CN110503021A - Fire hazard smoke detecting method based on time compression track characteristic identification - Google Patents
Fire hazard smoke detecting method based on time compression track characteristic identification Download PDFInfo
- Publication number
- CN110503021A CN110503021A CN201910763584.0A CN201910763584A CN110503021A CN 110503021 A CN110503021 A CN 110503021A CN 201910763584 A CN201910763584 A CN 201910763584A CN 110503021 A CN110503021 A CN 110503021A
- Authority
- CN
- China
- Prior art keywords
- network
- fire
- time
- sub
- smog
- 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
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
- G06V20/41—Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
- G06V20/46—Extracting features or characteristics from the video content, e.g. video fingerprints, representative shots or key frames
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- Computational Linguistics (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of fire hazard smoke detecting methods based on time compression track characteristic identification, the time that length is T before current time t is taken to compress track as trace image, by the slice convolution sub-network extraction feature based on vgg16, using the temporal signatures sub-network based on RNNs, then by the differentiation sub-network connected entirely, differentiate in scene cigarette, fire whether occurs.The present invention extracts time upper incidence relation using Recognition with Recurrent Neural Network, and resolution capability is improved by the way of piecemeal, using Weakly supervised training method, smog can be accurately and rapidly detected and issue fire alarm.Experimental result shows that performance is good in the accuracy rate, sensibility and false detection rate that smog identifies for this method.
Description
Technical field
The present invention relates to smog identification technology fields, and in particular to a kind of fire based on time compression track characteristic identification
Smog detection method.
Background technique
Fire alarm is always the important topic in security against fire field.Fire occurs that a large amount of smog would generally be generated early period.
If smog can be timely detected, so that it may provide fire alarm earlier, reduce casualties and property loss.Tradition
Smog detection method be typically based on the features such as color, texture, shape, movement, achieve certain achievement, but in practical application
In still have problem, it is often effective in certain occasions, and changed application scenarios less effective.This essentially consists in the static state of smog
It is too big that feature such as color, texture etc. by weather, illumination, time etc. is influenced variation range, and behavioral characteristics such as background modeling or
Inter-frame difference is typically based on threshold value, and threshold value is affected to testing result.
In recent years, deep neural network recognition of face, in terms of achieve breakthrough.Depth mind
Through network using original image as input, more abstract high-level characteristic is formed by combination shallow-layer feature, it can be found that data
Profound distributed nature avoids manually extracting the complicated lengthy and jumbled of feature work.Hohberg in 2015 etc. uses convolutional Neural
Network (convolutional neural networks, CNN) (Hohberg etc., 2015), Frizzi in 2016 etc. are used
LeNet (Frizzi etc., 2016), Tao etc. identify smog using Alex Net (Tao etc., 2016).Frizzi etc. uses sliding
Window samples video frame, and carries out flame and cigarette detection with convolutional neural networks, improves the complexity of algorithm, reduces
The range of smoke detection, so that network can only see regional area, and cannot only see the overall situation, to increase false alarm
Risk (Frizzi etc., 2016).The it is proposeds such as Chen Junzhou are based on concatenated convolutional neural network cigarette texture recognition frame, and fusion is static
It is using original image as static texture input, the light stream sequence of original image is defeated as dynamic texture with dynamic texture information
Enter (Chen Junzhou etc., 2016).The use Res Net extraction static characteristic identification such as Filonenko smog (Filonenko etc.,
2017), Yin etc. using RNNs (Recurrent Neural Networks) extract dynamic characteristic identification smog (M.Yin etc.,
2017).Yin etc. proposes 14 layers of standardization convolutional neural networks D-NCNN (deep normalization
Andconvolutional neural network), the convolutional layer in traditional CNN is improved to batch standardization convolutional layer, effectively
Ground solves the problems, such as gradient disperse and over-fitting in network training process, accelerates training process with this and improves detection effect
(Z.Yin etc., 2017).Zhang etc. uses Faster R-CNN, with the classification image block of sliding window (Zhang etc.,
2018).But these detect the smog that figurate method is not particularly suited for coming in every shape.The it is proposeds such as Luo are based on background
Dynamic updates and dark channel prior detects moving target, then automatically extracts suspicious region feature by CNN, then carry out smog identification
(Y.Luo etc., 2018).Dung etc. also first extracts moving region, and it is a series of to reuse color, growth, size, edge energy etc.
