CN109325439A - A kind of recognition methods of the driving unlawful practice based on CNN - Google Patents
A kind of recognition methods of the driving unlawful practice based on CNN Download PDFInfo
- Publication number
- CN109325439A CN109325439A CN201811087757.3A CN201811087757A CN109325439A CN 109325439 A CN109325439 A CN 109325439A CN 201811087757 A CN201811087757 A CN 201811087757A CN 109325439 A CN109325439 A CN 109325439A
- Authority
- CN
- China
- Prior art keywords
- layer
- cnn
- network
- training
- picture
- 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.)
- Granted
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
- G06V20/59—Context or environment of the image inside of a vehicle, e.g. relating to seat occupancy, driver state or inner lighting conditions
- G06V20/597—Recognising the driver's state or behaviour, e.g. attention or drowsiness
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24133—Distances to prototypes
- G06F18/24137—Distances to cluster centroïds
- G06F18/2414—Smoothing the distance, e.g. radial basis function networks [RBFN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
Abstract
The recognition methods for the driving unlawful practice based on CNN that the invention discloses a kind of, is arranged training parameter with the artificial intelligence frame of independent research, and every layer is built CNN network, carries out each layer of image recognition training.This stage is the process of iteration, and each layer network is relatively independent training mission, can be with controlled training effect in every layer so that it is convenient to be regulated and controled for weak link.Finally trained three networks are spelled, just be completed from initial monitor video to driver's unlawful practice.Solve the confinement problems of traditional Open Framework.
Description
Technical field
The present invention relates to video identification field, in particular to the recognition methods of a kind of driving unlawful practice based on CNN.
Background technique
In China's road traffic accident certainly caused by driver's operation error and fatigue driving, i.e., most accidents are
Be that outstanding experienced driver is also different surely keeps good driving condition for a long time.Therefore, driving for driver is monitored
Behavior is sailed, accurately can quickly identify the unlawful practice in driving, inherently reduces the generation of traffic accident situation with important
Meaning.
Currently, having been achieved for some research achievements in terms of monitoring driver's driving behavior again both at home and abroad, it is broadly divided into
Two kinds: one is based on hardware device, according to driver breathe in alcohol content judge whether to drink;According to driver's
The relative reflection principle of eyelid and eyeball come judge driver whether fatigue driving;According to the brain of driver that wave or electrocardiogram
To determine whether some monitoring drivers such as fatigue driving come in the device for physiologically whether being in normal condition to driver's
Driving condition is evaluated.Another kind is analyzed based on Computer Image Processing and mode identification technology, for driver's head
Portion's activity condition, facial characteristics (such as eyes, head face) change feature to judge driving behavior and the spirit shape of driver
State.For the latter since technical level is limited, the technical application of artificial intelligence is less, and accuracy of identification is not high and speed is slower, often needs
It is aided with artificial.For ten hundreds of monitor videos, it is also less scientific for being monitored by the way of semi-artificial and driving violation.
Therefore in a kind of method of artificial intelligence, computer high-precision is allowed to go identification violation driving behavior necessary.
In recent years, it is based on the feature learning method of CNN (Convolution Neural Network, convolutional neural networks)
Immense success is achieved in terms of image classification, causes the very big concern of computer vision field.
The limitation of traditional Open Framework.Current all well-known artificial intelligence learning frameworks by the U.S. company and
Mechanism publication, function, feature, the tendentiousness of these artificial intelligence frames are held by these publication tissues, this is also entire
AI industry is difficult to realize commercial one of the major reasons in different field, because small-to-medium business can only almost use this
A little disclosed frames carry out the research and development of upper layer packaging type, are limited technical strength, and it is fixed according to real demand and different application depth to lack
The ability of artificial intelligence frame processed.
Summary of the invention
It is an object of the invention to: a kind of recognition methods of driving unlawful practice based on CNN is provided, solves tradition
Open Framework confinement problems.
