CN108573283A - A kind of anti-design method failed to report of notch of switch machine monitoring - Google Patents
A kind of anti-design method failed to report of notch of switch machine monitoring Download PDFInfo
- Publication number
- CN108573283A CN108573283A CN201810338172.8A CN201810338172A CN108573283A CN 108573283 A CN108573283 A CN 108573283A CN 201810338172 A CN201810338172 A CN 201810338172A CN 108573283 A CN108573283 A CN 108573283A
- Authority
- CN
- China
- Prior art keywords
- data
- personal computer
- industrial personal
- alarm
- image
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
Abstract
The invention belongs to embedded computer technology fields, disclose a kind of anti-design method failed to report of notch of switch machine monitoring, are realized by industrial personal computer processing module, data receiver and memory module, machine learning identification module.Industrial personal computer processing module is wrapped by industrial personal computer routine data packet, video data, gap data is acquired and passes to NVR servers, while updating host computer and showing information.Data receiver and memory module reception industrial personal computer upload data to host computer and are displayed and stored in NVR servers.Machine learning identification module establishes image classification data library, introduce convolutional neural networks, the unit of classification sampling is established to image classification data library, finally by Fast R-CNN output video frames, video frame is handled, type of alarm is exported, warning message is obtained after finally comparing alarm types after detection and industrial personal computer type of alarm.The present invention can effectively prevent goat alert to fail to report, report by mistake, improve the reliability of notch of switch machine video monitoring.
Description
Technical field
The invention belongs to embedded computer fields, are related to a kind of anti-design method failed to report of notch of switch machine monitoring.
Background technology
In recent years, with the fast development of China's economy, the continuous propulsion of urbanization, urban population constantly increases.Train
It is continued to increase as most important traffic machine tool, passenger traffic, freight-transport capacity in the world today, in order to allow railway communication equipment to protect
Good working order is held, tightened up requirement is proposed to later maintenance work.Currently, goat is in each iron in China
The station of road interlocking electric is widely used, it is mainly characterized by that work is frequent, operating condition is severe, job site dispersion, no
Easily centralized management.Notch of switch machine offset reflects the degree of track switch fitting, weighing apparatus whether can precisely being shifted as track switch
Measure foundation.
The electric traction equipment that goat is precisely converted as train track switch.To under normal, mistake vehicle, goat working condition
The accurately monitoring of notch is most important.With the continuous development of rail traffic signal system, notch of switch machine monitoring system is increasingly
Maturation, but there are possibility alert wrong report and failed to report.
The anti-design method failed to report of notch of switch machine monitoring is to rely on network technology, the information of notch of switch machine monitoring system
The method that acquisition technique and image processing techniques are realized.The overall structure of notch of switch machine monitoring system is as shown in Figure 1.Goat
Notch monitoring system includes each goat of train yard, industrial personal computer, power carrier circuit, network commutator, ADSL
(Asymmetric Digital Subscriber Line) communication line, host computer, NVR (Network Video
Recorder) server, dedicated network.Industrial personal computer is present in each goat, and industrial personal computer includes mainly that goat temperature passes
Sensor, goat humidity sensor, goat vibrating sensor, goat image process controller, network-extension.Wherein image
Processing controller includes power supply unit, video storage unit, image analyzing unit, image acquisition units.Image acquisition units packet
Include LED light, camera, notch of switch machine position.
