome camera is internally provided with a 4G module which is used for shooting pictures and acquiring field video information of a monitored area and sending the shot pictures and the acquired video information to the server host through a 4G wireless network;
the server host is used for receiving the pictures shot by the camera of the dome camera and storing the pictures into a designated folder;
the server host is used for starting analysis after detecting the updated photos in the folder; identifying an external damage event by using a target detection technology related to computer vision and image processing, identifying a picture containing the external damage event, and determining the type of an invading object; generating an identification picture, and displaying the identification picture on the server host;
the server host is also internally provided with AI image recognition program software and is also used for driving a camera of the ball control machine, actively awakening the camera, acquiring on-site video information and shot pictures and analyzing the video information and the pictures;
the alarm is continuously sounded by a buzzer to give an alarm after the occurrence of an external damage event is confirmed;
the processing steps of the server host computer on the video information and the pictures are specifically as follows:
the method comprises the following steps: firstly, preprocessing video information or pictures and adjusting the size of the pictures to be proper;
step two: then the signal is transmitted into a convolutional neural network to obtain a plurality of candidate results, and finally a non-maximum suppression algorithm is used to obtain a final result; the AI image recognition program software unifies the bounding boxes of the target detection, the confidence score and the conditional category probability into an independent neural network, and the neural network extracts the characteristics of the whole image to predict the position parameters of each bounding box and the target category to which the bounding box belongs;
the method for training the deep neural network to construct the target recognition model comprises the following specific steps:
step (1): adjusting the resolution of the sample pictures on the ImageNet competition dataset to 448 x 448, and dividing the adjusted pictures into 7 x 7 grids; selecting a boundary frame in a sample picture as a standard frame of a prediction boundary frame for target detection, and marking a label on a grid unit of the sample picture, wherein the label contains the central position, the width, the height and the category of a target;
step (2): constructing a YOLO target detection model; the specific process is as follows:
s1: pre-training with the first 20 convolutional layers and 4 maximum pooling layers;
s2: migrating the classification task to target detection;
s3: adding 4 convolutional layers and 2 full-connection layers on the basis of the original 20 convolutional layers and 4 maximum pooling layers, starting to initialize network parameters randomly, forming a YOLO target detection model, and obtaining an original neural network model; the convolutional layer is responsible for extracting the characteristics of the detection target; the first 20 convolutional layers, the 4 largest pooling layers, and the last 4 convolutional layers and 2 fully-connected layers form a complete classification network;
s4: the last layer of full connection layer predicts the coordinate and the class probability of the bounding box; the YOLO algorithm normalizes the width and height of the bounding box by the width and height of the image to obtain w, h, and converts the center position coordinates of the bounding box into the offset of the center position relative to the corresponding network grid position, namely, the x, y is obtained by the same normalization;
predicting a boundary frame of the target, wherein the boundary frame of the target is output by the sample picture through a neural network model, namely a YOLO model, and comprises a central coordinate, a width and a height of the target relative to a boundary frame of a network unit; and the confidence of the intersection ratio between the bounding box and the standard box, namely the probability and the accuracy of whether the bounding box of the reaction prediction contains the target or not;
wherein the cross-over ratio is defined as: the intersection of the coverage ranges of the prediction frame and the standard frame and the ratio of the union of the coverage ranges of the prediction frame and the standard frame are determined, wherein the prediction frame is a detection frame output by the model, and the standard frame is a detection frame marked manually; if the prediction frame and the standard frame are completely overlapped, the intersection ratio is 1;
confidence is defined as
Figure FDA0002960095540000031
If the target is not in the grid, the confidence coefficient is 0, and if the target is in the grid, the intersection ratio is the value of the confidence coefficient;
predicting conditional Class probability Pr (Class) per grid celliI Object) and multiplies the conditional class probability by the confidence of the single bounding box prediction:
Figure FDA0002960095540000032
wherein, Pr (object) represents the probability that a certain grid cell contains the detection target, and Pr (Class)iI Object) represents a conditional category probability of the lattice cell for the detection target; pr (Class)i) Representing the probability that a certain grid cell contains a certain class of detection target;
Figure FDA0002960095540000033
the intersection ratio of the coverage ranges of a prediction box and an actual detection box which represent the output of the model;
providing a class-specific confidence score of each bounding box by the multiplied value, expressing the probability of the class appearing in the bounding box and evaluating the quality degree of the bounding box containing the target;
s5: adopting a linear activation function as an activation function of the last full-connection layer, wherein the activation functions of other convolution layers, pooling layers and full-connection layers are leakage correction linear activation functions;
and (3): correcting classification and positioning errors;
adjusting the resolution of the sample picture to 448 × 448, dividing the adjusted picture into 7 × 7 grids, inputting the 7 × 7 grids into the YOLO target detection model obtained in the step (2), namely the original neural network model, obtaining the LOSS value of the original neural network, and appropriately changing the network weight value initialized randomly to obtain a new neural network model; fine-tuning the network weight value through multiple training iterations, observing the change of the error under the promotion of training and fine tuning by adopting an LOSS change curve of a neural network, namely the change curve of the error between the learned parameter and the standard parameter in the model training process, and indicating that the error of the model cannot be further reduced when the error vibrates around a constant value; this indicates that the classification and positioning errors have been substantially corrected successfully;