The technical solution adopted by the invention is as follows:
A kind of recognition methods of the driving unlawful practice based on CNN, comprising the following steps:
S1, multilayer driving behavior CNN identification network is established, the CNN identification network includes three-layer network;
S2, interception monitor video image;
S3, the monitor video classmate for intercepting step S2 are sent into the first layer for the CNN identification network that step S1 is established, CNN
The first layer identification video image of identification network belongs to interior scene or car scene, will belong to the video image of interior scene
It is sent into step S4, the image for belonging to car scene is eliminated;
The second layer of S4, CNN identification network identifies the step S3 video image being sent into, and extracts in video image
Driver head part is distributed into step S5;
The third layer of S5, CNN identification network identifies the step S4 video image being sent into, and judges the behavior of driver
With the presence or absence of in violation of rules and regulations.
Training parameter is set with the artificial intelligence frame of independent research, every layer is built CNN network, carries out each layer of image
Recognition training.This stage is the process of iteration, and each layer network is relatively independent training mission, can be to control in every layer
Training effect for weak link so that it is convenient to be regulated and controled.Finally by trained three networks spell get up, just complete from
Initial monitor video is to driver's unlawful practice.
Further, the method that multilayer driving behavior CNN identification network is established in the step S1 includes following step
It is rapid:
S101, first layer CNN identification network is established;
S102, second layer CNN identification network is established;
S103, third layer CNN identification network is established;
S104, three layers of CNN for establishing S101, S102, S10 identify series network, form final network structure.
The method of multitiered network training.Sophisticated identification task is split as multitiered network substep to identify, using multitiered network
Structure is realized that individually training both can guarantee recognition speed to each layer network, moreover it is possible to accurate compared with greatly improving under small sample
Rate, and can be adjusted according to the training result of each layer network.The present invention will identify that driver makes a phone call, smokes, eats east
Western, fatigue driving behavior is split into outside the interior vehicle of identification, and identification drives the number of people, identification concrete behavior three parts.
Further, every layer of CNN identification network foundation the following steps are included:
S105, sample prepare, and sample is divided into training set and test set;
S106, the CNN network structure for constructing respective layer;
S107, it will be trained in the CNN network structure of training set image steps for importing S106 foundation.
S108, training are completed, and the CNN network of respective layer is obtained.
Further, the step S107 will be carried out in CNN network structure that training set image steps for importing S106 is established
It further include being converted to training set image before training.Whether image mapping function is settable opens, in the training of every wheel, if
The transformation of training set image is opened, training set sample will be done a round transformation again.Directly sample size is expanded.Expanding data
Collection realizes industrial application.
Further, the method for described image transformation includes transverse translation, longitudinal translation, rotation, change picture contrast,
Change brightness, setting fuzzy region range and fog-level and adjustment noise size, and the transformation of every class can all be controlled in detail
System transformation quantity.
Further, the training frame that recognition training is carried out in the step S107 uses the calculation of dynamical learning rate
Method sentences convergence algorithm with automatic.In order to realize intelligent training, alternatively referred to as on-hook training.This CNN training frame uses dynamic
The algorithm of learning rate sentences convergence algorithm with automatic.With the expansion of training round, learning rate can be according in reversed gradient algorithm
Change of gradient and dynamic adjust, gradually reduce to preset value.Change of gradient changes within a certain period of time is less than threshold value, then being
System will be completed voluntarily in deconditioning, mark training.After the completion of training, with test program obtained network file to unknown sample
Collection carries out Classification and Identification, then is slightly aided with artificial correction, just can very easily expand training dataset, be iterated training.Most
The network class accuracy rate realized eventually can reach 99.8%.
Further, each layer of sample preparation includes:
First layer training: need to prepare the picture of interior picture, car scene;
Second layer training: need be in interior scene driver's number of people and be not the number of people picture;
Third layer training: the picture of violation and the not picture of violation are needed in driver's number of people.
In conclusion by adopting the above-described technical solution, the beneficial effects of the present invention are:
1. a kind of recognition methods of the driving unlawful practice based on CNN of the present invention, is fully automated, using artificial intelligence
Mode is installed machines to replace manual labor.