Current each Products can be realized:It obtains notch of switch machine value 1. train is crossed under vehicle or equipment switching state and reports
It is alert.2. obtaining the real-time parameters such as temperature, humidity, vibration.3. can arrange, display, store to related data.But current market production
There are the deficiencies on some an open questions and design for product.Chinese invention patent application, publication number CN102673611A are public
It has opened a kind of goat and has indicated rod notch video frequency monitoring method and system.This system is compared, the design method uses more
The design of more feasibilities:(1) camera of the design method in industrial personal computer uses white light camera, is taken the photograph compared to traditional infrared
As head has the clear superiorities such as reduction object real image, light regulating technology, longer life expectancy.(2) the design method uses host computer
Processor-server have high performance computation ability, compared to industrial computer A RM processors processing capacity improve hundred times.(3)
Network bandwidth is larger to the limitation of uploaded videos network speed, and the design waits for the mode of idle period uploaded videos using being first locally stored
It avoids train and spends the special occasions such as vehicle, video image more can guarantee using the pattern uploaded after video compress compared to conventional products
Really.(4) in the gap analysis to image, each video frame images gap analysis of the design method to complete video can
Notch caused by possible to train instant jitter, which goes beyond the scope, to be monitored.Conventional monitor mode is before notch of switch machine variation
Sectional drawing is distinguished to video when stablizing constant with notch, stablizes constant notch value by finally obtaining and is to weigh notch of switch machine
It is no to go beyond the scope.There are larger loopholes for the method, go beyond the scope when there is shake in notch variation, and in normal model after stablizing
It encloses, the breach scope in None- identified variation.The basic reason of loophole be the device space in industrial personal computer, cost, power consumption and
The factors such as network bandwidth limit algorithm process ability of the industrial personal computer to Online Video.(5) anti-in the design method is failed to report logical
It crosses and presets image classification data library, local feature point analysis is carried out to image to be detected.The design is in image algorithm simultaneously
It is upper to use machine learning to great amount of images resource classification, optimized using convolutional neural networks, passes through Fast R-CNN (Fast
Regions Convolutional Neural Networks) carry out classification detection.Its effect is better than SOBEL (discrete single orders
Difference operator) boundary operator.
Invention content
The technical problem to be solved in the present invention is to provide a kind of anti-design methods failed to report of notch of switch machine monitoring, improve and turn
The reliability of rut machine notch monitoring.
Technical scheme of the present invention:
A kind of anti-design method failed to report of notch of switch machine monitoring, including industrial personal computer processing module, data receiver and storage
Module, machine learning identification module.Overall flow is as shown in Figure 2.
Start, initialize communication protocol, time timer is arranged in industrial personal computer processing module, and industry control is acquired within the setting time
The routine datas such as machine real time temperature, humidity, vibration simultaneously combine upload.Meanwhile whether industrial personal computer moment monitoring train is crossed vehicle or is turned
Whether rut machine works.Under crossing vehicle or goat working condition in train, notch inspection is carried out to image using SOBEL boundary operators
It surveys, and obtains the gap datas such as notch of switch machine value, location information, antiposition information and type of alarm.Judge that train is crossed vehicle or turned
Whether rut machine operating status stops, and combines upload gap data if stopping, and video is stored in local.Wait for network idle
When uploaded videos.
Then, data receiver and memory module receive the data that industrial personal computer uploads, the video resource that will be uploaded from industrial personal computer
It is stored in NVR servers.This module receives the data packet that industrial personal computer uploads routine data, calculates and shows in host computer.Work as mould
When block receives the data packet of gap data, to gap data resolve packet, gap data and type of alarm are extracted respectively, is protected
Gap data and type of alarm are deposited in NVR servers, while host computer shows gap data and type of alarm, by type of alarm with
The type of alarm exported after machine learning identification module analysis video compares.
Then, NVR is stored in advance in using the image local feature database of multiple target in machine learning identification module to take
It is engaged on device, converts video to be measured to video frame picture, be separately added into the sample set in image classification data library, wherein depositing in advance
The feature samples of storage contain the characteristic of division such as notch crosses the border, oil droplet pollutes, camera visual angle is crooked.It establishes inartificial
The function of server automatic identification in the case of manual identification characteristics of image, wherein with a shared convolutional neural networks with before
Feedback neural fusion target's feature-extraction and target signature detect two parts content.Image classification data library is established, by video
Image classification data library sample set is added in frame picture.Convolutional neural networks are established, process of convolution is carried out to relevant parameter.Using
Regularization obtains optimal fitting neural network to convolutional neural networks optimization processing.W sampling is carried out to image classification data library
Processing, is obtained W autonomous learning element, is combined to W unit in the way of ballot.It is calculated by Fast R-CNN
Method carries out classification detection, exports type of alarm.Type of alarm and industrial personal computer type of alarm are compared, may occur in which following four
Possible situation:(1) if being judged to alarming at this time, and industrial personal computer is also alarm, then is considered as Normal Alarm.(2) if judging at this time
To alarm, but industrial personal computer is normal, then is considered as and fails to report police.(3) if being determined as at this time normally, and industrial personal computer is alarm, then is considered as
False alarm.(4) if being judged to normally, being considered as normal at this time.Last host computer is calculated and be shown as a result, unicast is back to work
Control machine is then returned to relevant control instruction.The computing capability that the present invention relies on background server powerful is realized to being lacked in variation
Each frame image detection of mouth video, can solve the alarm appeared above, false alarm, fail to report police.When there is alert, system
Host computer operator on duty can be prompted.
