CN108549862A - Abnormal scene detection method and device - Google Patents
Abnormal scene detection method and device Download PDFInfo
- Publication number
- CN108549862A CN108549862A CN201810319654.9A CN201810319654A CN108549862A CN 108549862 A CN108549862 A CN 108549862A CN 201810319654 A CN201810319654 A CN 201810319654A CN 108549862 A CN108549862 A CN 108549862A
- Authority
- CN
- China
- Prior art keywords
- image
- network model
- neural network
- convolution neural
- abnormal
- 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
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- 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
Abstract
Exception scene detection method and device provided by the invention, the video data sent by receiving first terminal, video data is sampled according to the default sampling interval to obtain multiple images data, each image data is input in preset first convolution neural network model, determine whether image is abnormal image according to the calibration result of the first convolution neural network model, if it is determined that there is abnormal image, the image information of the abnormal image is then sent to second terminal, emergency response measure is taken in time so as to the user of second terminal, avoids the generation of second accident.By the above method, solve the problems, such as existing in the prior art low to railway or Highways ' exception scene Detection accuracy.
Description
Technical field
The present invention relates to technical field of video image processing more particularly to a kind of abnormal scene detection method and device.
Background technology
Current state-owned railroads construction has become the ring to hold the balance in national development strategy, and easily the railway network is country
Development provides guarantee, and more people's lives are provided convenience.When the natural calamity such as mud-rock flow, landslide etc. for encountering burst
When the problems such as causing rail burial and rail road foundation to cave in, very big prestige is constituted to the safety of Rail Transit System
Therefore the side of body in railway operation management, carries out safety patrol inspection to railway and context and is very important link, to ensure
The safety traffic of train.
Currently, the common method of railway department is manual inspection, worker patrols along rail, excludes different around rail
Reason condition, it is ensured that railway passes through safe.This method is time-consuming and laborious, and is difficult to hidden to the natural calamity that may occur in time
Trouble is accurately investigated.Therefore correlative theses propose the detecting system that the autonomic monitoring based on space base is realized by unmanned plane,
Inspection is carried out to the environment of Along Railway.Unmanned plane can acquire the image data on ground by the camera of carrying, and with meter
Calculation machine vision technique carries out intellectual analysis and processing, judges the scenario of Along Railway, realizes autonomous inspection, it is ensured that
Railway security.
However, being influenced by natural conditions such as illumination, Changes in weather, traditional computer vision technique is difficult accurately
Detection and identification are made to the natural calamity of Along Railway, therefore, detecting system demands perfection urgently.
Invention content
The present invention provides a kind of abnormal scene detection method and device, solves existing in the prior art to railway or public affairs
The low problem of curb line exception scene Detection accuracy.
The first aspect of the present invention provides a kind of abnormal scene detection method, including：
Receive the video data that first terminal is sent；
The video data is sampled according to the default sampling interval, obtains multiple images data；
Each described image data are input in the first convolution neural network model；The first convolution neural network model
Whether it is abnormal image for uncalibrated image；
Determine whether described image is abnormal image according to the calibration result of the first convolution neural network model, if
It is that the information of the abnormal image is then sent to second terminal.
Further, it is described each described image data are input in the first convolution neural network model before, further include：
The image pattern of first scene database is imported into convolutional neural networks, and the convolutional neural networks are carried out pre-
Training, obtains the second convolution neural network model；First scene database includes different types of scene；
The image pattern of second scene database is imported into the second convolution neural network model；The second scene number
Include natural calamity scene according to library；
Second training is carried out to the second convolution neural network model, obtains the first convolution neural network model.
Further, the image pattern by the second scene database import the second convolution neural network model it
Before, further include：
The image pattern of second scene database is divided into training set and test set；
Multi-scale transform and multi-angle transformation, the training after being expanded are carried out to the image pattern in the training set
Collection；
The image pattern by the second scene database imports the second convolution neural network model, including：By institute
It states the training set after expanding and imports the second convolution neural network model；
It is described that second training is carried out to the second convolution neural network model, the first convolution neural network model is obtained,
Including：
Second training is carried out to the second convolution neural network model in training set after the expansion, obtains third
Convolutional neural networks model；
The test set is input to the third convolutional neural networks model, obtains test result；
According to the calibration accuracy rate of third convolutional neural networks model described in the test result calculations；
It determines that the calibration accuracy rate meets and presets accuracy rate, then using the third convolutional neural networks model as described in
First convolution neural network model.
Further, the multi-scale transform is to be zoomed in and out according to preset ratio to image pattern, and cut center
Domain；
The multi-angle transformation is to be rotated according to predetermined angle to image pattern, and cut central area.