2. a kind of recognition methods of the driving unlawful practice based on CNN of the present invention, recognition speed is fast, and accuracy of identification is high, solution
It has determined the confinement problems of traditional Open Framework.
Detailed description of the invention
Examples of the present invention will be described by way of reference to the accompanying drawings, in which:
Fig. 1 is the multitiered network realization procedure chart of use of the invention;
Fig. 2 is repetitive exercise procedure chart of the invention;
Fig. 3 is the CNN structure chart of the first layer network of the invention;
Fig. 4 is the CNN structure chart of the second layer network of the invention;
Fig. 5 is the CNN structure chart of third layer network of the invention.
Specific embodiment
All features disclosed in this specification or disclosed all methods or in the process the step of, in addition to mutually exclusive
Feature and/or step other than, can combine in any way.
It elaborates below with reference to Fig. 1 to Fig. 5 to the present invention.
Embodiment 1
A kind of recognition methods of the driving unlawful practice based on CNN, comprising the following steps:
S1, multilayer driving behavior CNN identification network is established, the CNN identification network includes three-layer network;
S2, interception monitor video image;
S3, the monitor video classmate for intercepting step S2 are sent into the first layer for the CNN identification network that step S1 is established, CNN
The first layer identification video image of identification network belongs to interior scene or car scene, will belong to the video image of interior scene
It is sent into step S4, the image for belonging to car scene is eliminated;
The second layer of S4, CNN identification network identifies the step S3 video image being sent into, and extracts in video image
Driver head part is distributed into step S5;
The third layer of S5, CNN identification network identifies the step S4 video image being sent into, and judges the behavior of driver
With the presence or absence of in violation of rules and regulations.
Training parameter is set with the artificial intelligence frame of independent research again, every layer is built CNN network, carries out each layer of figure
As recognition training.This stage is the process of iteration, and each layer network is relatively independent training mission, can be to control in every layer
Training effect processed for weak link so that it is convenient to be regulated and controled.Finally trained three networks are spelled, just be completed
From initial monitor video to driver's unlawful practice.
Embodiment 2
The present embodiment the difference from embodiment 1 is that, further, establish multilayer driving behavior in the step S1
CNN identify network method the following steps are included:
S101, first layer CNN identification network is established;
S102, second layer CNN identification network is established;
S103, third layer CNN identification network is established;
S104, three layers of CNN for establishing S101, S102, S10 identify series network, form final network structure.
The method of multitiered network training.Sophisticated identification task is split as multitiered network substep to identify, using multitiered network
Structure is realized that individually training both can guarantee recognition speed to each layer network, moreover it is possible to accurate compared with greatly improving under small sample
Rate, and can be adjusted according to the training result of each layer network.The present invention will identify that driver makes a phone call, smokes, eats east
Western, fatigue driving behavior is split into outside the interior vehicle of identification, and identification drives the number of people, identification concrete behavior three parts.
Further, every layer of CNN identification network foundation the following steps are included:
S105, sample prepare, and sample is divided into training set and test set;
S106, the CNN network structure for constructing respective layer;
S107, it will be trained in the CNN network structure of training set image steps for importing S106 foundation.
S108, training are completed, and the CNN network of respective layer is obtained.
Further, the step S107 will be carried out in CNN network structure that training set image steps for importing S106 is established
It further include being converted to training set image before training.Whether image mapping function is settable opens, in the training of every wheel, if
The transformation of training set image is opened, training set sample will be done a round transformation again.Directly sample size is expanded.Expanding data
Collection realizes industrial application.
Further, the method for described image transformation includes transverse translation, longitudinal translation, rotation, change picture contrast,
Change brightness, setting fuzzy region range and fog-level and adjustment noise size, and the transformation of every class can all be controlled in detail
System transformation quantity.