It is smog that cascade classifier, which is differentiated,.Cascade classifier afterbody used convolutional neural networks CNN (Dung etc.,
2018).If the frame that smog occurs in certain time is more than setting quantity, it is considered as that smog has occurred.Li et al. is by Smoke Detection
It regards as and is partitioned into target from the 3D space-time body of monitor video, using the frame similar with FCN, examined with the full convolutional network of 3D
It surveys smog (Li et al., 2018).2019 Nian Yuanfei oxen etc., which propose, forms basic block, Mei Gejuan by multiple dimensioned, multiple parallel convolutional layer
Lamination has the filter of identical quantity, but has different scale invariabilities.Batch processing standardizes after each convolutional layer, so
Summarize all normalized outputs from multiple dimensioned parallel layer afterwards, and activate and as block final output (Yuan Feiniu etc.,
2019).Luo etc. is by image to u, v direction projection, cumulative, it is believed that in compression image smog track tendency, shape, frequency,
Polymerization etc. is different from (S.Luo etc., 2015) with non-cigarette target.Using conventional method, such as SVM, detect individual
Smog track is relatively difficult.
Summary of the invention
To solve the above problems, the present invention provides the fire hazard aerosol fog detection sides based on time compression track characteristic identification
Method is extracted time upper incidence relation using Recognition with Recurrent Neural Network, resolution capability is improved by the way of piecemeal, using Weakly supervised
Training method can accurately and rapidly detect smog and issue fire alarm.
To achieve the above object, the technical scheme adopted by the invention is as follows:
Based on the fire hazard smoke detecting method of time compression track characteristic identification, take length before current time t be T when
Between compression track as trace image, by the slice convolution sub-network extraction feature based on vgg16, using with RNNs
Based on temporal signatures sub-network differentiate in scene cigarette, fire whether occurs then by the differentiation sub-network that connects entirely.
Further, every frame u-v image is added up to the direction u and v, time pressure is then obtained using time t as abscissa
Contracting trace image, i.e.,
F in formulat(u, v) is the t frame in video, and w, h are that frame is wide and vertical frame dimension, ft(v) be video v-t time compressed tracks
Mark, ftIt (u) is u-t time of video to compress track；V-t, u-t are spliced, the time compression track f of video is obtainedt。
Further, vgg16 is used to be sliced convolution sub-network for basic construction, the feature of extraction is through average max
512 dimensional features are exported after pooling, using two full articulamentum dimensionality reductions to 156,32 dimensions, are output to the time domain of next stage
Recycle sub-network.
Further, temporal signatures sub-network introduces 3 door hold modes, receives the output result h of last momentT-1、
The system mode c at current timet-1X is inputted with current systemt, pass through input gate ct, forget door rtWith out gate ztMore new system
State ctAnd by final result htIt is exported, therefore
In formula, Wxz、Wxr、Wxc、Whz、Whr、WhcIndicate the weighting parameters of each door, bz、br、bcIndicate biasing；GRU passes through anti-
Learn these parameters to propagating, because the number of parameters of RNNs is numerous, in order to reduce the requirement to sample, temporal signatures sub-network
Hiding layer unit only takes 64.
Further, while cigarette, fire are detected, two three full articulamentums (64,24,7,1) is added after time domain cyclic network
Differentiation network judge whether there is cigarette, fire in scene respectively.
It further, is p there are the probability of smog in scenes, fume loss function is
In formula, N is training samples number, LsFor the label of smog,For the L2 model of parameter on Smoke Detection access
Number, λsFor the weighting coefficient of Smoke Detection access L2 norm.
Probability in the presence of fire is pf, fiery loss function is
In formula, LfFor the label of flame,For the L2 norm of parameter on fire defector access, λfFor fire defector
The weighting coefficient of access L2 norm.
Total loss function is
λ in formula1、λ2Respectively smog, fire defector access L2 norm weighting coefficient.
The invention has the following advantages:
Time upper incidence relation is extracted using Recognition with Recurrent Neural Network, resolution capability is improved by the way of piecemeal, use is weak
The training method of supervision can accurately and rapidly detect smog and issue fire alarm.Experimental result shows that this method is in cigarette
Performance is good on accuracy rate, sensibility and the false detection rate that mist identifies.