Further, it is described each described image data are input in the first convolution neural network model before, further include：
According to the input type of the first convolution neural network model, each described image data are zoomed in and out and cut
Processing.
Further, the first convolution neural network model connects entirely including 16 convolutional layers, five pond layers, two
Connect layer and one softmax layers.
Further, before the information by the abnormal image is sent to second terminal, further include：
The attribute information of the abnormal image is obtained, the attribute information includes shooting time and the bat of described image data
It acts as regent and sets；
The information by the abnormal image is sent to second terminal, including：
The attribute information of the abnormal image and the abnormal image is sent to second terminal.
The second aspect of the present invention provides a kind of abnormal scene detection device, including：
Receiving module, the video data for receiving first terminal transmission；
Sampling module obtains multiple images data for being sampled to the video data according to the default sampling interval；
Import modul, for each described image data to be input in the first convolution neural network model；The first volume
Whether product neural network model is abnormal image for uncalibrated image；
Determining module, for according to the calibration result of the first convolution neural network model determine described image whether be
Abnormal image；
If it is determined that described image is abnormal image, sending module, for the information of the abnormal image to be sent to second
Terminal.
The third aspect of the present invention provides a kind of computer readable storage medium, is stored thereon with computer program, described
The method described in any one of first aspect is realized when program is executed by processor.
The fourth aspect of the present invention provides a kind of electronic equipment, including：
Processor；And
Memory, the executable instruction for storing the processor；
Wherein, the processor is configured to execute described in any one of first aspect via the executable instruction is executed
Method.
Exception scene detection method and device provided in an embodiment of the present invention, the video counts sent by receiving first terminal
According to being sampled to obtain multiple images data to video data according to the default sampling interval, each image data is input to default
The first convolution neural network model in, determine whether image is abnormal according to the calibration result of the first convolution neural network model
Image, however, it is determined that have abnormal image, then the image information of the abnormal image is sent to second terminal, so as to the use of second terminal
Emergency response measure is taken at family in time, avoids the generation of second accident.By the above method, solve existing in the prior art
The problem low to railway or Highways ' exception scene Detection accuracy.
Description of the drawings
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technology description to be briefly described, it should be apparent that, the accompanying drawings in the following description is this hair
Some bright embodiments for those of ordinary skill in the art without having to pay creative labor, can be with
Obtain other attached drawings according to these attached drawings.
Fig. 1 is the flow diagram for the abnormal scene detection method that one embodiment of the invention provides；
Fig. 2 is the modeling procedure figure for the first convolutional neural networks that one embodiment of the invention provides；
Fig. 3 is the flow diagram for the first convolution neural network model features localization that one embodiment of the invention provides；
Fig. 4 is the structural schematic diagram for the abnormal scene detection device that one embodiment of the invention provides；
Fig. 5 is the structural schematic diagram of electronic equipment embodiment provided by the invention.
Specific implementation mode
In order to make the object, technical scheme and advantages of the embodiment of the invention clearer, below in conjunction with the embodiment of the present invention
In attached drawing, technical scheme in the embodiment of the invention is clearly and completely described, it is clear that described embodiment is
A part of the embodiment of the present invention, instead of all the embodiments.Based on the embodiments of the present invention, those of ordinary skill in the art
The every other embodiment obtained without creative efforts, shall fall within the protection scope of the present invention.
Term " first ", " second ", " third " in description and claims of this specification and above-mentioned attached drawing etc. are
For distinguishing similar object, without being used to describe specific sequence or precedence.It should be appreciated that the data used in this way
It can be interchanged in the appropriate case, so that the embodiment of the present invention described herein can be in addition to illustrating or describing herein
Sequence other than those is implemented.
It should be appreciated that term " comprising " and " having " used herein and their any deformation, it is intended that cover
It includes to be not necessarily limited to for example, containing the process of series of steps or unit, method, system, product or equipment to cover non-exclusive
Those of clearly list step or unit, but may include not listing clearly or for these processes, method, product
Or the other steps or unit that equipment is intrinsic.
The present invention provides a kind of abnormal scene detection method and device, existing in the prior art to railway or public affairs to solve
The low problem of the natural calamity scene detection accuracy rate of curb line.
In order to keep the above objects, features and advantages of the present invention more obvious and easy to understand, below in conjunction with the accompanying drawings and specifically
The present invention is described in detail for embodiment.
Fig. 1 is the flow diagram for the abnormal scene detection method that one embodiment of the invention provides, as shown in Figure 1, this reality
The method for applying example is applied to the railway system or highway system, is not especially limited to this present embodiment.