Further, the training frame that recognition training is carried out in the step S107 uses the calculation of dynamical learning rate
Method sentences convergence algorithm with automatic.In order to realize intelligent training, alternatively referred to as on-hook training.This CNN training frame uses dynamic
The algorithm of learning rate sentences convergence algorithm with automatic.With the expansion of training round, learning rate can be according in reversed gradient algorithm
Change of gradient and dynamic adjust, gradually reduce to preset value.Change of gradient changes within a certain period of time is less than threshold value, then being
System will be completed voluntarily in deconditioning, mark training.After the completion of training, with test program obtained network file to unknown sample
Collection carries out Classification and Identification, then is slightly aided with artificial correction, just can very easily expand training dataset, be iterated training.Most
The network class accuracy rate realized eventually can reach 99.8%.
Further, each layer of sample preparation includes:
First layer training: need to prepare the picture of interior picture, car scene;
Second layer training: need be in interior scene driver's number of people and be not the number of people picture;
Third layer training: the picture of violation and the not picture of violation are needed in driver's number of people.
Embodiment 3
Referring to Fig.1, unlawful practice recognition methods is entirely driven to realize by three-layer network.
First layer identifies that the video is the scene outside interior or vehicle;
The second layer identifies wherein according to the interior scene that first layer sorts out comprising driver's number of people again;
Third layer in the interior scene there are driver's number of people sorted out according to the second layer, then identifies driver respectively
Whether make a phone call, smoke, eating, the behaviors such as fatigue driving.
Referring to Fig. 2, driving the training process that unlawful practice identifies, specific step is as follows:
S105 prepares training sample
Sample data of the invention is generated based on third party's monitor supervision platform passenger-cargo carriage Vehicular video, the monitor video of acquisition
Sample resolution generally 352*288.It can certainly obtain by other means, resolution ratio is also not necessarily limited to 352*288.To this
A little videos carry out key frame interception.
Since the present invention uses three-layer network, to prepare the training sample of three types.
First layer, interior car scene.By the compression of images of original 352*288 at 72*88.
The second layer, driver's Head recognition.Prepare the sample data that size is 32*32, being divided into is the one kind for driving the number of people
With the non-class for driving the number of people.This sample is to be come out by the image of former 352*288 according to the size cut of 64*64, writes down the upper left corner
Coordinate recompresses the size for 32*32.
Third layer, driver make a phone call to identify.According to the picture in second layer training sample, selecting is the figure for driving the number of people
Whether further according to making a phone call, smoking, eat etc., behaviors are divided into positive and negative two class.Size is 64*64.
After classification gets out sample, image preprocessing is carried out.Using OpenCV shape library by RGB255 color-map representation at ash
Degree figure, the grayscale image pixel value being converted to further increase efficiency, then are uniformly processed on the section 0-1 between 0 to 255.
S106 constructs CNN network structure in the neural network framework of independent research.
Use CNN network structure as follows outside interior vehicle as shown in figure 3, first layer of the present invention identifies:
This training network has ten layers, is made of an input layer, five convolutional layers, three pond layers and an output layer.
First layer is input layer, the map that input parameter is a 72*88.
The second layer is convolutional layer, and using Relu activation primitive, convolution mode is band edge convolution, and it is 3x3's that 6 sizes, which are arranged,
Convolution kernel, moving step length 1.Input dimension of picture is 72*88, and output dimension of picture is 72*88.
Third layer is pond layer, and according to maximum value pond, the core that 6 sizes are 2x2, moving step length 2 does not cover pond
Change.Input dimension of picture is 72*88, and output dimension of picture is 36*44.
4th layer is convolutional layer, and using Relu activation primitive, convolution mode is band edge convolution, and it is 3x3 that 12 sizes, which are arranged,
Convolution kernel, moving step length 1.Input dimension of picture is 36*44, and output dimension of picture is 36*44.
Layer 5 is pond layer, and according to maximum value pond, the core that 12 sizes are 2x2, moving step length 2 does not cover pond
Change.Input dimension of picture is 36*44, and output dimension of picture is 18*22.