The machine learning identification module realization is as follows:
Step 1:Establish image classification data library
Detection video is converted to video frame picture first, establishes the image classification data library of an image;By video frame
Picture, notch cross the border, oil droplet pollutes, the crooked N number of characteristic of division in camera visual angle extracts L sample respectively, and from feature samples
In respectively extract S test set, if remainder set M=L-S is training set, wherein M in each sample<L, S<<L, using independently adopting
Sample method is split training set M, and verifies and collect respectively as sample;Next the compression of images concentrated to sample verification converts
At the image of 64*64 sizes, the mean value of each pixels of M is found out, the image that obtains that treated;
Step 2:Convolutional neural networks are built, deconvolution parameter optimization is carried out, obtain the convolutional neural networks after optimization;
There are two types of convolutional neural networks structures:The first be convolutional layer, convolutional layer, Relu layers, Relu layers, it is sample level, complete
Articulamentum and softmax layers;Be for second convolutional layer, Relu layers, sample level, full articulamentum and softmax layers;Two kinds of convolution god
3 kinds of convolutional neural networks frames are corresponded to through network structure, convolutional neural networks frame in 3 kinds will be inputted in training set, and allow three
Kind convolutional neural networks frame carries out successive ignition operation respectively, and the accurate journey of verification collection image recognition is compared according to training set
Degree, then take accuracy rate it is high as an optimization after convolutional neural networks frame;
Step 3:To a regular coefficient is added in each convolutional layer of convolutional neural networks after optimization and full articulamentum, and
Over-fitting can be effectively reduced plus a random inactivation coefficient in each full articulamentum, wherein random inactivation is next time
It can be by the network structure of unlatching neuron again;And training set M is divided into small samples repetitive exercise;
Step 4:Step 1 image classification data library is subjected to W sampling, obtains W learner Ai(i=1,2...W);
Each learner carries out prediction ballot with identical ballot probability to test set S, obtains the most classification of poll as final prediction
As a result it is exported,
WhereinIt is the N kind samples correspondence prediction value set of taxonomy database,To scheme after processing
Prediction output on the direction vector of picture;Occur multiple test images in test set S and obtain high ticket, then therefrom randomly selects one
It is a;
Step 5:Will in step 2 handle after image convolutional layer and Relu layer, obtain one with map characteristic pattern,
For multiple rectangle frames are arranged in mappings characteristics figure, and marked one by one using binary label;Rectangle frame is demarcated into region and front
The content of step identification is overlap proportion Z, and foreground overlap proportion Z is arranged according to actual parameter1, patterning ratio Z2;It takes and is more than
Z1Be set as foreground sample, then obtain prediction block;It takes and is less than Z2Rectangle frame is labeled as background sample, remaining frame is given up.
The beneficial effects of the present invention are being crossed under vehicle or goat working condition when train, industrial personal computer reports alert, and will
Video is stored to local hard drive, upload server when network idle, and alarm video frame, meter are exported by machine learning algorithm
Notch value is calculated, alert is exported.Type of alarm and the comparison of industrial personal computer alert are judged whether to fail to report, it is hidden to eliminate railway website
Suffer from.The local Security Personnel of calling in time overhauls on the spot.
Description of the drawings
Fig. 1 is the composition frame chart of the notch of switch machine monitoring system of the present invention.
Fig. 2, which is that the bright notch of switch machine monitoring of this hair is anti-, fails to report design overall flow figure.
Fig. 3 is the industrial personal computer processing module flow chart of the present invention.
Fig. 4 is the data receiver and memory module flow chart of the present invention.
Fig. 5 is the machine learning identification module block diagram of the present invention.
Fig. 6 is the video detection block diagram of the present invention.
Specific implementation mode
Below in conjunction with invention content and the Figure of description specific implementation mode that the present invention will be described in detail.
The present invention includes industrial personal computer processing module, data receiver and memory module, machine learning identification module.At industrial personal computer
Reason module is responsible for acquiring goat conventional parameter, video, gap parameters, alarm types, be respectively combined after analysis be uploaded to it is upper
Machine.The data and video that industrial personal computer uploads are stored in NVR servers by data receiver and memory module, and machine learning identifies mould
Block output alarm video frame calculates notch value, output type of alarm, the type of alarm for later uploading type of alarm and industrial personal computer
It compares and exports result.