meanwhile, observing and calculating the overall performance indexes of the model along with the increase of the training times, wherein the overall performance indexes comprise an intersection ratio IOU, Precision, Recall rate Recall and accurate Precision AP; when the overall performance index gradually rises and finally is stabilized to fluctuate near a higher value, and the average accuracy of the precision, the recall rate and the model prediction reaches more than 0.9; the required deep neural network model is obtained through representation;
and (4): adjusting the resolution of an effective video or picture of an intelligent detection system obtained by a camera of the linked alarm relay triggering ball machine to 448 multiplied by 448, dividing the adjusted picture into 7 multiplied by 7 grids, inputting the 7 multiplied by 7 grids into the deep neural network model obtained in the step (3) for target detection to obtain a plurality of candidate results, finally obtaining a final result by using a non-maximum suppression algorithm, and outputting position parameters, categories and positioning results of a boundary frame of the target detection;
the defense area optical fiber is a distributed single-mode optical fiber;
the optical fiber host generates an alarm signal by adopting the Michelson interference principle;
the specific calculation algorithm of the corresponding center position, width and height in the step (1) of training the deep neural network to construct the target recognition model is as follows:
firstly, determining the absolute height and the absolute width of a standard frame and the absolute coordinate of the central position of the standard frame;
secondly, the height of the standard frame is the ratio of the absolute height to the resolution, and the width of the standard frame is the ratio of the absolute width to the resolution in the same way; the width and the height of the standard frame can be obtained according to the standard frame;
the abscissa of the central position is the ratio of the difference between the absolute abscissa and the grid cell width to the grid cell width, and similarly, the ordinate of the central position is the ratio of the difference between the absolute ordinate and the grid cell height to the grid cell height; accordingly, the horizontal coordinate and the vertical coordinate of the central position of the standard frame can be obtained;
in step S5 of constructing the YOLO target detection model in step (2), the leakage correction linear activation function is:
Figure FDA0002960095540000051
in the formula, q is the output of the previous layer, and a mapping f (q) is output as the input of the next layer under the action of an activation function;
the Precision is specifically the probability of truly being the positive type in the samples predicted to be the positive type;
the Recall rate Recall is the probability of being predicted as the positive class in all the positive classes;
the accurate precision AP is the accurate precision of model prediction;
in the second step of processing the video information and the pictures by the server host, the non-maximum suppression algorithm is to search local maximum and suppress non-maximum elements; when the non-maximum suppression algorithm generates a plurality of candidate frames for a target, all the frames are sorted from large to small according to score, and for each frame, if the frame is not suppressed, all the frames with IOU larger than thresh are set as suppression, and finally the frames which are not suppressed are returned;
in the step (3) of correcting the classification and positioning errors, calculating an LOSS value according to a LOSS function of a YOLO algorithm, wherein the specific LOSS value is as follows:
Figure FDA0002960095540000052
in the formula (I), the compound is shown in the specification,
Figure FDA0002960095540000053
indicating whether the target is present in grid cell i,
Figure FDA0002960095540000054
the jth bounding box predictor represented in grid cell i is "responsible" for the prediction; that is, if there is an object in grid cell i, the value of the jth bounding box predictor is valid for that prediction
Figure FDA0002960095540000055
Otherwise
Figure FDA0002960095540000056
Figure FDA0002960095540000061
If there is no target in grid cell i, then
Figure FDA0002960095540000062
S2B represents the number of divided grid cells, and B represents the number of bounding boxes; (x, y) is the position of the prediction target relative to the grid cell bounding box,
Figure FDA0002960095540000063
is the actual position of the target relative to the grid cell bounding box; (w, h) is the width and height of the predicted bounding box,
Figure FDA0002960095540000064
is the actual width and height of the bounding box; c represents the score of the confidence level and,
Figure FDA0002960095540000065
representing the intersection of the predicted bounding box and the standard box; p is a radical ofi(c) The category of the prediction is represented by,
Figure FDA0002960095540000066
representing the actual category;
λcoordloss weight, λ, for bounding box confidence prediction involving objectsnoobjA loss weight predicted for the bounding box confidence that does not contain the target; in this model λ is usedcoord＝5，λnoobj＝0.5；
The loss function consists of three parts, namely coordinate loss, confidence coefficient loss and category loss;
the coordinate loss comprises a position and a width and a height, and specifically comprises the following steps:
Figure FDA0002960095540000067
namely to S2Calculating the loss of the predicted positions, widths and heights of B bounding boxes of the divided cells; because the error of the width and the height of the large object prediction boundary box is larger than the error of the width and the height of the small object prediction boundary box, the root sign is carried out on the width and the height in the loss function, and the tiny position deviation of the large boundary box is solved;
the confidence loss is specifically:
Figure FDA0002960095540000068
namely to S2Loss of confidence scores for B bounding box predictions for each divided cell; the former represents the loss of confidence score predicted for each bounding box of a cell containing a detection target, and the latter represents the loss of confidence score predicted for each bounding box of a cell not containing a detection target;
the category loss is specifically:
Figure FDA0002960095540000071
i.e. represent pairs S2The class loss of the detection target predicted by each divided cell.
CN201910604299.4A 2019-07-05 2019-07-05 Distributed cable external-damage-prevention system based on video intelligent detection Active CN110321853B (en)

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 CN110321853A (en) 2019-10-11
CN110321853B true 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)

Families Citing this family (22)

* Cited by examiner, † Cited by third party
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
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
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
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
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
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
CN111914737A (en) * 2020-07-29 2020-11-10 姚广元 Return cable district section early warning intelligent host computer with identification function
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
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
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
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
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

Citations (10)

* Cited by examiner, † Cited by third party
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

Patent Citations (10)

* Cited by examiner, † Cited by third party
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 (3)

* Cited by examiner, † Cited by third party
Title
电力电缆线路回流缆防盗报警装置的研制与应用;黄肖为 等;《科技创新与应用》;20181231;第182-183页 *
电缆线路防外破预警系统应用;陆斌;《中国电力企业管理》;20171231;第42页 *
高压电缆线路现场智能移动巡检终端设计;黄肖为 等;《科技创新导报》;20181231;第1-2页 *

Also Published As

Publication number Publication date
CN110321853A (en) 2019-10-11

Similar Documents

Publication Publication Date Title
CN110321853B (en) Distributed cable external-damage-prevention system based on video intelligent detection
CN109872483B (en) Intrusion alert photoelectric monitoring system and method
US20230005176A1 (en) Throwing position acquisition method and apparatus, computer device and storage medium
Stula et al. Intelligent forest fire monitoring system
CN111339826A (en) Landslide unmanned aerial vehicle linear sensor network frame detection system
KR20200054733A (en) Facility monitoring system using artificial intelligence automatic flight drone and intelligence video analysis
CN112560619A (en) Multi-focus image fusion-based multi-distance bird accurate identification method
US20210104059A1 (en) Method for size estimation by image recognition of specific target using given scale
CN115248873A (en) Cable tunnel safety monitoring method and system based on data fusion
CN116846059A (en) Edge detection system for power grid inspection and monitoring
CN113965733A (en) Binocular video monitoring method, system, computer equipment and storage medium
CN115880598B (en) Ground image detection method and related device based on unmanned aerial vehicle
CN115912183B (en) Ecological measure inspection method and system for high-voltage transmission line and readable storage medium
CN113784085B (en) Monitoring device and method for preventing external damage of power transmission line
CN115546710A (en) Method, device and equipment for locating personnel in hydraulic power plant and readable storage medium
CN112541403B (en) Indoor personnel falling detection method by utilizing infrared camera
CN115294048A (en) Foreign matter detection method, device, equipment and storage medium
CN115187880A (en) Communication optical cable defect detection method and system based on image recognition and storage medium
CN114387391A (en) Safety monitoring method and device for transformer substation equipment, computer equipment and medium
CN111079642B (en) Line movable monitoring method and device and computer readable medium
CN114200468A (en) Positioning method and system of underwater netting inspection robot and storage medium
KR102375380B1 (en) Apparatus for monitoring animals with low power and method thereof
CN113758927B (en) Encoder-decoder small bridge bottom crack detection system and method based on reverse residual bottleneck
CN116013018B (en) Forest fire prevention early warning analysis method and system based on unmanned aerial vehicle detection
CN116739357B (en) Multi-mode fusion perception city existing building wide area monitoring and early warning method and device

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
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