Layer 6 is convolutional layer, and using Relu activation primitive, convolution mode is band edge convolution, and it is 3x3 that 24 sizes, which are arranged,
Convolution kernel, moving step length 1.Input dimension of picture is 18*22, and output dimension of picture is 18*22.
Layer 7 is convolutional layer, and using Relu activation primitive, convolution mode is band edge convolution, and it is 3x3 that 24 sizes, which are arranged,
Convolution kernel, moving step length 1.Input dimension of picture is 18*22, and output dimension of picture is 18*22
8th layer is convolutional layer, and using Relu activation primitive, convolution mode is band edge convolution, and it is 3x3 that 24 sizes, which are arranged,
Convolution kernel, moving step length 1.Input dimension of picture is 18*22, and output dimension of picture is 18*22
9th layer is pond layer, and using Relu activation primitive, the core that 24 sizes are 2x2, moving step length 2 is not covered.
Input dimension of picture is 18*22, and output dimension of picture is 9*11.
Tenth layer is output layer, and using Softmax activation primitive, number of classifying is 2.
As shown in figure 4, second layer identification of the present invention drives the number of people and uses CNN network structure as follows:
This training network has seven layers, is made of an input layer, four convolutional layers, a pond layer and an output layer.
First layer is input layer, the map that input parameter is a 32*32.
The second layer is convolutional layer, and using Relu activation primitive, convolution mode is band edge convolution, and it is 3x3 that 16 sizes, which are arranged,
Convolution kernel, moving step length 1.Input dimension of picture is 32*32, and output dimension of picture is 32*32.
Third layer is pond layer, and according to maximum value pond, the core that 16 sizes are 2x2, moving step length 2 does not cover pond
Change.Input dimension of picture is 32*32, and output dimension of picture is 16*16.
4th layer is convolutional layer, and using Relu activation primitive, convolution mode is band edge convolution, and it is 3x3 that 32 sizes, which are arranged,
Convolution kernel, moving step length 1.Input dimension of picture is 16*16, and output dimension of picture is 16*16.
Layer 5 is convolutional layer, and using Relu activation primitive, convolution mode is band edge convolution, and it is 3x3 that 32 sizes, which are arranged,
Convolution kernel, moving step length 1.Input dimension of picture is 16*16, and output dimension of picture is 16*16.
Layer 6 is convolutional layer, and using Relu activation primitive, convolution mode is band edge convolution, and it is 3x3 that 32 sizes, which are arranged,
Convolution kernel, moving step length 1.Input dimension of picture is 16*16, and output dimension of picture is 16*16
Layer 7 is output layer, and using Softmax activation primitive, number of classifying is 2.
As shown in figure 5, third layer identification of the present invention drives the number of people and uses CNN network structure as follows:
This training network has eight layers, is made of an input layer, four convolutional layers, two pond layers and an output layer.
First layer is input layer, the map that input parameter is a 64*64.
The second layer is convolutional layer, and using Relu activation primitive, convolution mode is band edge convolution, and it is 3x3 that 16 sizes, which are arranged,
Convolution kernel, moving step length 1.Input dimension of picture is 64*64, and output dimension of picture is 64*64.
Third layer is pond layer, and according to maximum value pond, the core that 16 sizes are 2x2, moving step length 2 does not cover pond
Change.Input dimension of picture is 64*64, and output dimension of picture is 32*32.
4th layer is convolutional layer, and using Relu activation primitive, convolution mode is band edge convolution, and it is 3x3 that 32 sizes, which are arranged,
Convolution kernel, moving step length 1.Input dimension of picture is 32*32, and output dimension of picture is 32*32.
Layer 5 is convolutional layer, and using Relu activation primitive, convolution mode is band edge convolution, and it is 3x3 that 32 sizes, which are arranged,
Convolution kernel, moving step length 1.Input dimension of picture is 32*32, and output dimension of picture is 32*32.