Detection method includes the following steps for exception scene provided in this embodiment：
S101, first terminal send video data to abnormal scene detection device；
In this step, first terminal is mobile terminal or fixed terminal with camera shooting or camera function, Ke Yishi
Unmanned machine equipment can also be the video camera mounted on railway or Highways '.Wherein, pacify on unmanned machine equipment or video camera
Equipped with high-definition camera, the video data for acquiring railway or Highways ' in real time, it will be understood by those skilled in the art that the
The video data of one terminal taking is made of each frame image data, and each image data includes shooting time and the bat of image frame
It acts as regent and the information such as sets, for when detecting abnormal image, time of origin and accident point coordinates to be determined according to image data.
The video data of shooting is real-time transmitted to abnormal scene detection device by the first terminal of the present embodiment, so that detection
Device is detected and demarcates to photographed data in time, and when there is burst accident, railway or highway system in time, can be obtained accurately
Accident point information is taken, the emergency capability of railway or highway system is promoted.
Optionally, the quantity of the first terminal of the present embodiment is at least one, and railway or highway system can use simultaneously
Different types of first terminal, i.e., using unmanned machine equipment in such a way that fixed video camera is combined, to expand railway or highway
The coverage area of data acquisition along the line.
S102, abnormal scene detection device sample video data according to the default sampling interval, obtain multiple images
Data；
Detection device is after the video data for receiving first terminal transmission, according to the preset sampling interval, for example, often
Two frames of second, sample video data, obtain multiple images data.
Specifically, detection device periodically can make above-mentioned sampling process to one section of video data, for example, detection device
With 1 minute or 30 seconds for the period, the video data of the duration is sampled.
Optionally, when the quantity of first terminal is two or more, detection device receives multiple first eventually simultaneously
The video data sent, the video data that multiple first terminals can be sent is held to pass through the parallel arrangement of sampling module of detection device
Data sampling is carried out respectively, finally each image data of each first terminal of same period is summarized, obtains image data set, is used
To execute S103.
Each image data is input in the first convolution neural network model by S103, abnormal scene detection device；
Wherein, whether the first convolution neural network model is abnormal image for uncalibrated image；
The first convolution neural network model in the present embodiment is by different type scene database and abnormal scene number
It repeatedly trains what is obtained there is the network model for determining accuracy rate compared with high standard according to library, can relatively accurately judge that first terminal acquires
Each image data it is whether abnormal.
Specifically, the first convolution neural network model is by the image of the feature of each image data and preset abnormal image
Feature is compared, if matching with the characteristics of image of preset abnormal image, it is determined that the image is abnormal image, if not having
The feature to match with the characteristics of image of preset abnormal image, it is determined that the image is normal picture.
Optionally, before each image data is input in the first convolution neural network model, detection device needs root
According to the input type of the first convolution neural network model, for example, input picture is the image of 224*224 pixels, to each picture number
According to zoom in and out and cutting processing.
S104, abnormal scene detection device determine that image is abnormal according to the calibration result of the first convolution neural network model
Image；
S105, abnormal scene detection device send the information of abnormal image to second terminal.
In the present embodiment, the calibration result of the first convolution neural network model is set as 0 and 1, specifically：
When it is abnormal image to determine image, then the calibration result for providing the image is 0；
When it is normal picture to determine image, then the calibration result for providing the image is 1.
Detection device determines whether image is abnormal image according to the calibration result of the first convolution neural network model, if figure
As being abnormal image, then the information of the abnormal image is sent to second terminal by detection device, for the work people of second terminal
Member takes responsive measures, the patrol officer's debugging nearby of notice fault point in time, while notifying the driver of online car, with
The generation to avoid traffic accident.
Specifically, the information of abnormal image includes the information such as shooting time and the camera site of image.
For unmanned machine equipment, GPRS positioning devices and clock are provided on unmanned machine equipment, in shooting video counts
According to when, automatically generate the shooting time and taking location information of each frame image.
For the video camera mounted on railway or Highways ', every video camera is internally provided with clock, is shooting
Records photographing time when video, and when sending video data, the mark of the video camera is carried, detection device passes through video camera
Mark inquire pre-stored location data table, determine the installation site of the video camera, and then can determine each image, especially extremely
The shooting time of image and camera site.
Second terminal in the present embodiment can be the portable terminals such as mobile terminal, such as smart mobile phone, laptop
End, can also be fixed terminal, such as railway or the server terminal of highway system command centre, does not make to have to this present embodiment
Body limits.
Exception scene detection method provided in this embodiment, the video data sent by receiving first terminal, according to pre-
If the sampling interval samples video data to obtain multiple images data, each image data is input to preset first convolution
In neural network model, determine whether image is abnormal image according to the calibration result of the first convolution neural network model, if really
Surely there is abnormal image, then the image information of the abnormal image is sent to second terminal, adopted in time so as to the user of second terminal
Emergency response measure is taken, the generation of second accident is avoided.The above method provided by the embodiment solves existing in the prior art
The problem low to railway or Highways ' exception scene Detection accuracy.