(1) industrial personal computer processing module
This module acquires the routine datas such as real time temperature, humidity, vibration in industrial personal computer;Also train crosses vehicle or goat work
When making, notch of switch machine video, gap data, type of alarm gap data, and above-mentioned data are respectively combined and are uploaded to
Position machine.As shown in Figure 3.First, communication protocol is initialized, multicast protocol is added in host computer and each industrial personal computer, using multicast and list
Broadcast the mode being combined.Industrial personal computer uses unicast return information.Then, judge whether cycle timer is opened, if it is not, then returning
It returns;If it is, acquire the routine datas such as industrial personal computer real time temperature, humidity, vibration using 1s as the period, and by above-mentioned data single broadcasting
It is back to host computer, and in server backup.Meanwhile industrial personal computer persistently monitors train or goat state, if state is constant,
Then continue to monitor, and constantly uploads real-time routine data.If state changes, white light camera is opened, and is calculated using the edges SOBEL
Son carries out notch detection to image;Judge whether current video records completion, if it is, the machine video is stored in hard disk, such as
Fruit is no, then the combination unicast of the gap datas such as notch of switch machine value, location information, antiposition information is uploaded to host computer, and taking
Business device backup.If train still crosses vehicle or goat state persistently changes, return;If state stops, network is monitored
State, if network idle, uploaded videos later.
(2) data receiver and memory module
The data packet that industrial personal computer uploads in data receiver and memory module receiving station, industrial personal computer video resource is stored in
NVR servers.Wherein host computer receives the routine data data packet that the fixed cycle sends from every industry control machine equipment, simultaneously
The notch of switch machine of also the gap data data packet and idle period of notch of switch machine data and type of alarm combination regards
Frequently, as shown in Figure 4.First, communication protocol is initialized, host computer establishes UDP cast communications with industrial personal computer, receives each industrial personal computer
Reported data.Then judge data type, if it is video data, then video is stored in NVR servers.If it is normal
Data are advised, after often receiving a bag data, judge whether data are effective by start bit, stop bits and exclusive or check value.If
Invalid then directly discarding, continues to receive data from industrial personal computer;Data packet is further parsed at this time, if data effectively if extract number
According to then judging data type, if it is routine data data packet, then preserve routine data in NVR servers, simultaneously
Host computer shows routine data.Then, to gap data resolve packet, gap data and type of alarm are extracted respectively, are preserved
Gap data and type of alarm are in NVR servers, while host computer shows gap data and type of alarm, by type of alarm and machine
The type of alarm exported after device study identification module analysis video compares.
(3) machine learning identification module
This module is used to handle video when train crosses vehicle or goat work.Video is resident locally by industrial personal computer, waits for net
The network free time is uploaded to NVR servers, and then this module uses machine learning method to the further image procossing of the video of upload.
As shown in figure 4, to content recognition basic procedure.(1) image classification data library is established.(2) convolutional Neural net is established
Network.(3) regularization optimization neural network is used.(4) W sampling processing is carried out to image classification data library, obtains W autonomous
Unit is practised, W unit is combined in the way of ballot.(5) by Fast R-CNN to the unit of extraction into
Row classification detects, and classification results are exported and assessed.It is as follows:
Step 1:Establish image classification data library.Detection video is converted to video frame picture first herein, establishes one
The image classification data library of a image.Video frame picture, notch are crossed the border, N number of classification such as oil droplet pollutes, camera visual angle is crooked
Feature extracts L sample respectively, and S test set is respectively extracted from feature samples, if remainder set M=L-S in each sample
For training set, wherein M<L, S<<L is split training set M using autonomous sampling method, and verifies and collect respectively as sample.It connects
The image for getting off to be transformed into the compression of images that verification is concentrated 64*64 sizes, finds out the mean value of each pixels of M, after obtaining processing
Image.
Step 2:Convolutional neural networks are built, deconvolution parameter optimization is carried out, obtain the convolutional neural networks after optimization.Volume
There are two types of product neural network structures.The first is convolutional layer, convolutional layer, Relu layers, Relu layers, sample level, full articulamentum,
Softmax layers;It is for second convolutional layer, Relu layers, sample level, full articulamentum, softmax layers.This two kinds of convolutional neural networks
Structure corresponds to 3 kinds of convolutional neural networks frames, convolutional neural networks frame in 3 kinds will be inputted in training set, and allow three kinds of frames
30 wheel iterative operations are carried out respectively, and the order of accuarcy of verification collection image recognition is compared according to training set, then takes accuracy rate high
Convolutional neural networks frame after as an optimization.