Layer 6 is convolutional layer, and using Relu activation primitive, convolution mode is band edge convolution, and it is 3x3 that 32 sizes, which are arranged,
Convolution kernel, moving step length 1.Input dimension of picture is 32*32, and output dimension of picture is 32*32
Layer 7 is pond layer, and according to maximum value pond, the core that 16 sizes are 2x2, moving step length 2 does not cover pond
Change.Input dimension of picture is 32*32, and output dimension of picture is 16*16.
8th layer is output layer, and using Softmax activation primitive, number of classifying is 2.
S109, the transformation of training set image.
The method of image transformation has: lateral, longitudinal translation, rotation change picture contrast, change brightness, and mould is arranged
Regional scope and fog-level are pasted, noise size is adjusted.
And transformation quantity can all be controlled in detail to the transformation of every class.Whether image mapping function is settable opens, in every wheel
In training, if opening the transformation of training set image, training set sample will be done a round transformation again.It can expand as needed
Data set is truly realized small sample and is trained, and realizes industrial application.
S107, training.
CNN training is begun to after the completion of image transformation.In order to realize intelligent training, alternatively referred to as on-hook training.This CNN instruction
Practice frame to use the algorithm of dynamical learning rate and sentence convergence algorithm automatically.With the expansion of training round, learning rate can basis
Change of gradient in reversed gradient algorithm and dynamic adjusts, gradually reduce to preset value.Change of gradient becomes within a certain period of time
Change and be less than threshold value, then system will be completed voluntarily in deconditioning, mark training.
The network finally realized, recognition accuracy of making a phone call reach 97%, smoking recognition accuracy 98%, identification of eating
Accuracy rate 97%.
Embodiment 4
The present invention is directed to the server-side hardware depth optimization of cross operating system, realizes in low-level hardware, with 4 core 2GHz's
The calculated performance of x86 framework CPU is under standard, for CIF resolution ratio Vehicular video, using 60s duration as standard interception five
The recognition speed of Zhang Tu, each behavior are less than 100ms.Accuracy of identification is high.Phone recognition accuracy reaches 97%, identification of smoking
Accuracy rate 98%, recognition accuracy 97% of eating.
The above, only the preferred embodiment of the present invention, but scope of protection of the present invention is not limited thereto, it is any
Those skilled in the art within the technical scope disclosed by the invention, can without the variation that creative work is expected or
Replacement, should be covered by the protection scope of the present invention.Therefore, protection scope of the present invention should be limited with claims
Subject to fixed protection scope.
Claims (7)
1. a kind of recognition methods of the driving unlawful practice based on CNN, it is characterised in that: the following steps are included:
S1, multilayer driving behavior CNN identification network is established, the CNN identification network includes three-layer network;
S2, interception monitor video image;
S3, the monitor video classmate for intercepting step S2 are sent into the first layer for the CNN identification network that step S1 is established, CNN identification
The first layer identification video image of network belongs to interior scene or car scene, and the video image for belonging to interior scene is sent into
Step S4 eliminates the image for belonging to car scene;
The second layer of S4, CNN identification network identifies the step S3 video image being sent into, and extracts in video image and drives
Member head part is distributed into step S5;
S5, CNN identification network third layer the step S4 video image being sent into is identified, judge driver behavior whether
In the presence of in violation of rules and regulations.
2. a kind of recognition methods of driving unlawful practice based on CNN according to claim 1, it is characterised in that: described
Established in step S1 multilayer driving behavior CNN identification network method the following steps are included:
S101, first layer CNN identification network is established;
S102, second layer CNN identification network is established;
S103, third layer CNN identification network is established;
S104, three layers of CNN for establishing S101, S102, S10 identify series network, form final network structure.
3. a kind of recognition methods of driving unlawful practice based on CNN according to claim 2, it is characterised in that: described
The foundation of every layer of CNN identification network the following steps are included:
S105, sample prepare, and sample is divided into training set and test set;
S106, the CNN network structure for constructing respective layer;
S107, it will be trained in the CNN network structure of training set image steps for importing S106 foundation.
