CN110032977A - A kind of safety warning management system based on deep learning image fire identification - Google Patents
A kind of safety warning management system based on deep learning image fire identification Download PDFInfo
- Publication number
- CN110032977A CN110032977A CN201910312300.6A CN201910312300A CN110032977A CN 110032977 A CN110032977 A CN 110032977A CN 201910312300 A CN201910312300 A CN 201910312300A CN 110032977 A CN110032977 A CN 110032977A
- Authority
- CN
- China
- Prior art keywords
- fire
- identification
- image
- fire identification
- module
- 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
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
- Alarm Systems (AREA)
Abstract
The invention discloses a kind of safety warning management systems based on deep learning image fire identification, from front to back successively include system management module, image pre-processing module, fire identification model module, forewarning management module and visualization model.Fire identification is the fire identification based on CNN (convolutional neural networks) and HSV algorithm;And/or it is based on the fire identification of faster-rcnn (target detection) algorithm;And/or, fire identification based on CNN (convolutional neural networks) and inter-frame difference algorithm, the present invention does image preprocessing by foreground detection techniques, it is completed based on deep learning modeling to fire detection, simultaneously, artificial model is established based on enterprise staff and practical experience, classifies to fire for endangering menace level, realizes grading forewarning system.When not only realizing high discrimination, the accuracy rate of fire, but also meeting multichannel high concurrent, the system Whole Response time is shorter.
Description
Technical field
The present invention relates to a kind of safety warning management systems more particularly to a kind of based on deep learning image fire identification
Safety warning management system.
Background technique
Existing fire detection method is broadly divided into two major classes: based on traditional images processing fire detection method and be based on
The fire detection method of depth network.Fire detection method based on traditional images processing is primarily intended to through analysis fire
The static natures such as visual signature such as color characteristic, textural characteristics, shape feature train fire hazard classification device;And it is based on depth network
Fire detection method the image in video frame is directly processed using convolutional neural networks model, save traditional study
Pretreated a large amount of previous works are done to training data in method, and due to convolutional neural networks model efficient " intelligence " and
" initiative " is compared with the traditional method the object detection method based on depth network, there is bigger advantage.
Fire detection method based on traditional images processing is mainly the visual characteristic such as color spy from detection fire image
The information such as sign, textural characteristics, shape feature are set out, and classifier is constructed.Colouring information is a key factor of fire detection, existing
The volume color model of some maturations has RGB and YCbCr color model, and researcher has done a large amount of research to fire color characteristic.
T.Celi, proposes a kind of adaptive video fire detection algorithm, which combines video foreground information and colouring information
To detect the class flame range domain in video.Adaptive background modeling is carried out first, then extracts the RGB color information of image, building
Statistical color model removes noise using the corrosion and expansive working of morphological method, finally uses Connected component labeling algorithm
Carry out fire detection.The real-time of this method is stronger, but causes algorithm accuracy rate not high due to being only extracted colouring information.With
Afterwards, Celik et al. proposition models fire distribution of color on YCbCr color space using multinomial.These color moulds
Type is by doing statistics credit from YCbCr color space to different classes of fire video sequence and fire picture concerned
What analysis obtained.The model increases to the accuracy of fire detection.But the behavioral characteristics of fire are not bound with, it is static to class fire
Target rate of false alarm is high.Lin proposes a kind of intelligent fire detection algorithm based on image procossing, which considers fire
Movement and static nature.Video motion region is obtained first;Secondly, by color it is complete between model obtain and wait upon favored area;Again,
Extract the interference that the shape features such as area, the perimeter of candidate region exclude some class fire objects;Finally, using irregular polygon
The shape features such as shape, circle carry out fire detection.This method has outstanding contributions on improving performance, reduction false detection rate.But away from
From remote or fire in image accounting hour, less effective.Kosmas proposes a real-time fire detection algorithm, using it is various when
Empty feature models fire-resistance behavior, and using dynamic texture analysis model to the temporal evolution of candidate image block image pixel intensities
It is modeled.The space-time characteristic that this method uses has: color probability information, blinking characteristics, wavelet analysis, space-time energy spectrometer,
Dynamic texture analysis etc. after the space time information feature for having extracted fire, carries out fire inspection using SVM (support vector machines) classifier
It surveys, improves the accuracy rate and robustness of algorithm.But time complexity is high, and delay is big.Duan Suolin is in order to overcome ambient lighting strong
The influence of degree proposes a kind of algorithm for effectively extracting fire characteristic, which uses vector correlation theory first, then sharp
Classified with feature of the neural network to extraction, this method achieves preferable classification knot under the complex situations of illumination variation
Fruit.But there is also ineffective under Small object discrimination is low and high concurrent.WangDC, CuiX, ParkE are proposed to be surveyed using random
The adaptive fire detection algorithm of examination and robust features, the algorithm construct YCbCr color space model first;Secondly, using
Approximate median method updates movement background image, and the two combines available some candidate video frames;Then, some fire are extracted
Feature, including flicker frequency, area, mass center etc. obtain one group of feature vector;Finally using random gloomy in machine learning algorithm
Woods algorithm carries out the building of fire hazard classification model, completes the fire detection in video.But this method nevertheless suffer from it is some external
The interference of condition, such as: resolution ratio, weather condition and the environmental factor of camera.Jin be directed to before based on motion information and
The frequent wrong report of the video fire hazard detection method of color model improves, and proposes a kind of logic-based and returns and random test
Real-time fire detection algorithm, experiment shows that the algorithm has remarkable effect in rate of false alarm, but could be improved to rate of failing to report.
Human brain nervous system is simulated by establishing profound model structure based on the fire detection method of depth network
Hierarchical signal treatment mechanism.Frizzi proposes a kind of fire hazard aerosol fog detection model based on convolutional neural networks.This method
It is no feature extraction phases with the maximum difference of other methods, directly obtains training mould by operating original rgb video frame
Type.In the training process, the time is handled required for detection fire hazard aerosol fog in order to reduce, used CNN (convolutional neural networks)
Photo structure in model simultaneously detects original image or reconstructed image using sliding window.These sliding windows pass through convolutional Neural
It network and is fully connected layer and obtains classification results.The shortcomings that algorithm is when image resolution ratio is high, and fire accounting hour effect is owed
It is good.Maksymiv proposes a kind of fire detection method, and Adaboost and local binary patterns are used in combination first for this method
Area-of-interest is obtained, to reduce time complexity;Then fire detection is carried out to alleviate wrong report using convolutional neural networks
Problem.This method can achieve the accuracy rate up to 95.2% on solution of emergent event test problems.But Adaboost was trained
Cheng Jiwei is time-consuming, and exactly algorithm needs one of improved critical issue for this.WangZ, ZhangH are by convolutional neural networks and branch
It holds vector machine and combines carry out fire detection, innovative point is that replace convolutional neural networks last with support vector machines connects entirely
Connect layer and Softmax classification, that is to say, that this method extracts the further feature of fire image using convolutional neural networks, then
The feature extracted SVM (support vector machines) algorithm is constructed into disaggregated model, realizes fire detection.Experiment shows this method
Better than the method that CNN (convolutional neural networks) or SVM (support vector machines) progress fire detection are used alone.But it is unable to satisfy
Multichannel high concurrent real time handling requirement.
Summary of the invention
The technical problem to be solved in the present invention is to provide a kind of safe early warning pipes based on deep learning image fire identification
Reason system, the safety warning management system can not only realize 90% or more the high discrimination of fire, 90% or more accuracy rate, and
And when meeting multichannel high concurrent, the system Whole Response time requires in the system performance within 8s.
In order to solve the above-mentioned technical problem, the present invention is based on the safe early warning management systems of deep learning image fire identification
System successively includes system management module, image pre-processing module, fire identification model module, forewarning management module from front to back
And visualization model.
The function of system management module is
1) the various configuration datas needed by management end system configuration video identification service and Image Acquisition end, match confidence
Breath includes:
A. manufacturer, the model, number, state, access path of hard disk video recorder are configured by hard disk video recorder management module
Equal main informations;
B. manufacturer, the carry hard disk video recorder, number, sequence number, subregion of video camera are configured by video camera management module
The main informations such as position, delay time, interval frame number;
2) parameter information relied on when the operation of configuration image preprocessing;
3) user information and user role are safeguarded, is all the access authority of role's setup module and operation;
The function of image pre-processing module is
1) camera information configured;
2) camera head monitor real-time pictures are obtained;
3) real-time pictures are carried out with gradation conversion, Gaussian Blur, frame are poor, corrosion expands, foreground extraction;
4) fire identification model module is transmitted to by the library url by 3 video requency frame datas extracted;
The function of fire identification model module is
Identification model is established using the image method for recognizing fire disaster based on deep learning algorithm, by the video received
It is identified, recognition result is passed into forewarning management module;
The function of forewarning management module is
Forewarning management module establishes artificial model in conjunction with person works' experience, carries out for the result of fire identification module
Analysis, further reduces the generation of wrong report;
The function of visualization model is
1) information such as the camera information and queue address that configure;
2) dynamic alert information is obtained, obtains warning message, and pop up the video pictures of alarm in camera window,
Alert locations are marked in corresponding video pictures.
Image method for recognizing fire disaster based on deep learning algorithm is carried out by fire identification convolutional neural networks structure
's;
Fire identification convolutional neural networks structure includes:
First layer pond: the core step-length of maxpooling2x2 be 2,
Second layer convolution: 64 step-lengths of convolution kernel of 5x5 size be 2,
Second layer pond: the core step-length of maxpooling2x2 be 2,
Third layer be full articulamentum: 512 neural units, dropout 0.6,
4th layer is sub-network output layer: being exported as two nodes, i.e. two classifications have fire, without fire;
Two nodal values of output, which are mapped to probability space i.e., with Softmax activation primitive fire, the probability without fire;
Softmax function expression are as follows:
Other than the last layer, other each layer activation primitives are all made of ReLU function;
It is ReLU (x)=max (0, x) up to formula;
The initialization of weight W uses μ=0, and the Gauss of std=0.1 is truncated;
Learning rate when training pattern is 0.001.
Fire identification is
Fire identification based on CNN (convolutional neural networks) and HSV algorithm;And/or
Fire identification based on faster-rcnn (target detection) algorithm;And/or
Fire identification based on CNN (convolutional neural networks) and inter-frame difference algorithm.
The present invention is based on the safety warning management system of deep learning image fire identification have compared with prior art with
Lower beneficial effect.
1, the technical program successively includes system management module, image pre-processing module, fire due to using from front to back
Calamity identification model module, the technological means of forewarning management module and visualization model, so, the present invention passes through foreground detection techniques
Image preprocessing is done, is completed based on deep learning modeling to fire detection, meanwhile, people is established based on enterprise staff and practical experience
Work model classifies to fire for endangering menace level, realizes grading forewarning system.Not only realize high discrimination, the standard of fire
True rate, and when meeting multichannel high concurrent, the system Whole Response time is shorter.
2, the technical program due to being using the function of system management module
1) the various configuration datas needed by management end system configuration video identification service and Image Acquisition end, match confidence
Breath includes:
A. manufacturer, the model, number, state, access path of hard disk video recorder are configured by hard disk video recorder management module
Equal main informations;
B. manufacturer, the carry hard disk video recorder, number, sequence number, subregion of video camera are configured by video camera management module
The main informations such as position, delay time, interval frame number;
2) parameter information relied on when the operation of configuration image preprocessing;
3) user information and user role are safeguarded, is all the access authority of role's setup module and operation;
The function of image pre-processing module is
1) camera information configured;
2) camera head monitor real-time pictures are obtained;
3) real-time pictures are carried out with gradation conversion, Gaussian Blur, frame are poor, corrosion expands, foreground extraction;
4) fire identification model module is transmitted to by the library url by 3 video requency frame datas extracted;
The function of fire identification model module is
Identification model is established using the image method for recognizing fire disaster based on deep learning algorithm, by the video received
It is identified, recognition result is passed into forewarning management module;
The function of forewarning management module is
Forewarning management module establishes artificial model in conjunction with person works' experience, carries out for the result of fire identification module
Analysis, further reduces the generation of wrong report;
The function of visualization model is
1) information such as the camera information and queue address that configure;
2) dynamic alert information is obtained, obtains warning message, and pop up the video pictures of alarm in camera window,
Alert locations are marked in corresponding video pictures;
Image method for recognizing fire disaster based on deep learning algorithm is carried out by fire identification convolutional neural networks structure
's;
Fire identification convolutional neural networks structure includes:
First layer pond: the core step-length of maxpooling2x2 be 2,
Second layer convolution: 64 step-lengths of convolution kernel of 5x5 size be 2,
Second layer pond: the core step-length of maxpooling2x2 be 2,
Third layer be full articulamentum: 512 neural units, dropout 0.6,
4th layer is sub-network output layer: being exported as two nodes, i.e. two classifications have fire, without fire;
Two nodal values of output, which are mapped to probability space i.e., with Softmax activation primitive fire, the probability without fire;
Softmax function expression are as follows:
Other than the last layer, other each layer activation primitives are all made of ReLU function;
It is ReLU (x)=max (0, x) up to formula;
The initialization of weight W uses μ=0, and the Gauss of std=0.1 is truncated;
The technological means that learning rate when training pattern is 0.001, so, discrimination can up to 90% or more, accuracy rate
Up to 90% or more, when sufficiently meeting multichannel high concurrent, the system Whole Response time requires in the system performance within 8s.This is
Other methods can not reach simultaneously at present.
3, the technical program is the fire based on CNN (convolutional neural networks) and HSV algorithm due to using fire identification
Identification;And/or it is based on the fire identification of faster-rcnn (target detection) algorithm;And/or it is based on CNN (convolutional Neural net
Network) and inter-frame difference algorithm fire identification technological means, so, different algorithm can be selected to carry out fire according to the actual situation
Calamity identification.
Detailed description of the invention
With reference to the accompanying drawings and detailed description to the present invention is based on the safe early warnings of deep learning image fire identification
Management system is described in further detail.
Fig. 1 is that the present invention is based on the schematic diagrames of the safety warning management system of deep learning image fire identification.
Fig. 2 is that the present invention is based on fire identification convolution minds in the safety warning management system of deep learning image fire identification
Through schematic network structure.
Fig. 3 is that the present invention is based on inter-frame difference algorithms in the safety warning management system of deep learning image fire identification to show
It is intended to.
Specific embodiment
As shown in Figure 1, present embodiments provide for a kind of safe early warning management based on deep learning image fire identification
System successively includes system management module, image pre-processing module, fire identification model module, forewarning management mould from front to back
Block and visualization model.
Present embodiment successively includes system management module, image pre-processing module, fire due to using from front to back
The technological means of identification model module, forewarning management module and visualization model, so, the present invention is done by foreground detection techniques
Image preprocessing is completed based on deep learning modeling to fire detection, meanwhile, it is established based on enterprise staff and practical experience artificial
Model classifies to fire for endangering menace level, realizes grading forewarning system.Not only realize the high discrimination, accurate of fire
Rate, and when meeting multichannel high concurrent, the system Whole Response time is shorter.
As shown in Figure 1, the function of system management module is
1) the various configuration datas needed by management end system configuration video identification service and Image Acquisition end, match confidence
Breath includes:
A. manufacturer, the model, number, state, access path of hard disk video recorder are configured by hard disk video recorder management module
Equal main informations;
B. manufacturer, the carry hard disk video recorder, number, sequence number, subregion of video camera are configured by video camera management module
The main informations such as position, delay time, interval frame number;
2) parameter information relied on when the operation of configuration image preprocessing;
3) user information and user role are safeguarded, is all the access authority of role's setup module and operation;
The function of image pre-processing module is
1) camera information configured;
2) camera head monitor real-time pictures are obtained;
3) real-time pictures are carried out with gradation conversion, Gaussian Blur, frame are poor, corrosion expands, foreground extraction;
4) fire identification model module is transmitted to by the library url by 3 video requency frame datas extracted;
The function of fire identification model module is
Identification model is established using the image method for recognizing fire disaster based on deep learning algorithm, by the video received
It is identified, recognition result is passed into forewarning management module;
The function of forewarning management module is
Forewarning management module establishes artificial model in conjunction with person works' experience, carries out for the result of fire identification module
Analysis, further reduces the generation of wrong report;
The function of visualization model is
1) information such as the camera information and queue address that configure;
2) dynamic alert information is obtained, obtains warning message, and pop up the video pictures of alarm in camera window,
Alert locations are marked in corresponding video pictures.
Image method for recognizing fire disaster based on deep learning algorithm is carried out by fire identification convolutional neural networks structure
's;
As shown in Fig. 2, fire identification convolutional neural networks structure includes:
First layer pond: the core step-length of maxpooling2x2 be 2,
Second layer convolution: 64 step-lengths of convolution kernel of 5x5 size be 2,
Second layer pond: the core step-length of maxpooling2x2 be 2,
Third layer be full articulamentum: 512 neural units, dropout 0.6,
4th layer is sub-network output layer: being exported as two nodes, i.e. two classifications have fire, without fire;
Two nodal values of output, which are mapped to probability space i.e., with Softmax activation primitive fire, the probability without fire;
Softmax function expression are as follows:
Other than the last layer, other each layer activation primitives are all made of ReLU function;
It is ReLU (x)=max (0, x) up to formula;
The initialization of weight W uses μ=0, and the Gauss of std=0.1 is truncated;
Learning rate when training pattern is 0.001.
Present embodiment due to being using the function of system management module
1) the various configuration datas needed by management end system configuration video identification service and Image Acquisition end, match confidence
Breath includes:
A. manufacturer, the model, number, state, access path of hard disk video recorder are configured by hard disk video recorder management module
Equal main informations;
B. manufacturer, the carry hard disk video recorder, number, sequence number, subregion of video camera are configured by video camera management module
The main informations such as position, delay time, interval frame number;
2) parameter information relied on when the operation of configuration image preprocessing;
3) user information and user role are safeguarded, is all the access authority of role's setup module and operation;
The function of image pre-processing module is
1) camera information configured;
2) camera head monitor real-time pictures are obtained;
3) real-time pictures are carried out with gradation conversion, Gaussian Blur, frame are poor, corrosion expands, foreground extraction;
4) fire identification model module is transmitted to by the library url by 3 video requency frame datas extracted;
The function of fire identification model module is
Identification model is established using the image method for recognizing fire disaster based on deep learning algorithm, by the video received
It is identified, recognition result is passed into forewarning management module;
The function of forewarning management module is
Forewarning management module establishes artificial model in conjunction with person works' experience, carries out for the result of fire identification module
Analysis, further reduces the generation of wrong report;
The function of visualization model is
1) information such as the camera information and queue address that configure;
2) dynamic alert information is obtained, obtains warning message, and pop up the video pictures of alarm in camera window,
Alert locations are marked in corresponding video pictures;
Image method for recognizing fire disaster based on deep learning algorithm is carried out by fire identification convolutional neural networks structure
's;
Fire identification convolutional neural networks structure includes:
First layer pond: the core step-length of maxpooling2x2 be 2,
Second layer convolution: 64 step-lengths of convolution kernel of 5x5 size be 2,
Second layer pond: the core step-length of maxpooling2x2 be 2,
Third layer be full articulamentum: 512 neural units, dropout 0.6,
4th layer is sub-network output layer: being exported as two nodes, i.e. two classifications have fire, without fire;
Two nodal values of output, which are mapped to probability space i.e., with Softmax activation primitive fire, the probability without fire;
Softmax function expression are as follows:
Other than the last layer, other each layer activation primitives are all made of ReLU function;
It is ReLU (x)=max (0, x) up to formula;
The initialization of weight W uses μ=0, and the Gauss of std=0.1 is truncated;
The technological means that learning rate when training pattern is 0.001, so, discrimination can up to 90% or more, accuracy rate
Up to 90% or more, when sufficiently meeting multichannel high concurrent, the system Whole Response time requires in the system performance within 8s.This is
Other methods can not reach simultaneously at present.
The fire identification of present embodiment can be the fire identification based on CNN (convolutional neural networks) and HSV algorithm.
It is of course also possible to be the fire identification based on faster-rcnn (target detection) algorithm;
It can also be the fire identification based on CNN (convolutional neural networks) and inter-frame difference algorithm.
Fire identification based on CNN (convolutional neural networks) and HSV algorithm is as follows.
Product design, C-terminal, that is, client are carried out with C/S model, main task is acquisition image and carries out to image data
Pretreatment reads real-time monitoring images with rtsp transport protocol and obtains the location information in all class flame ranges domain with HSV algorithm,
And obtained small figure is transferred to the end S i.e. server-side with http with byte stream files format, using the small figure received as defeated
Enter and predicted with convolutional neural networks model, will in prediction result be that the location information of fire is echoed back to C-terminal, and bullet
Frame alarm.
This embodiment finds in testing, and recognition effect is very the case where to high fire or fire characteristic clearly
Good accuracy rate can reach 99% or more, but when recognition effect only has 70% left side to fiery positional distance camera farther out or when flare is small
The right side, and because HSV algorithm can occupy a large amount of cpu resources, it is delayed and reaches 7-10 seconds in the case where 10 road camera, add
The time loss of server-side identification, is more unable to satisfy the demand of timeliness in actual conditions.
Fire identification based on faster-rcnn (target detection) algorithm is as follows.
In view of HSV algorithm be unable to satisfy timeliness requirement, determine with depth learning technology to product technology framework into
Row improves, and C-terminal only acquires image and do not do any pretreatment, directly sends the end S for whole monitoring image and identify, the end S fortune
Task is identified and positioned with the completion of faster-rcnn (target detection) algorithm.
This embodiment is found in testing, although accelerating and reducing pretreated time loss solution by GPU
High concurrent of having determined imeliness problem, but prediction effect is not satisfactory.Reason is that faster-rcnn (target detection) algorithm is used for
When target detection, big target is not so good as the discrimination of Small object.And accounting of the fire in monitoring all will not be very big.Training in this way
Model out, discrimination can reach 90% or so, but rate of false alarm is very high.It is unable to satisfy the demand of confidence level in actual conditions.
Fire identification based on CNN (convolutional neural networks) and inter-frame difference algorithm is as follows.
It is found by second scheme test result analysis, main wrong report source is background parts in image, then first
Foreground extraction is made a decision again out will largely reduce wrong report, while consider timeliness, determine to use performance consumption
For relatively small number of frame difference method as Image Pretreatment Algorithm, frame difference method thought is to detected to change in adjacent two field pictures
Region.This method is to carry out difference with the two continuous frames image in image sequence, and then the binaryzation grey scale difference image comes
Extract motion information.The image divided by interframe Changing Area Detection distinguishes background area and moving region, Jin Erti
Take the target to be detected.It is passed through by comparing the difference of front and back two field pictures corresponding pixel points gray value in image sequence
Two frames subtract each other, the threshold value comparison with setting, if gray value is less than threshold value, it is believed that this without motion object passes through;Conversely,
Then think there is moving object.Kth frame and k+1 frame image fk (x, y), one two-value difference diagram of variation between fk+l (x, y)
As D (x, y) expression, such as formula:
The 0 corresponding unchanged place in front and back in binary map, the place of 1 corresponding variation.
It is as shown in Figure 3:
The characteristics of frame difference method is to realize simple, and arithmetic speed is fast, be for dynamic environment adaptivity it is very strong, to light
Variation be not very sensitive.
Then the end S is sent by foreground picture, with deep learning CNN (convolutional neural networks) model prediction.And in technology
Warning module is added in framework, which is based on personnel on site's experience and models, and realizes to fire grading forewarning system, further mentions
The confidence level of high product.
Present embodiment is that the fire based on CNN (convolutional neural networks) and HSV algorithm is known due to using fire identification
Not;And/or it is based on the fire identification of faster-rcnn (target detection) algorithm;And/or it is based on CNN (convolutional neural networks)
And the technological means of the fire identification of inter-frame difference algorithm, so, different algorithms can be selected to carry out fire according to the actual situation
Identification.
Claims (4)
1. a kind of safety warning management system based on deep learning image fire identification, it is characterised in that: from front to back successively
Including system management module, image pre-processing module, fire identification model module, forewarning management module and visualization model.
2. the safety warning management system according to claim 1 based on deep learning image fire identification, it is characterised in that:
The function of system management module is
1) the various configuration datas needed by management end system configuration video identification service and Image Acquisition end, configuration information packet
It includes:
A. it is main that manufacturer, model, number, state, access path of hard disk video recorder etc. is configured by hard disk video recorder management module
Want information;
B. by video camera management module configure the manufacturer of video camera, carry hard disk video recorder, number, sequence number, district location,
The main informations such as delay time, interval frame number;
2) parameter information relied on when the operation of configuration image preprocessing;
3) user information and user role are safeguarded, is all the access authority of role's setup module and operation;
The function of image pre-processing module is
1) camera information configured;
2) camera head monitor real-time pictures are obtained;
3) real-time pictures are carried out with gradation conversion, Gaussian Blur, frame are poor, corrosion expands, foreground extraction;
4) fire identification model module is transmitted to by the library url by 3 video requency frame datas extracted;
The function of fire identification model module is
Identification model is established using the image method for recognizing fire disaster based on deep learning algorithm, by carrying out to the video received
Identification, passes to forewarning management module for recognition result;
The function of forewarning management module is
Forewarning management module establishes artificial model in conjunction with person works' experience, is divided for the result of fire identification module
Analysis, further reduces the generation of wrong report;
The function of visualization model is
1) information such as the camera information and queue address that configure;
2) dynamic alert information is obtained, warning message is obtained, and pop up the video pictures of alarm in camera window, corresponding
Video pictures in mark alert locations.
3. the safety warning management system according to claim 2 based on deep learning image fire identification, it is characterised in that:
Image method for recognizing fire disaster based on deep learning algorithm is carried out by fire identification convolutional neural networks structure;
Fire identification convolutional neural networks structure includes:
First layer pond: the core step-length of maxpooling2x2 be 2,
Second layer convolution: 64 step-lengths of convolution kernel of 5x5 size be 2,
Second layer pond: the core step-length of maxpooling2x2 be 2,
Third layer be full articulamentum: 512 neural units, dropout 0.6,
4th layer is sub-network output layer: being exported as two nodes, i.e. two classifications have fire, without fire;
Two nodal values of output, which are mapped to probability space i.e., with Softmax activation primitive fire, the probability without fire;
Softmax function expression are as follows:
Other than the last layer, other each layer activation primitives are all made of ReLU function;
It is ReLU (x)=max (0, x) up to formula;
The initialization of weight W uses μ=0, and the Gauss of std=0.1 is truncated;
Learning rate when training pattern is 0.001.
4. stating the safety warning management system based on deep learning image fire identification according to claim 2, it is characterised in that:
Fire identification is
Fire identification based on CNN and HSV algorithm;And/or
Fire identification based on faster-rcnn algorithm;And/or
Fire identification based on CNN and inter-frame difference algorithm.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910312300.6A CN110032977A (en) | 2019-04-18 | 2019-04-18 | A kind of safety warning management system based on deep learning image fire identification |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910312300.6A CN110032977A (en) | 2019-04-18 | 2019-04-18 | A kind of safety warning management system based on deep learning image fire identification |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110032977A true CN110032977A (en) | 2019-07-19 |
Family
ID=67238964
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910312300.6A Pending CN110032977A (en) | 2019-04-18 | 2019-04-18 | A kind of safety warning management system based on deep learning image fire identification |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110032977A (en) |
Cited By (17)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110633675A (en) * | 2019-09-18 | 2019-12-31 | 东北大学 | System and method for identifying fire in video based on convolutional neural network |
CN110853287A (en) * | 2019-09-26 | 2020-02-28 | 华南师范大学 | Flame real-time monitoring system and method based on Internet of things distributed architecture |
CN110956611A (en) * | 2019-11-01 | 2020-04-03 | 武汉纺织大学 | Smoke detection method integrated with convolutional neural network |
CN110991243A (en) * | 2019-11-01 | 2020-04-10 | 武汉纺织大学 | Straw combustion identification method based on combination of color channel HSV and convolutional neural network |
CN111191575A (en) * | 2019-12-27 | 2020-05-22 | 国网江苏省电力有限公司电力科学研究院 | Naked flame detection method and system based on flame jumping modeling |
CN111274896A (en) * | 2020-01-15 | 2020-06-12 | 深圳市守行智能科技有限公司 | Smoke and fire identification algorithm |
CN111462451A (en) * | 2019-11-01 | 2020-07-28 | 武汉纺织大学 | Straw burning detection alarm system based on video information |
CN111639620A (en) * | 2020-06-08 | 2020-09-08 | 深圳航天智慧城市系统技术研究院有限公司 | Fire disaster analysis method and system based on visible light image recognition |
CN111784964A (en) * | 2020-07-03 | 2020-10-16 | 深圳爱克莱特科技股份有限公司 | Street lamp, street lamp system and street lamp fire alarm method |
CN111814725A (en) * | 2020-07-20 | 2020-10-23 | 北京华正明天信息技术股份有限公司 | Early warning method for judging ignition of monitoring video based on CNN + LSTM + MLP combined neural network |
CN111860646A (en) * | 2020-07-20 | 2020-10-30 | 北京华正明天信息技术股份有限公司 | Forest fire detection method based on neural network |
CN111882810A (en) * | 2020-07-31 | 2020-11-03 | 广州市微智联科技有限公司 | Fire identification and early warning method and system |
CN112297033A (en) * | 2020-11-03 | 2021-02-02 | 泰州市华仕达机械制造有限公司 | Automatic change robotic arm actuating system |
CN112418004A (en) * | 2020-11-05 | 2021-02-26 | 烟台艾睿光电科技有限公司 | Fire safety detection method and related components |
CN112907886A (en) * | 2021-02-07 | 2021-06-04 | 中国石油化工股份有限公司 | Refinery plant fire identification method based on convolutional neural network |
CN114299445A (en) * | 2021-12-11 | 2022-04-08 | 南方电网大数据服务有限公司 | Power distribution room danger source identification method and device, computer equipment, medium and product |
CN114604708A (en) * | 2022-05-13 | 2022-06-10 | 深圳正为格智能科技有限公司 | Elevator fire risk early warning method and system |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101873472A (en) * | 2009-04-22 | 2010-10-27 | 北京中星微电子有限公司 | Video monitoring system and video monitoring method |
CN103106766A (en) * | 2013-01-14 | 2013-05-15 | 广东赛能科技有限公司 | Forest fire identification method and forest fire identification system |
CN104441526A (en) * | 2014-09-24 | 2015-03-25 | 上海智觉光电科技有限公司 | Online mold monitoring and protecting system and method based on contour matching |
CN105976365A (en) * | 2016-04-28 | 2016-09-28 | 天津大学 | Nocturnal fire disaster video detection method |
CN109377713A (en) * | 2018-09-26 | 2019-02-22 | 石化盈科信息技术有限责任公司 | A kind of fire alarm method and system |
CN109587242A (en) * | 2018-12-05 | 2019-04-05 | 华润置地控股有限公司 | Platform of internet of things system and its cloud platform and local terminal |
-
2019
- 2019-04-18 CN CN201910312300.6A patent/CN110032977A/en active Pending
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101873472A (en) * | 2009-04-22 | 2010-10-27 | 北京中星微电子有限公司 | Video monitoring system and video monitoring method |
CN103106766A (en) * | 2013-01-14 | 2013-05-15 | 广东赛能科技有限公司 | Forest fire identification method and forest fire identification system |
CN104441526A (en) * | 2014-09-24 | 2015-03-25 | 上海智觉光电科技有限公司 | Online mold monitoring and protecting system and method based on contour matching |
CN105976365A (en) * | 2016-04-28 | 2016-09-28 | 天津大学 | Nocturnal fire disaster video detection method |
CN109377713A (en) * | 2018-09-26 | 2019-02-22 | 石化盈科信息技术有限责任公司 | A kind of fire alarm method and system |
CN109587242A (en) * | 2018-12-05 | 2019-04-05 | 华润置地控股有限公司 | Platform of internet of things system and its cloud platform and local terminal |
Non-Patent Citations (1)
Title |
---|
雷杨: "林火识别中不同环境识别算法的探讨", 《广东通信技术》 * |
Cited By (23)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110633675A (en) * | 2019-09-18 | 2019-12-31 | 东北大学 | System and method for identifying fire in video based on convolutional neural network |
CN110853287A (en) * | 2019-09-26 | 2020-02-28 | 华南师范大学 | Flame real-time monitoring system and method based on Internet of things distributed architecture |
CN111462451A (en) * | 2019-11-01 | 2020-07-28 | 武汉纺织大学 | Straw burning detection alarm system based on video information |
CN110991243A (en) * | 2019-11-01 | 2020-04-10 | 武汉纺织大学 | Straw combustion identification method based on combination of color channel HSV and convolutional neural network |
CN111462451B (en) * | 2019-11-01 | 2022-04-26 | 武汉纺织大学 | Straw burning detection alarm system based on video information |
CN110956611A (en) * | 2019-11-01 | 2020-04-03 | 武汉纺织大学 | Smoke detection method integrated with convolutional neural network |
CN111191575A (en) * | 2019-12-27 | 2020-05-22 | 国网江苏省电力有限公司电力科学研究院 | Naked flame detection method and system based on flame jumping modeling |
CN111191575B (en) * | 2019-12-27 | 2022-09-23 | 国网江苏省电力有限公司电力科学研究院 | Naked flame detection method and system based on flame jumping modeling |
CN111274896A (en) * | 2020-01-15 | 2020-06-12 | 深圳市守行智能科技有限公司 | Smoke and fire identification algorithm |
CN111274896B (en) * | 2020-01-15 | 2023-09-26 | 深圳市守行智能科技有限公司 | Smoke and fire recognition algorithm |
CN111639620A (en) * | 2020-06-08 | 2020-09-08 | 深圳航天智慧城市系统技术研究院有限公司 | Fire disaster analysis method and system based on visible light image recognition |
CN111639620B (en) * | 2020-06-08 | 2023-11-10 | 深圳航天智慧城市系统技术研究院有限公司 | Fire analysis method and system based on visible light image recognition |
CN111784964A (en) * | 2020-07-03 | 2020-10-16 | 深圳爱克莱特科技股份有限公司 | Street lamp, street lamp system and street lamp fire alarm method |
CN111860646A (en) * | 2020-07-20 | 2020-10-30 | 北京华正明天信息技术股份有限公司 | Forest fire detection method based on neural network |
CN111814725A (en) * | 2020-07-20 | 2020-10-23 | 北京华正明天信息技术股份有限公司 | Early warning method for judging ignition of monitoring video based on CNN + LSTM + MLP combined neural network |
CN111882810B (en) * | 2020-07-31 | 2022-07-01 | 广州市微智联科技有限公司 | Fire identification and early warning method and system |
CN111882810A (en) * | 2020-07-31 | 2020-11-03 | 广州市微智联科技有限公司 | Fire identification and early warning method and system |
CN112297033A (en) * | 2020-11-03 | 2021-02-02 | 泰州市华仕达机械制造有限公司 | Automatic change robotic arm actuating system |
CN112418004A (en) * | 2020-11-05 | 2021-02-26 | 烟台艾睿光电科技有限公司 | Fire safety detection method and related components |
CN112418004B (en) * | 2020-11-05 | 2022-11-11 | 烟台艾睿光电科技有限公司 | Fire safety detection method and related components |
CN112907886A (en) * | 2021-02-07 | 2021-06-04 | 中国石油化工股份有限公司 | Refinery plant fire identification method based on convolutional neural network |
CN114299445A (en) * | 2021-12-11 | 2022-04-08 | 南方电网大数据服务有限公司 | Power distribution room danger source identification method and device, computer equipment, medium and product |
CN114604708A (en) * | 2022-05-13 | 2022-06-10 | 深圳正为格智能科技有限公司 | Elevator fire risk early warning method and system |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110032977A (en) | A kind of safety warning management system based on deep learning image fire identification | |
CN109271554B (en) | Intelligent video identification system and application thereof | |
US8744125B2 (en) | Clustering-based object classification | |
Çetin et al. | Video fire detection–review | |
Bertini et al. | Multi-scale and real-time non-parametric approach for anomaly detection and localization | |
CN101833838B (en) | Large-range fire disaster analyzing and early warning system | |
CN102136059B (en) | Video- analysis-base smoke detecting method | |
RU2393544C2 (en) | Method and device to detect flame | |
CN105426820B (en) | More people's anomaly detection methods based on safety monitoring video data | |
CN112270331A (en) | Improved billboard detection method based on YOLOV5 | |
CN111126293A (en) | Flame and smoke abnormal condition detection method and system | |
Malhi et al. | Vision based intelligent traffic management system | |
CN108898098A (en) | Early stage video smoke detection method based on monitor supervision platform | |
Marchisio et al. | fakeWeather: Adversarial attacks for deep neural networks emulating weather conditions on the camera lens of autonomous systems | |
Tao et al. | Smoke vehicle detection based on robust codebook model and robust volume local binary count patterns | |
Huo et al. | 3DVSD: An end-to-end 3D convolutional object detection network for video smoke detection | |
CN108965804A (en) | A kind of video structural technology for city security protection | |
CN114997279A (en) | Construction worker dangerous area intrusion detection method based on improved Yolov5 model | |
Shi et al. | Moving cast shadow detection in video based on new chromatic criteria and statistical modeling | |
KR101467307B1 (en) | Method and apparatus for counting pedestrians using artificial neural network model | |
CN117789255A (en) | Pedestrian abnormal behavior video identification method based on attitude estimation | |
CN117475353A (en) | Video-based abnormal smoke identification method and system | |
CN110796008A (en) | Early fire detection method based on video image | |
Karishma et al. | Artificial Intelligence in Video Surveillance | |
Anusiya et al. | Density map based estimation of crowd counting using Vgg-16 neural network |
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 | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190719 |
|
RJ01 | Rejection of invention patent application after publication |