Step 3:To P is added in each convolutional layer of convolutional neural networks after optimization and full articulamentum2=0.05 canonical system
Number, and be that the over-fitting that can effectively reduce that K=0.5 is inactivated at random shows plus a coefficient after each full articulamentum
As wherein random inactivation can be by the network structure of unlatching neuron again for next time.And by training set M be divided into one every group it is each
100 small-sized training sample iteration 100 times, setting step-length 5e-3。
Step 4:Step 1 image classification data library is subjected to W sampling, obtains W learner Ai(i=1,2...W);
Each learner carries out prediction ballot with identical ballot probability to test set S, obtains the most classification of poll as final prediction
As a result it is exported,
WhereinIt is the N kind samples correspondence prediction value set of taxonomy database,To scheme after processing
Prediction output on the direction vector of picture;Occur multiple test images in test set S and obtain high ticket, then therefrom randomly selects one
It is a;
Step 5:The convolutional layer of image and Relu layer after being handled in step 2, one mappings characteristics figure of acquisition, selection reflects
Penetrate rectangle frame in characteristic pattern.It, will be each in nine kinds of rectangle frames to each one binary label of rectangle frame in nine kinds of rectangle frames
The candidate region of rectangle frame calibration is labeled as with the maximum rectangle frame of picture material overlap proportion by the identification of first four step
Foreground sample;Then, the surplus rectangle frame candidate region to removing labeled as foreground sample is analyzed again, will pass through first four step
Rectangle frame of the picture material overlap proportion of middle identification more than 0.7 is also denoted as foreground sample.Then prediction block is obtained, by prediction block
It is demarcated as the corresponding rectangle frame of foreground sample.By the candidate region of the rectangle frame calibration except above-mentioned label foreground sample rectangle frame
It is denoted as background sample with rectangle frame of the picture material overlap proportion less than 0.3 identified in first four step.Remaining background frame is given up
It abandons.Then,
A) image ingress area generates network model after the processing for obtaining step 1, to the Zone-network model of generation into
80000 repetitive exercise optimization of row, preserves the prediction block that target detection network model generates;
B) the prediction block information obtained by image after processing that step 1 obtains and a) imports target detection network model, right
Target detection network model carries out the training optimization of 40000 iteration, and preserves target detection network model parameter;
C) the target detection network model parameter obtained in b) is imported into the Zone-network model generated, to the region of generation
Network model carries out 80000 iteration, preserves the prediction block that target detection network model generates;
D) target detection network model parameter b) obtained and the prediction block information (c) obtained are imported into target detection net
Network model carries out 40000 iteration to target detection network model, preserves the prediction that target detection network model finally generates
Frame.
The included video frame extraction functions of opencv are wherein used when being extracted to video characteristic values.Such as Fig. 5, acquisition waits for
Video frame is added taxonomy database image pattern model, the instruction of above-mentioned steps is carried out to sample set by the video frame for detecting video
Practice final output video frame, to different types of sample set preset threshold value, if current video frame is more than threshold value, directly
It is connected on image and draws rectangle frame and output video frame;If being less than threshold value, marked by the rectangle frame with binary label
Characteristic value, then further output video frame preservation.Video frame is handled, type of alarm is exported.Last host computer is by work
The type of alarm that control machine prestores is compared with this type of alarm, is exported and is shown final result.
Claims (2)
1. a kind of anti-design method failed to report of notch of switch machine monitoring, passes through industrial personal computer processing module, data receiver and storage mould
Block, machine learning identification module are realized, which is characterized in that
Start, initialize communication protocol, time timer is arranged in industrial personal computer processing module, and acquisition industrial personal computer is real within the setting time
Shi Wendu, humidity, vibration data simultaneously combine upload;Meanwhile industrial personal computer moment monitoring train whether cross vehicle or goat whether work
Make;Under crossing vehicle or goat working condition in train, notch detection is carried out to image using SOBEL boundary operators, and obtain
The gap data of notch of switch machine value, location information, antiposition information and type of alarm;Judge that train crosses vehicle or goat operation shape
Whether state stops, and combines upload gap data if stopping, and video is stored in local;
Then, data receiver and memory module receive the data that industrial personal computer uploads, and the video resource uploaded from industrial personal computer is preserved
In NVR servers;This module receives the data packet that industrial personal computer uploads routine data, calculates and shows in host computer;When module connects
When receiving the data packet of gap data, to gap data resolve packet, gap data and type of alarm are extracted respectively, preserves and lacks
Mouth data and type of alarm are in NVR servers, while host computer shows gap data and type of alarm, by type of alarm and machine
The type of alarm exported after study identification module analysis video compares;
Then, NVR servers are stored in advance in using the image local feature database of multiple target in machine learning identification module
On, it converts video to be measured to video frame picture, is separately added into the sample set in image classification data library;Wherein, it prestores
Feature samples cross the border comprising notch, oil droplet pollution, the crooked characteristic of division in camera visual angle;It establishes in inartificial manual identification
The function of server automatic identification in the case of characteristics of image, it is real with a shared convolutional neural networks and feedforward neural network
Existing target's feature-extraction and target signature detection;Image classification data library is established, image classification data is added in video frame picture
Library sample set;Convolutional neural networks are established, process of convolution is carried out;Using regularization to convolutional neural networks optimization processing, obtain
Go out optimal fitting neural network;W sampling processing is carried out to image classification data library, W autonomous learning element is obtained, utilizes throwing
Ticket mode is combined W unit;Classification detection is carried out by Fast R-CNN algorithms, exports type of alarm;It will report
Alert type and industrial personal computer type of alarm are compared, and following four occur may situation:(1) if being judged to alarming at this time, and work
Control machine is also alarm, then is considered as Normal Alarm;(2) if being judged to alarming at this time, but industrial personal computer is normal, then is considered as and fails to report
It is alert;(3) if being determined as at this time normally, and industrial personal computer is alarm, then is considered as false alarm;(4) if be determined as at this time it is normal, depending on
It is normal;Last host computer is calculated and be shown as a result, unicast is back to industrial personal computer, is then returned to relevant control instruction.
2. the anti-design method failed to report of notch of switch machine monitoring according to claim 1, which is characterized in that the machine
Study identification module realization is as follows:
Step 1:Establish image classification data library
Detection video is converted to video frame picture first, establishes the image classification data library of an image;By video frame picture,
Notch crosses the border, oil droplet pollutes, the crooked N number of characteristic of division in camera visual angle extracts L sample, and respectively carried from feature samples respectively
S test set is taken, if remainder set M=L-S is training set, wherein M in each sample<L, S<<L, using autonomous sampling method pair
Training set M is split, and is verified and collected respectively as sample;Next the compression of images concentrated to sample verification is transformed into 64*
The image of 64 sizes finds out the mean value of each pixels of M, the image that obtains that treated;
Step 2:Convolutional neural networks are built, deconvolution parameter optimization is carried out, obtain the convolutional neural networks after optimization;
There are two types of convolutional neural networks structures:The first is convolutional layer, convolutional layer, Relu layers, Relu layers, sample level, full connection
Layer and softmax layers;Be for second convolutional layer, Relu layers, sample level, full articulamentum and softmax layers;Two kinds of convolutional Neural nets
Network structure corresponds to 3 kinds of convolutional neural networks frames, convolutional neural networks frame in 3 kinds will be inputted in training set, and allow three kinds of volumes
Product neural network framework carries out successive ignition operation respectively, the order of accuarcy of verification collection image recognition is compared according to training set, so
Take afterwards accuracy rate it is high as an optimization after convolutional neural networks frame;
Step 3:To a regular coefficient is added in each convolutional layer of convolutional neural networks after optimization and full articulamentum, and every
A full articulamentum can effectively reduce over-fitting plus a random inactivation coefficient, wherein random inactivation is that next time can quilt
Again the network structure of unlatching neuron;And training set M is divided into small samples repetitive exercise;
Step 4:Step 1 image classification data library is subjected to W sampling, obtains W learner Ai(i=1,2...W);Each
Learner carries out prediction ballot with identical ballot probability to test set S, obtains the most classification of poll as final prediction result
It is exported,
WhereinIt is the N kind samples correspondence prediction value set of taxonomy database,For image after processing
Prediction output on direction vector;Occur multiple test images in test set S and obtain high ticket, then therefrom randomly selects one;
Step 5:Will in step 2 handle after image convolutional layer and Relu layer, obtain one with map characteristic pattern, to reflect
It penetrates in characteristic pattern and multiple rectangle frames is set, and marked one by one using binary label;Rectangle frame is demarcated into region and preceding step
The content of identification is overlap proportion Z, and foreground overlap proportion Z is arranged according to actual parameter1, patterning ratio Z2;It takes and is more than Z1's
It is set as foreground sample, then obtains prediction block;It takes and is less than Z2Rectangle frame is labeled as background sample, remaining frame is given up.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810338172.8A CN108573283A (en) | 2018-04-12 | 2018-04-12 | A kind of anti-design method failed to report of notch of switch machine monitoring |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810338172.8A CN108573283A (en) | 2018-04-12 | 2018-04-12 | A kind of anti-design method failed to report of notch of switch machine monitoring |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108573283A true CN108573283A (en) | 2018-09-25 |
Family
ID=63574929
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810338172.8A Pending CN108573283A (en) | 2018-04-12 | 2018-04-12 | A kind of anti-design method failed to report of notch of switch machine monitoring |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108573283A (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109583295A (en) * | 2018-10-19 | 2019-04-05 | 河南辉煌科技股份有限公司 | A kind of notch of switch machine automatic testing method based on convolutional neural networks |
CN109887343A (en) * | 2019-04-04 | 2019-06-14 | 中国民航科学技术研究院 | It takes to a kind of flight and ensures node automatic collection monitoring system and method |
CN110310255A (en) * | 2019-05-24 | 2019-10-08 | 同济大学 | Notch of switch machine detection method based on target detection and image procossing |
CN110363742A (en) * | 2019-04-19 | 2019-10-22 | 上海铁大电信科技股份有限公司 | A kind of notch of switch machine detection method based on CNN and image procossing |
CN111127496A (en) * | 2019-12-10 | 2020-05-08 | 电子科技大学 | Turnout switch machine notch edge positioning method based on two-dimensional histogram and adaptive classification |
CN111950494A (en) * | 2020-08-20 | 2020-11-17 | 宁波广播电视集团 | Monitoring method, device and system based on convolutional neural network and readable storage medium |
CN113177614A (en) * | 2021-05-27 | 2021-07-27 | 广州白云电器设备股份有限公司 | Image recognition system and method for power supply switch cabinet of urban rail transit |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US6970183B1 (en) * | 2000-06-14 | 2005-11-29 | E-Watch, Inc. | Multimedia surveillance and monitoring system including network configuration |
CN201597606U (en) * | 2009-12-21 | 2010-10-06 | 杭州慧景科技有限公司 | Switch machine gap monitoring device |
CN102673611A (en) * | 2011-12-19 | 2012-09-19 | 济南铁路天龙高新技术开发有限公司 | Video monitoring method and video monitoring system of indication rod gaps of railway point switch |
CN104228881A (en) * | 2014-09-09 | 2014-12-24 | 宁波思高信通科技有限公司 | Switch machine gap representation video monitoring system and method |
CN105120217A (en) * | 2015-08-21 | 2015-12-02 | 上海小蚁科技有限公司 | Intelligent camera motion detection alarm system and method based on big data analysis and user feedback |
CN107437099A (en) * | 2017-08-03 | 2017-12-05 | 哈尔滨工业大学 | A kind of specific dress ornament image recognition and detection method based on machine learning |
-
2018
- 2018-04-12 CN CN201810338172.8A patent/CN108573283A/en active Pending
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US6970183B1 (en) * | 2000-06-14 | 2005-11-29 | E-Watch, Inc. | Multimedia surveillance and monitoring system including network configuration |
CN201597606U (en) * | 2009-12-21 | 2010-10-06 | 杭州慧景科技有限公司 | Switch machine gap monitoring device |
CN102673611A (en) * | 2011-12-19 | 2012-09-19 | 济南铁路天龙高新技术开发有限公司 | Video monitoring method and video monitoring system of indication rod gaps of railway point switch |
CN104228881A (en) * | 2014-09-09 | 2014-12-24 | 宁波思高信通科技有限公司 | Switch machine gap representation video monitoring system and method |
CN105120217A (en) * | 2015-08-21 | 2015-12-02 | 上海小蚁科技有限公司 | Intelligent camera motion detection alarm system and method based on big data analysis and user feedback |
CN107437099A (en) * | 2017-08-03 | 2017-12-05 | 哈尔滨工业大学 | A kind of specific dress ornament image recognition and detection method based on machine learning |
Non-Patent Citations (1)
Title |
---|
张富春: "转辙机表示缺口视频监测系统", 《铁道通信信号》 * |
Cited By (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109583295A (en) * | 2018-10-19 | 2019-04-05 | 河南辉煌科技股份有限公司 | A kind of notch of switch machine automatic testing method based on convolutional neural networks |
CN109583295B (en) * | 2018-10-19 | 2022-12-06 | 河南辉煌科技股份有限公司 | Automatic detection method for switch machine notch based on convolutional neural network |
CN109887343A (en) * | 2019-04-04 | 2019-06-14 | 中国民航科学技术研究院 | It takes to a kind of flight and ensures node automatic collection monitoring system and method |
CN110363742A (en) * | 2019-04-19 | 2019-10-22 | 上海铁大电信科技股份有限公司 | A kind of notch of switch machine detection method based on CNN and image procossing |
CN110310255A (en) * | 2019-05-24 | 2019-10-08 | 同济大学 | Notch of switch machine detection method based on target detection and image procossing |
CN110310255B (en) * | 2019-05-24 | 2023-01-10 | 同济大学 | Point switch notch detection method based on target detection and image processing |
CN111127496A (en) * | 2019-12-10 | 2020-05-08 | 电子科技大学 | Turnout switch machine notch edge positioning method based on two-dimensional histogram and adaptive classification |
CN111127496B (en) * | 2019-12-10 | 2022-07-29 | 电子科技大学 | Turnout switch machine notch edge positioning method based on two-dimensional histogram and adaptive classification |
CN111950494A (en) * | 2020-08-20 | 2020-11-17 | 宁波广播电视集团 | Monitoring method, device and system based on convolutional neural network and readable storage medium |
CN111950494B (en) * | 2020-08-20 | 2024-01-16 | 宁波广播电视集团 | Monitoring method, device and system based on convolutional neural network and readable storage medium |
CN113177614A (en) * | 2021-05-27 | 2021-07-27 | 广州白云电器设备股份有限公司 | Image recognition system and method for power supply switch cabinet of urban rail transit |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108573283A (en) | A kind of anti-design method failed to report of notch of switch machine monitoring | |
CN105791774A (en) | Surveillance video transmission method based on video content analysis | |
CN103096121B (en) | A kind of camera movement detection method and device | |
CN106331636A (en) | Intelligent video monitoring system and method of oil pipelines based on behavioral event triggering | |
CN107071344A (en) | A kind of large-scale distributed monitor video data processing method and device | |
CN111325048B (en) | Personnel gathering detection method and device | |
CN107194396A (en) | Method for early warning is recognized based on the specific architecture against regulations in land resources video monitoring system | |
CN110808945A (en) | Network intrusion detection method in small sample scene based on meta-learning | |
CN104966304A (en) | Kalman filtering and nonparametric background model-based multi-target detection tracking method | |
CN108182416A (en) | A kind of Human bodys' response method, system and device under monitoring unmanned scene | |
CN110096945B (en) | Indoor monitoring video key frame real-time extraction method based on machine learning | |
CN102984503A (en) | Network video storage server system | |
CN115116004A (en) | Office area abnormal behavior detection system and method based on deep learning | |
CN111353436A (en) | Super store operation analysis method and device based on image deep learning algorithm | |
CN211184122U (en) | Intelligent video analysis system for linkage of railway operation safety prevention and control and large passenger flow early warning | |
CN112883755A (en) | Smoking and calling detection method based on deep learning and behavior prior | |
CN117440018A (en) | Network camera equipment fortune pipe platform based on thing networking | |
CN108491759A (en) | A kind of process detection device and its process detection method based on deep learning | |
CN112181331A (en) | Consistency detection system for published contents of variable information publishing terminal | |
CN116761049A (en) | Household intelligent security monitoring method and system | |
CN116419059A (en) | Automatic monitoring method, device, equipment and medium based on behavior label | |
CN115309871A (en) | Industrial big data processing method and system based on artificial intelligence algorithm | |
CN115439933A (en) | Garbage classification release site detection method based on multiple model recognition strategies | |
CN114037937A (en) | Real-time refrigerator food material identification method based on multi-target tracking | |
CN109543505B (en) | Target detection system and method based on video image |
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 | ||
WD01 | Invention patent application deemed withdrawn after publication | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20180925 |