S108, training are completed, and the CNN network of respective layer is obtained.
4. a kind of recognition methods of driving unlawful practice based on CNN according to claim 3, it is characterised in that: described
Step S107 further includes to training set figure before being trained in CNN network structure that training set image steps for importing S106 is established
As being converted.
5. a kind of recognition methods of driving unlawful practice based on CNN according to claim 4, it is characterised in that: described
The method of image transformation includes transverse translation, longitudinal translation, and rotation changes picture contrast, changes brightness, setting confusion region
Domain range and fog-level and adjustment noise size, and transformation quantity can all be controlled in detail to the transformation of every class.
6. a kind of recognition methods of driving unlawful practice based on CNN according to claim 3, it is characterised in that: described
The training frame that recognition training is carried out in the step S107 uses the algorithm of dynamical learning rate and sentences convergence algorithm automatically.
7. a kind of recognition methods of driving unlawful practice based on CNN according to claim 3, it is characterised in that: described
Each layer of sample prepares
First layer training: need to prepare the picture of interior picture, car scene;
Second layer training: need be in interior scene driver's number of people and be not the number of people picture;
Third layer training: the picture of violation and the not picture of violation are needed in driver's number of people.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811087757.3A CN109325439B (en) | 2018-09-18 | 2018-09-18 | CNN-based driving violation identification method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811087757.3A CN109325439B (en) | 2018-09-18 | 2018-09-18 | CNN-based driving violation identification method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109325439A true CN109325439A (en) | 2019-02-12 |
CN109325439B CN109325439B (en) | 2022-04-08 |
Family
ID=65266193
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811087757.3A Active CN109325439B (en) | 2018-09-18 | 2018-09-18 | CNN-based driving violation identification method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109325439B (en) |
Citations (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102542257A (en) * | 2011-12-20 | 2012-07-04 | 东南大学 | Driver fatigue level detection method based on video sensor |
CN102592143A (en) * | 2012-01-09 | 2012-07-18 | 清华大学 | Method for detecting phone holding violation of driver in driving |
CN106295799A (en) * | 2015-05-12 | 2017-01-04 | 核工业北京地质研究院 | A kind of implementation method of degree of depth study multilayer neural network |
CN106530730A (en) * | 2016-11-02 | 2017-03-22 | 重庆中科云丛科技有限公司 | Traffic violation detection method and system |
US9619748B1 (en) * | 2002-09-30 | 2017-04-11 | Michael Lamport Commons | Intelligent control with hierarchical stacked neural networks |
CN106682602A (en) * | 2016-12-16 | 2017-05-17 | 深圳市华尊科技股份有限公司 | Driver behavior identification method and terminal |
CN106682648A (en) * | 2017-01-20 | 2017-05-17 | 博康智能信息技术有限公司北京海淀分公司 | Test method and device of user making or receiving mobile phone behavior |
CN107491764A (en) * | 2017-08-25 | 2017-12-19 | 电子科技大学 | A kind of violation based on depth convolutional neural networks drives detection method |
CN107862296A (en) * | 2017-11-20 | 2018-03-30 | 深圳市深视创新科技有限公司 | The monitoring method and system of driving behavior, computer-readable recording medium |
US20180113458A1 (en) * | 2016-10-24 | 2018-04-26 | International Business Machines Corporation | Method of effective driving behavior extraction using deep learning |
-
2018
- 2018-09-18 CN CN201811087757.3A patent/CN109325439B/en active Active
Patent Citations (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9619748B1 (en) * | 2002-09-30 | 2017-04-11 | Michael Lamport Commons | Intelligent control with hierarchical stacked neural networks |
CN102542257A (en) * | 2011-12-20 | 2012-07-04 | 东南大学 | Driver fatigue level detection method based on video sensor |