By above-described embodiment, it is known that the first convolution neural network model is preset convolutional neural networks model, for marking
Determine whether image is abnormal image, how to establish and train with the network model for determining accuracy rate compared with high standard, is that the present invention carries
The key technology point of the abnormal scene detection method of confession.With reference to specific embodiment to the preset first convolution nerve net
The process of establishing of network model elaborates.
Fig. 2 is the modeling procedure figure for the first convolutional neural networks that one embodiment of the invention provides, as shown in Fig. 2, this is built
Mold process includes the following steps：
S201, the image pattern of the first scene database is imported to convolutional neural networks, and convolutional neural networks is carried out
Pre-training obtains the second convolution neural network model；
Wherein, the first scene database includes different types of scene；
The first scene database in the present embodiment can be specifically Places365 databases, and Places databases are mesh
Preceding larger scene image data library, the database contain the picture of a large amount of different type scenes.
Pre-training, the second obtained convolution nerve net are carried out to convolutional neural networks by introducing Places365 databases
Network model has the ability of higher scene classification.
Specifically, the forward-propagating by convolutional neural networks and backpropagation mechanism, continuous iteration updates network model
Parameter optimizes the performance of network model.During pre-training, in order to alleviate the inefficiencies of training stage, over-fitting is prevented
Problem can improve classification performance by Dropout methods, accelerate trained convergence rate.Dropout methods refer to each
During wheel training, the neuron of several ratios is randomly selected, makes its connection weight with other nodes without update, it is this
Method trains the model come and has more generalization ability.
S202, the image pattern for obtaining the second scene database, and the image pattern of the second scene database is divided into instruction
Practice collection and test set；
Wherein, the second scene database includes natural calamity scene；
Specifically, on the basis of S201, image pattern pair the second convolution nerve net of the second scene database is utilized
Network model is finely adjusted, with Optimized model parameter.Wherein, the second scene database is specially the video counts of railway or Highways '
Include the normal scene image of railway or Highways ' and abnormal scene image, abnormal scene image according to library, in the database
Specially natural calamity scene image.
By the fine tuning to the second convolution neural network model, obtained convolutional neural networks model can be preferably applicable in
In the special scenes of the present embodiment, preferably abnormal scene detection result is obtained.
In this step, it, for example, training set accounts for 70% and is only used for training, can be tested according to preset classifying rules
Collection accounts for 30%, for test model and calculates calibration accuracy rate, classifying rules is set according to specific demand, to this present embodiment
It is not especially limited.
It should be pointed out that the image pattern in the second scene database is the image pattern demarcated in advance, including just
Sample and negative sample, wherein positive sample are the image pattern of normal scene, and negative sample is the image pattern of abnormal scene.
S203, multi-scale transform and multi-angle transformation, the training after being expanded are carried out to the image pattern in training set
Collection；
Since the image pattern of the second scene database preservation in S202 is usually from different terminals, the shooting of each terminal
There may be differences for parameter, therefore, it is necessary to be pre-processed to the image pattern of the second scene database, for example, according to pre-
If the image pattern of the second scene database of size pair carries out the adjustment of picture size.
On the basis of above-mentioned pretreated, multiple dimensioned and multi-angle is carried out to the image pattern in training set and is converted, to expand
Fill training set.
Specifically, multi-scale transform is to be zoomed in and out according to preset ratio to image pattern, and cut central area.Example
Such as, 1/2 times of image pattern full size, the 1/4 times expansion sample for cutting central area after zooming in and out as multi-scale transform is used
This.
Multi-angle transformation is to be rotated according to predetermined angle to image pattern, and cut central area.For example, by image
Sample rotates the exptended sample for cutting that central area is multi-angle transformation after -90 °, -45 ° ,+45 ° ,+90 °.
S204, the training set after expansion is imported into the second convolution neural network model；
Second training is carried out to the second convolution neural network model in S205, the training set after expansion, obtains third volume
Product neural network model；
By the second training to the second convolution neural network model, third convolutional neural networks model, the network are obtained
Model has certain abnormal capacity rating, and whether calibration accuracy is qualified, needs the verification by S206-S208.
S206, test set is input to third convolutional neural networks model, obtains test result；
It should be pointed out that the image pattern in test set is the image pattern of known calibration result, it is above-mentioned in order to verify
The third convolutional neural networks model that step is established whether can accurate calibration provide the image of off-note, will be in test set
Multiple images sample be input to third convolutional neural networks model, obtain the test result of multiple images sample, test result
For the calibration result of multiple series of images sample.
S207, according to the calibration accuracy rate of test result calculations third convolutional neural networks model；
According to above-mentioned test result, the number of the image pattern of the quantity for demarcating correct image pattern and calibration mistake is determined
Amount, to which the calibration accuracy rate of the third convolutional neural networks model under the test set be calculated.Wherein, demarcate accuracy rate=
Demarcate the total quantity of image pattern in quantity/test set of correct image pattern.
S208, judge to demarcate whether accuracy rate meets default accuracy rate, if satisfied, then by third convolutional neural networks model
As the first convolution neural network model.
According to the result of calculation of S208, determine whether calibration accuracy rate is up to standard, for example, preset calibration accuracy rate is
95%, then only third convolutional neural networks model of the calibration accuracy rate more than or equal to 95% could be used as final preset first
Convolutional neural networks model, for the abnormality detection to practical railway or highway system.
Optionally, periodically increase new abnormal image data, and the training on preset first convolution neural network model,
To continue to optimize the performance of network model, it is made preferably to play a role.
The modeling process for present embodiments providing the first convolution neural network model, the first volume established by the above method
Product neural network model has higher calibration accuracy rate, and by being continuously increased new abnormal image sample, and based on update
Image data base pair the first convolution neural network model afterwards is trained, and can further promote the first convolution neural network model
Calibration performance.
On the basis of the various embodiments described above, it should be pointed out that the first convolution neural network model of the present embodiment has
Body includes：16 convolutional layers, five pond layers, two full articulamentums and one softmax layers.Wherein,
For convolutional layer, the size of the convolution kernel of each convolutional layer is 3*3, and each convolution kernel presses a pixel separation pair
Input image data (input vector) carries out convolution operation.The activation primitive used in each convolutional layer is ReLu functions yi=
Max { xi, 0 }, wherein xi are the inner product of i-th of weight matrix and current layer input that previous output is connected, i.e. convolution results, yi
For the output of corresponding convolutional layer.
For pond layer, the pond method of generally use has mean value pond method, random pool method and maximum pondization side
Method, the pond layer of the present embodiment is using maximum pond method.
For full articulamentum and softmax layers, dimension is respectively 4096 and 2.Wherein, softmax layers of final output mark
Determine result (0 or 1).
The features localization for carrying out image data to first volume product neural network model below elaborates.
Fig. 3 is the flow diagram for the first convolution neural network model features localization that one embodiment of the invention provides, such as
Shown in Fig. 3, which is unfolded by taking an image pattern as an example, specifically includes following steps：
S301, image pattern is continued through to two the first convolutional layer convolution, then maximum pond is carried out by the first pond layer
Change；
Image pattern in the present embodiment should meet the input type of the first convolution neural network model.Therefore, detection dress
It sets before image pattern is inputted the first convolution neural network model, needs to pre-process image pattern, pretreatment is specific
Including scaling and cutting.For example, image pattern is zoomed to 256*256 pixel sizes, and further it is cut to 224*224 pixels
Size.Then, it will be input in the first convolution neural network model by pretreated image pattern, and execute S301.
Two the first convolutional layers in this step are the convolutional layer in 64 channels.
S302, the output result of the first pond layer is continued through to two the second convolutional layer convolution again, then passes through the second pond
Change layer and carries out maximum pond；
Two the second convolutional layers in this step are the convolutional layer in 128 channels.
S303, the output result of the second pond layer is continued through to three third convolutional layer convolution again, then passes through third pond
Change layer and carries out maximum pond；
Three third convolutional layers in this step are the convolutional layer in 256 channels.
S304, the output result of third pond layer is continued through to three Volume Four lamination convolution again, then passes through the 4th pond
Change layer and carries out maximum pond；
Three Volume Four laminations in this step are the convolutional layer in 512 channels.
S305, the output result of the 4th pond layer is continued through to three the 5th convolutional layer convolution again, then passes through the 5th pond
Change layer and carries out maximum pond；
Three the 5th convolutional layers in this step are the convolutional layer in 512 channels.
S306, the output result of the 5th pond layer is continued through to two full articulamentums and one softmax layers again, obtained
The features localization result of image pattern.
The features localization result of the image pattern of the present embodiment is 0 or 1.Wherein,
0 represents abnormal image feature, and 1 represents normal picture feature.
By above-mentioned calibration process, realize that the abnormal of the image pattern to being input to the first convolution neural network model is examined
It surveys, the calibration of characteristics of image is carried out using the first convolution neural network model of the present embodiment, there is higher accuracy, be convenient for
Detection device in time makes a response abnormal scene image according to the output result of the first convolution neural network model, improves detection
The detection efficiency of device.
Fig. 4 is the structural schematic diagram for the abnormal scene detection device that one embodiment of the invention provides, as shown in figure 4, this reality
The abnormal scene detection device 40 of example is applied, including：
Receiving module 41, the video data for receiving first terminal transmission；
Sampling module 42 obtains multiple images number for being sampled to the video data according to the default sampling interval
According to；
Import modul 43, for each described image data to be input in the first convolution neural network model；Described first
Whether convolutional neural networks model is abnormal image for uncalibrated image；
Determining module 44, for whether determining described image according to the calibration result of the first convolution neural network model
For abnormal image；
If it is determined that described image is abnormal image, sending module 45, for the information of the abnormal image to be sent to the
Two terminals.
Exception scene detection device provided in this embodiment, can execute the technical solution of above method embodiment, in fact
Existing principle is similar with technique effect, and details are not described herein again for the present embodiment.
The present embodiment provides a kind of computer readable storage mediums, are stored thereon with computer program, the computer journey
The technical solution as described in any one of aforementioned embodiment of the method, implementing principle and technical effect are realized when sequence is executed by processor
Similar, details are not described herein again.
Fig. 5 is the structural schematic diagram of electronic equipment embodiment provided by the invention, as shown in figure 5, the electronics of the present embodiment
Equipment, including：
Processor 51；And
Memory 52, the executable instruction for storing the processor 51；
Wherein, the processor 51 is configured to execute any one of aforementioned method implementation via the executable instruction is executed
Technical solution described in example, implementing principle and technical effect are similar, and details are not described herein again.
The function of each module in above system can be realized by processor 51.
One of ordinary skill in the art will appreciate that：Realize that all or part of step of above-mentioned each method embodiment can lead to
The relevant hardware of program instruction is crossed to complete.Program above-mentioned can be stored in a computer read/write memory medium.The journey
When being executed, execution includes the steps that above-mentioned each method embodiment to sequence；And storage medium above-mentioned includes：ROM、RAM、SRAM、
The various media that can store program code such as DRAM, FLASH, magnetic disc or CD.
Finally it should be noted that：The above embodiments are only used to illustrate the technical solution of the present invention., rather than its limitations；To the greatest extent
Present invention has been described in detail with reference to the aforementioned embodiments for pipe, it will be understood by those of ordinary skill in the art that：Its according to
So can with technical scheme described in the above embodiments is modified, either to which part or all technical features into
Row equivalent replacement；And these modifications or replacements, various embodiments of the present invention technology that it does not separate the essence of the corresponding technical solution
The range of scheme.
Claims (10)
1. a kind of exception scene detection method, which is characterized in that including：
Receive the video data that first terminal is sent；
The video data is sampled according to the default sampling interval, obtains multiple images data；
Each described image data are input in the first convolution neural network model；The first convolution neural network model is used for
Whether uncalibrated image is abnormal image；
Determine whether described image is abnormal image according to the calibration result of the first convolution neural network model, if so,
The information of the abnormal image is sent to second terminal.
2. according to the method described in claim 1, it is characterized in that, described be input to the first convolution god by each described image data
Before in network model, further include：
The image pattern of first scene database is imported into convolutional neural networks, and the convolutional neural networks are instructed in advance
Practice, obtains the second convolution neural network model；First scene database includes different types of scene；
The image pattern of second scene database is imported into the second convolution neural network model；Second scene database
Including natural calamity scene；
Second training is carried out to the second convolution neural network model, obtains the first convolution neural network model.
3. according to the method described in claim 2, it is characterized in that, the image pattern by the second scene database imports institute
Before stating the second convolution neural network model, further include：
The image pattern of second scene database is divided into training set and test set；
Multi-scale transform and multi-angle transformation, the training set after being expanded are carried out to the image pattern in the training set；
The image pattern by the second scene database imports the second convolution neural network model, including：By the expansion
Training set after filling imports the second convolution neural network model；
It is described that second training is carried out to the second convolution neural network model, the first convolution neural network model is obtained, including：
Second training is carried out to the second convolution neural network model in training set after the expansion, obtains third convolution
Neural network model；
The test set is input to the third convolutional neural networks model, obtains test result；
According to the calibration accuracy rate of third convolutional neural networks model described in the test result calculations；
It determines that the calibration accuracy rate meets and presets accuracy rate, then using the third convolutional neural networks model as described first
Convolutional neural networks model.
4. according to the method described in claim 3, it is characterized in that,
The multi-scale transform is to be zoomed in and out according to preset ratio to image pattern, and cut central area；
The multi-angle transformation is to be rotated according to predetermined angle to image pattern, and cut central area.
5. according to the method described in claim 1, it is characterized in that, described be input to the first convolution god by each described image data
Before in network model, further include：
According to the input type of the first convolution neural network model, to each described image data zoom in and out and cutting at
Reason.
6. according to claim 1-5 any one of them methods, which is characterized in that the first convolution neural network model includes
16 convolutional layers, five pond layers, two full articulamentums and one softmax layers.
7. according to the method described in claim 1, it is characterized in that, the information by the abnormal image is sent to second eventually
Before end, further include：
The attribute information of the abnormal image is obtained, the attribute information includes shooting time and the shooting position of described image data
It sets；
The information by the abnormal image is sent to second terminal, including：
The attribute information of the abnormal image and the abnormal image is sent to second terminal.
8. a kind of exception scene detection device, which is characterized in that including：
Receiving module, the video data for receiving first terminal transmission；
Sampling module obtains multiple images data for being sampled to the video data according to the default sampling interval；
Import modul, for each described image data to be input in the first convolution neural network model；The first convolution god
For uncalibrated image whether it is abnormal image through network model；
Determining module, for determining whether described image is abnormal according to the calibration result of the first convolution neural network model
Image；
If it is determined that described image is abnormal image, sending module, for the information of the abnormal image to be sent to second terminal.
9. a kind of computer readable storage medium, is stored thereon with computer program, which is characterized in that described program is by processor
Claim 1-7 any one of them methods are realized when execution.
10. a kind of electronic equipment, which is characterized in that including：
Processor；And
Memory, the executable instruction for storing the processor；
Wherein, the processor is configured to carry out perform claim requirement 1-7 any one of them via the execution executable instruction
Method.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810319654.9A CN108549862A (en) | 2018-04-11 | 2018-04-11 | Abnormal scene detection method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810319654.9A CN108549862A (en) | 2018-04-11 | 2018-04-11 | Abnormal scene detection method and device |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108549862A true CN108549862A (en) | 2018-09-18 |
Family
ID=63514429
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810319654.9A Pending CN108549862A (en) | 2018-04-11 | 2018-04-11 | Abnormal scene detection method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108549862A (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109653798A (en) * | 2018-12-21 | 2019-04-19 | 浙江合作力科技股份有限公司 | A kind of tunnel safety detection warning system and method |
| CN109685131A (en) * | 2018-12-20 | 2019-04-26 | 斑马网络技术有限公司 | Automobile vehicle device system exception recognition methods and device |
| CN109726771A (en) * | 2019-02-27 | 2019-05-07 | 深圳市赛梅斯凯科技有限公司 | Abnormal driving detection model method for building up, device and storage medium |
| CN110008916A (en) * | 2019-04-11 | 2019-07-12 | 深圳云朵数据科技有限公司 | A kind of video image storage method, apparatus, equipment and storage medium |
| CN110851349A (en) * | 2019-10-10 | 2020-02-28 | 重庆金融资产交易所有限责任公司 | Page abnormal display detection method, terminal equipment and storage medium |
| CN110956783A (en) * | 2019-12-04 | 2020-04-03 | 江河瑞通（北京）技术有限公司 | Urban waterlogging monitoring method and electronic equipment |
| CN111160289A (en) * | 2019-12-31 | 2020-05-15 | 欧普照明股份有限公司 | Method and device for detecting accident of target user and electronic equipment |
| CN111460909A (en) * | 2020-03-09 | 2020-07-28 | 兰剑智能科技股份有限公司 | Vision-based goods location management method and device |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104134080A (en) * | 2014-08-01 | 2014-11-05 | 重庆大学 | Method and system for automatically detecting roadbed collapse and side slope collapse of road |
| CN104657706A (en) * | 2015-01-27 | 2015-05-27 | 北京大学 | Image-based high-speed railway line pole breakage abnormality and connecting structural body abnormality detection method |
| US20170262962A1 (en) * | 2016-03-11 | 2017-09-14 | Qualcomm Incorporated | Systems and methods for normalizing an image |
| CN107563356A (en) * | 2017-09-29 | 2018-01-09 | 西安因诺航空科技有限公司 | A kind of unmanned plane inspection pipeline target analysis management method and system |
-
2018
- 2018-04-11 CN CN201810319654.9A patent/CN108549862A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104134080A (en) * | 2014-08-01 | 2014-11-05 | 重庆大学 | Method and system for automatically detecting roadbed collapse and side slope collapse of road |
| CN104657706A (en) * | 2015-01-27 | 2015-05-27 | 北京大学 | Image-based high-speed railway line pole breakage abnormality and connecting structural body abnormality detection method |
| US20170262962A1 (en) * | 2016-03-11 | 2017-09-14 | Qualcomm Incorporated | Systems and methods for normalizing an image |
| CN107563356A (en) * | 2017-09-29 | 2018-01-09 | 西安因诺航空科技有限公司 | A kind of unmanned plane inspection pipeline target analysis management method and system |
Non-Patent Citations (3)
| Title |
|---|
| 刘文祺: "基于深度神经网络的铁路异物检测算法", 《中国优秀硕士学位论文全文数据库工程科技Ⅱ辑》 * |
| 常燕敏: "无人机影像在地震灾区道路损毁应急评估中的应用研究", 《中国优秀硕士学位论文全文数据库基础科学辑》 * |
| 胡耀: "自然场景中物体识别算法研究", 《中国优秀硕士学位论文全文数据库信息科技辑》 * |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109685131A (en) * | 2018-12-20 | 2019-04-26 | 斑马网络技术有限公司 | Automobile vehicle device system exception recognition methods and device |
| CN109653798A (en) * | 2018-12-21 | 2019-04-19 | 浙江合作力科技股份有限公司 | A kind of tunnel safety detection warning system and method |
| CN109653798B (en) * | 2018-12-21 | 2020-04-28 | 浙江合作力科技股份有限公司 | Safety detection alarm system and method for tunnel |
| CN109726771A (en) * | 2019-02-27 | 2019-05-07 | 深圳市赛梅斯凯科技有限公司 | Abnormal driving detection model method for building up, device and storage medium |
| CN110008916A (en) * | 2019-04-11 | 2019-07-12 | 深圳云朵数据科技有限公司 | A kind of video image storage method, apparatus, equipment and storage medium |
| CN110851349A (en) * | 2019-10-10 | 2020-02-28 | 重庆金融资产交易所有限责任公司 | Page abnormal display detection method, terminal equipment and storage medium |
| CN110956783A (en) * | 2019-12-04 | 2020-04-03 | 江河瑞通（北京）技术有限公司 | Urban waterlogging monitoring method and electronic equipment |
| CN110956783B (en) * | 2019-12-04 | 2021-11-16 | 江河瑞通（北京）技术有限公司 | Urban waterlogging monitoring method and electronic equipment |
| CN111160289A (en) * | 2019-12-31 | 2020-05-15 | 欧普照明股份有限公司 | Method and device for detecting accident of target user and electronic equipment |
| CN111460909A (en) * | 2020-03-09 | 2020-07-28 | 兰剑智能科技股份有限公司 | Vision-based goods location management method and device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108549862A (en) | Abnormal scene detection method and device | |
| US20200043324A1 (en) | Method for obtaining road condition information, apparatus thereof, and storage medium | |
| CN111563557B (en) | Method for detecting target in power cable tunnel | |
| CN107481526A (en) | System and method for drive a vehicle lane change detection record and lane change violating the regulations report control | |
| CN105913604A (en) | Fire occurrence determining method and device based on unmanned aerial vehicle | |
| CN103838772B (en) | A kind of multi-source traffic data fusion method | |
| CN107220633A (en) | A kind of intelligent mobile enforcement system and method | |
| CN110414359B (en) | Long-distance pipeline unmanned aerial vehicle inspection data analysis and management method and system | |
| CN110516529A (en) | It is a kind of that detection method and system are fed based on deep learning image procossing | |
| CN109034036A (en) | A kind of video analysis method, Method of Teaching Quality Evaluation and system, computer readable storage medium | |
| CN107730993A (en) | The parking lot intelligent vehicle-tracing system and method identified again based on image | |
| CN106485923B (en) | A kind of public transport crowding real-time status acquisition method and device | |
| Giyenko et al. | Application of convolutional neural networks for visibility estimation of CCTV images | |
| CN111126411B (en) | Abnormal behavior identification method and device | |
| CN109359169A (en) | A kind of retrograde behavior real-time identification method of the shared bicycle based on probability graph model | |
| CN115083112B (en) | Intelligent early warning emergency management system and deployment method thereof | |
| CN109961056A (en) | Traffic accident responsibility identification, system and equipment based on decision Tree algorithms | |
| CN106205135B (en) | A kind of detection method, apparatus and system and a kind of ball machine of the vehicle behavior violating the regulations that turns around | |
| CN110087041A (en) | Video data processing and transmission method and system based on the base station 5G | |
| Guo et al. | Dense traffic detection at highway-railroad grade crossings | |
| CN203324805U (en) | Intelligent driving school supervision system | |
| CN112989916A (en) | Crowd counting method combining density estimation and target detection | |
| CN112364864A (en) | License plate recognition method and device, electronic equipment and storage medium | |
| CN115439933A (en) | Garbage classification release site detection method based on multiple model recognition strategies | |
| CN115482473A (en) | Graph convolution network model for extracting aerial image features and method for detecting abnormity |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180918 |