Detailed description of the invention
Fig. 1 is the traj+CNN+RNNs network frame schematic diagram in the embodiment of the present invention.
Specific embodiment
The present invention is described in detail combined with specific embodiments below.Following embodiment will be helpful to the technology of this field
Personnel further understand the present invention, but the invention is not limited in any way.It should be pointed out that the ordinary skill of this field
For personnel, without departing from the inventive concept of the premise, various modifications and improvements can be made.These belong to the present invention
Protection scope.
Based on the fire hazard smoke detecting method of time compression track characteristic identification, take length before current time t be T when
Between compression track as trace image, by the slice convolution sub-network extraction feature based on vgg16, using with RNNs
Based on temporal signatures sub-network differentiate in scene cigarette, fire whether occurs then by the differentiation sub-network that connects entirely.Tool
Body:
Firstly, taking video in the frame F of current time tt(u, v), the projection compression on u and v coordinate, namely to u and the side v
To adding up, i.e.,
Obtain trace image ft(v)、ft(u)。ft(v)、ftIt (u) is the vector of w × 3, h × 3, wherein w, h are the width of frame
And height；
By ft(v)、ft(u) building slice convolution sub-network of the input based on vgg16 extracts track characteristic after connecting,
This network architecture is known as traj+CNN+RNNs, such as Fig. 1.The feature of extraction exports 512 after average max pooling
Dimensional feature is output to the time domain cyclic sub-network of next stage using two full articulamentum dimensionality reductions to 156,32 dimensions.
Temporal signatures sub-network introduces 3 door hold modes, receives the output result h of last momentT-1, current time
System mode ct-1X is inputted with current systemt, pass through input gate ct, forget door rtWith out gate ztUpdate system mode ctAnd it will
Final result htIt is exported, therefore
In formula, Wxz、Wxr、Wxc、Whz、Whr、WhcIndicate the weighting parameters of each door, bz、br、bcIndicate biasing；GRU passes through anti-
Learn these parameters to propagating, because the number of parameters of RNNs is numerous, in order to reduce the requirement to sample, GRU hides layer unit
Only take 64.
In order to improve the detection accuracy of smog, while cigarette, fire are detected, two three full connections are added after time domain cyclic network
The differentiation network of layer (64,24,7,1) judges whether there is cigarette, fire in scene respectively.
In the present embodiment, in scene there are the probability of smog be ps, fume loss function is
In formula, N is training samples number, LsFor the label of smog,For the L2 of parameter on Smoke Detection access
Norm, λsFor the weighting coefficient of Smoke Detection access L2 norm.
Probability in the presence of fire is pf, fiery loss function is
In formula, LfFor the label of flame,For the L2 norm of parameter on fire defector access, λfFor fire defector
The weighting coefficient of access L2 norm.
Total loss function is
λ in formula1、λ2Respectively smog, fire defector access L2 norm weighting coefficient.
Experimental example
Evaluation index
It (1) is the degree of dependence to hardware of verification algorithm, we are provided with frame per second rf, maximum consumption memory RAM and net
Network weight storage size tri- indexs of W.Frame per second is the frame number of processing per second, i.e.,
Rf=Nsx150/Ta (5)
Wherein, sample number when Ns is test, Ta are the time for having handled the cost of Ns sample.
It (2) is the accuracy of verification algorithm, the index of experiment are as follows:
In formula, ACC is accuracy rate, and TPR is real rate, and TNR is very negative rate, and total number of samples when Ns is test, TP is true
Positive sample number, TN are true negative sample number, and FP is false positive sample number, and FN is false negative sample number.
(3) Smoke Detection is time-effectiveness, and loss can be reduced by being detected in early stage fire.It is detection algorithm to smog
Sensitivity, be provided with sensibility index ff.Sensibility index is the frame detected when the probability that cigarette occurs in scene is greater than 0.5
Start the difference for occurring between the frame ordinal number of cigarette in ordinal number and label,
Ff=Nps > 0.5-Nl (7)
In formula, ps is the probability that cigarette occurs, and Nl is to start the frame ordinal number for cigarette occur in label.
Average sensitivity index is
Fa=Σ ffi/Ns (8)
In formula, i-th of sample number sensibility index when ffi is test.Sensibility index reflects resolution of the algorithm to weak cigarette
Ability.
Test data does not participate in training process as unknown scene.Test data set1 include VSD 10 sections of videos 1,
15 sections of videos 1 of Bilkent and 16 sections of videos 2 of KMU and 12 sections of videos of laboratory shooting, including 33 sections of cigarettes, 20 sections of nothings
Cigarette video amounts to 53 sections of videos.For the sensibility of testing algorithm, weak cigarette that test data is concentrated (from scratch, from it is small to
Big and remote small smog) it picks out, including 6 sections of cigarettes, 5 sections of smokeless videos, as set2.
The left-half of part Experiment video is positive sample, and right half part is negative sample.It is indoor smog on the left of the first row,
It is night wagon flow on the right side of the first row；It is forest smog on the left of second row, is haze on the right side of the second row；It is outdoor cigarette on the left of the third line
Mist, the third line right side are flame；It is city fog on the left of fourth line, is flag on the right side of fourth line；It is open ring on the left of fifth line
Smog under border, fifth line right side are cloud.The strong cigarette of four behaviors before left-half, fifth line are weak cigarette.
Allocation of computer is Intel Core i7-6700CPU 2.60Hz, and video card is NVIDIA Geforce GTX
1060。
Compare algorithm
For the robustness of comparison algorithm, sensibility and fails to report, reports situation, more fully assessment models performance by mistake, it is right
Five kinds of CNN, C3D convolutional network, traj+SVM, traj+RNNs and traj+CNN+RNNs algorithms are compared.CNN uses migration instruction
Practice, 512 dimension shape features of video every frame image are extracted on the basis of vgg16, then judge scene by three layers of full articulamentum
In whether there is smog.C3D convolutional network extracts scene dynamics feature directly from video, by two-dimensional convolution nuclear convolution image
Mode Three dimensional convolution nuclear convolution video.Whether traj+SVM is compressed in track using SVM analysis time comprising smog.traj+
Time is compressed track dimensionality reduction after three layers of full connection and is fed directly into RNNs at 32 dimensional features by RNNs, judge scene whether include
Smog；Traj+CNN+RNNs is then first to be sliced, then extract shape feature with CNN, is judged again with RNNs after full connection dimensionality reduction
Method.In these methods, the part CNN in CNN and traj+CNN+RNNs all uses vgg16 and carries out as basic network
Transfer training, C3D convolutional network transfer training on the basic network of (2015) Tran D etc..It is limited by sample size, simultaneously
It is compared also on same baseline, RNNs uses the GRU that historic state hidden unit is remembered with 64.
Test result on data set set1
Test result on data set set1 is as shown in table 1.Relative to the conventional method of traj+SVM, traj+CNN+
Accuracy, the sensibility of RNNs is all improved largely, and improves 35.2% than traj+SVM in accuracy rate, in very negative rate
15.6% is improved, this shows that the method for deep learning reduces false detection rate while guaranteeing positive inspection rate.But it is based on depth
The algorithm complexity of study is relatively high, and frame per second, maximum consumption memory and the network weight of traj+CNN+RNNs is respectively
49fps, 2.31G and 261M, and the frame per second of traditional traj+SVM is 178fps.
Test result on 1 data set set1 of table
*: when the time span Ts of slice takes 32,1 respectively, corresponding susceptibility.
The multinomial performance of traj+CNN+RNNs is better than traj+RNNs, and ACC, TPR, TNR are respectively increased 15.9%, 25.1%
With 39.6%, show that extract shape feature again after slice is of great benefit to for improving false alarm rate.But traj+CNN+RNNs
Sensibility is not so good as traj+RNNs, this is because span Ts is set as 32 when slice.If the span Ts of slice is set as 1, traj+
The sensibility of CNN+RNNs is then the highest in various algorithms, only 19.
Other than traj+CNN+RNNs method, the effect of C3D is the most excellent.This may have benefited from the airspace of C3D, time domain
It handles simultaneously, the feature for having taste can be extracted more.But on the other hand, the algorithm complexity of C3D is also highest.
In method based on deep learning, the effect of CNN is worst, only slightly better than traditional traj+SVM.This may is that
Because CNN basic network is the vgg16 for predominantly detecting rigid-object, and is not suitable with translucent smog；And the method for CNN does not have
Have in view of temporal signatures, therefore sensibility is also poor.
Test result on data set set2 is as shown in table 2.Contrast table 1, it can be seen that CNN, C3D two methods exist
It is greatly lowered in the indexs such as ACC, TPR, TNR and sensibility；And three kinds of methods based on time compression track, traj+
SVM, traj+RNNs, traj+CNN+RNNs are also maintained at higher level.Especially traj+CNN+RNNs algorithm, ACC, TPR,
TNR and sensibility also reach 0.853,0.847,0.872 and 52/26.
Test result on 2 data set set2 of table
*: when the time span Ts of slice takes 32,1 respectively, corresponding susceptibility.
Experimental result shows that method of the invention shows in the accuracy rate, sensibility and false detection rate that smog identifies
Well.
Specific embodiments of the present invention are described above.It is to be appreciated that the invention is not limited to above-mentioned
Particular implementation, those skilled in the art can make a variety of changes or modify within the scope of the claims, this not shadow
Ring substantive content of the invention.In the absence of conflict, the feature in embodiments herein and embodiment can any phase
Mutually combination.
Claims (6)
1. the fire hazard smoke detecting method based on time compression track characteristic identification, it is characterised in that: take long before current time t
The time compression track that degree is T is as trace image, by the slice convolution sub-network extraction feature based on vgg16, then
By the temporal signatures sub-network based on RNNs, then by the differentiation sub-network connected entirely, differentiate whether send out in scene
Make cigarette, fire.
2. the fire hazard smoke detecting method as described in claim 1 based on time compression track characteristic identification, it is characterised in that:
Every frame u-v image is added up to the direction u and v, time compression trace image is then obtained using time t as abscissa, i.e.,
F in formulat(u, v) is the t frame in video, and w, h are that frame is wide and vertical frame dimension, ftIt (v) is v-t time of video to compress track, ft
It (u) is u-t time of video to compress track；V-t, u-t are spliced, the time compression track f of video is obtainedt。
3. the fire hazard smoke detecting method as described in claim 1 based on time compression track characteristic identification, it is characterised in that:
Vgg16 is used to be sliced convolution sub-network for basic construction, the feature of extraction exports 512 dimensions after average max pooling
Feature is output to the time domain cyclic sub-network of next stage using two full articulamentum dimensionality reductions to 156,32 dimensions.
4. the fire hazard smoke detecting method as described in claim 1 based on time compression track characteristic identification, it is characterised in that:
Temporal signatures sub-network introduces 3 door hold modes, receives the output result h of last momentT-1, current time system shape
State ct-1X is inputted with current systemt, pass through input gate ct, forget door rtWith out gate ztUpdate system mode ctAnd by final knot
Fruit htIt is exported, therefore
In formula, Wxz、Wxr、Wxc、Whz、Whr、WhcIndicate the weighting parameters of each door, bz、br、bcIndicate biasing；GRU is by reversely passing
It broadcasts and learns these parameters, temporal signatures sub-network hides layer unit and only takes 64.
5. the fire hazard smoke detecting method as described in claim 1 based on time compression track characteristic identification, it is characterised in that:
Cigarette, fire are detected simultaneously, the differentiation network that two three full articulamentums (64,24,7,1) are added after time domain cyclic network judges respectively
Whether cigarette, fire are had in scene.
6. the fire hazard smoke detecting method as described in claim 1 based on time compression track characteristic identification, it is characterised in that:
In scene there are the probability of smog be ps, fume loss function is
In formula, N is training samples number, LsFor the label of smog, //W1 A//2For the L2 norm of parameter on Smoke Detection access, λs
For the weighting coefficient of Smoke Detection access L2 norm；
Probability in the presence of fire is pf, fiery loss function is
In formula, LfFor the label of flame,For the L2 norm of parameter on fire defector access, λfFor fire defector access L2
The weighting coefficient of norm；
Total loss function is
λ in formula1、λ2Respectively smog, fire defector access L2 norm weighting coefficient.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910763584.0A CN110503021A (en) | 2019-08-19 | 2019-08-19 | Fire hazard smoke detecting method based on time compression track characteristic identification |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910763584.0A CN110503021A (en) | 2019-08-19 | 2019-08-19 | Fire hazard smoke detecting method based on time compression track characteristic identification |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110503021A true CN110503021A (en) | 2019-11-26 |
Family
ID=68588313
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910763584.0A Pending CN110503021A (en) | 2019-08-19 | 2019-08-19 | Fire hazard smoke detecting method based on time compression track characteristic identification |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110503021A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111079572A (en) * | 2019-11-29 | 2020-04-28 | 南京恩博科技有限公司 | Forest smoke and fire detection method based on video understanding, storage medium and equipment |
| CN111145222A (en) * | 2019-12-30 | 2020-05-12 | 浙江中创天成科技有限公司 | Fire detection method combining smoke movement trend and textural features |
| CN111639610A (en) * | 2020-06-03 | 2020-09-08 | 北京思湃德信息技术有限公司 | Fire recognition method and system based on deep learning |
| CN111797761A (en) * | 2020-07-02 | 2020-10-20 | 温州智视科技有限公司 | Three-stage smoke detection system, method and readable medium |
| CN116503715A (en) * | 2023-06-12 | 2023-07-28 | 南京信息工程大学 | Forest fire detection method based on cascade network |
-
2019
- 2019-08-19 CN CN201910763584.0A patent/CN110503021A/en active Pending
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111079572A (en) * | 2019-11-29 | 2020-04-28 | 南京恩博科技有限公司 | Forest smoke and fire detection method based on video understanding, storage medium and equipment |
| CN111145222A (en) * | 2019-12-30 | 2020-05-12 | 浙江中创天成科技有限公司 | Fire detection method combining smoke movement trend and textural features |
| CN111639610A (en) * | 2020-06-03 | 2020-09-08 | 北京思湃德信息技术有限公司 | Fire recognition method and system based on deep learning |
| CN111797761A (en) * | 2020-07-02 | 2020-10-20 | 温州智视科技有限公司 | Three-stage smoke detection system, method and readable medium |
| CN111797761B (en) * | 2020-07-02 | 2023-05-16 | 温州智视科技有限公司 | Three-stage smoke detection system, method and readable medium |
| CN116503715A (en) * | 2023-06-12 | 2023-07-28 | 南京信息工程大学 | Forest fire detection method based on cascade network |
| CN116503715B (en) * | 2023-06-12 | 2024-01-23 | 南京信息工程大学 | Forest fire detection method based on cascade network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110503021A (en) | Fire hazard smoke detecting method based on time compression track characteristic identification | |
| Li et al. | Adaptively constrained dynamic time warping for time series classification and clustering | |
| EP3869459B1 (en) | Target object identification method and apparatus, storage medium and electronic apparatus | |
| CN105868689B (en) | A kind of face occlusion detection method based on concatenated convolutional neural network | |
| Çetin et al. | Video fire detection–review | |
| Zhao et al. | SVM based forest fire detection using static and dynamic features | |
| CN105787472B (en) | A kind of anomaly detection method based on the study of space-time laplacian eigenmaps | |
| CN106682635A (en) | Smoke detecting method based on random forest characteristic selection | |
| CN107506734A (en) | One kind of groups unexpected abnormality event detection and localization method | |
| CN111046827A (en) | Video smoke detection method based on convolutional neural network | |
| CN110852441A (en) | Fire early warning method based on improved naive Bayes algorithm | |
| Huang et al. | A fire-alarming method based on video processing | |
| Chowdhury et al. | Computer vision and smoke sensor based fire detection system | |
| CN113963301A (en) | Space-time feature fused video fire and smoke detection method and system | |
| CN115690564A (en) | Outdoor fire smoke image detection method based on Recursive BIFPN network | |
| CN108334902A (en) | A kind of track train equipment room smog fireproof monitoring method based on deep learning | |
| CN105894022A (en) | Adaptive hierarchical association multi-target tracking method | |
| CN103699888A (en) | Human face detection method and device | |
| CN110991245A (en) | Real-time smoke detection method based on deep learning and optical flow method | |
| Chen et al. | A novel smoke detection algorithm based on improved mixed Gaussian and YOLOv5 for textile workshop environments | |
| CN110503092A (en) | The improvement SSD monitor video object detection method adapted to based on field | |
| CN106384089A (en) | Human body reliable detection method based on lifelong learning | |
| Liang et al. | Random forest feature selection and back propagation neural network to detect fire using video | |
| Asatryan et al. | Method for fire and smoke detection in monitored forest areas | |
| CN114926764A (en) | Method and system for detecting remnants in industrial scene |
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 |