CN102592143A (en) * | 2012-01-09 | 2012-07-18 | 清华大学 | Method for detecting phone holding violation of driver in driving |
CN106295799A (en) * | 2015-05-12 | 2017-01-04 | 核工业北京地质研究院 | A kind of implementation method of degree of depth study multilayer neural network |
US20180113458A1 (en) * | 2016-10-24 | 2018-04-26 | International Business Machines Corporation | Method of effective driving behavior extraction using deep learning |
CN106530730A (en) * | 2016-11-02 | 2017-03-22 | 重庆中科云丛科技有限公司 | Traffic violation detection method and system |
CN106682602A (en) * | 2016-12-16 | 2017-05-17 | 深圳市华尊科技股份有限公司 | Driver behavior identification method and terminal |
CN106682648A (en) * | 2017-01-20 | 2017-05-17 | 博康智能信息技术有限公司北京海淀分公司 | Test method and device of user making or receiving mobile phone behavior |
CN107491764A (en) * | 2017-08-25 | 2017-12-19 | 电子科技大学 | A kind of violation based on depth convolutional neural networks drives detection method |
CN107862296A (en) * | 2017-11-20 | 2018-03-30 | 深圳市深视创新科技有限公司 | The monitoring method and system of driving behavior, computer-readable recording medium |
Non-Patent Citations (4)
Title |
---|
MINGQI LU等: "Driver Detection Based on Deep Learning", 《JOURNAL OF PHYSICS CONFERENCE SERIES》 * |
李琪等: "电子商务优秀设计方案与分析", 《电子商务优秀设计方案与分析》 * |
王忠民等: "SF-CNN在驾驶行为识别中的应用研究", 《计算机工程与应用》 * |
耿磊: "基于多形态红外特征与深度学习的实时驾驶员疲劳检测", 《红外与激光工程》 * |
Also Published As
Publication number | Publication date |
---|---|
CN109325439B (en) | 2022-04-08 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108921227B (en) | Glaucoma medical image classification method based on capsule theory | |
CN112132817B (en) | Retina blood vessel segmentation method for fundus image based on mixed attention mechanism | |
CN108961245A (en) | Picture quality classification method based on binary channels depth parallel-convolution network | |
CN110516539A (en) | Remote sensing image building extracting method, system, storage medium and equipment based on confrontation network | |
CN109034119A (en) | A kind of method for detecting human face of the full convolutional neural networks based on optimization | |
CN110070107A (en) | Object identification method and device | |
CN110009095A (en) | Road driving area efficient dividing method based on depth characteristic compression convolutional network | |
CN108156519A (en) | Image classification method, television equipment and computer readable storage medium | |
CN109805963A (en) | The determination method and system of one Endometrium parting | |
CN107977605A (en) | Ocular Boundary characteristic extraction method and device based on deep learning | |
CN108921169B (en) | A kind of eye fundus image blood vessel segmentation method | |
CN113160062A (en) | Infrared image target detection method, device, equipment and storage medium | |
CN107958219A (en) | Image scene classification method based on multi-model and Analysis On Multi-scale Features | |
CN107633196A (en) | A kind of eyeball moving projection scheme based on convolutional neural networks | |
CN111461218A (en) | Sample data labeling system for fundus image of diabetes mellitus | |
CN113065431A (en) | Human body violation prediction method based on hidden Markov model and recurrent neural network | |
CN111625457A (en) | Virtual automatic driving test optimization method based on improved DQN algorithm | |
CN111626969A (en) | Attention mechanism-based corn disease image processing method | |
CN107067005A (en) | A kind of method and device of Sino-British mixing OCR Character segmentations | |
CN110472673A (en) | Parameter regulation means, method for processing fundus images, device, medium and equipment | |
CN109325439A (en) | A kind of recognition methods of the driving unlawful practice based on CNN | |
CN109215036A (en) | Human body segmentation's method based on convolutional neural networks | |
CN111784687A (en) | Glaucoma fundus image detection method based on deep learning | |
CN109102018B (en) | The multi-tag classification method of Gynecological secretion image | |
CN111160327A (en) | Expression recognition method based on lightweight convolutional 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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |