CN109376663A - A kind of human posture recognition method and relevant apparatus - Google Patents
A kind of human posture recognition method and relevant apparatus Download PDFInfo
- Publication number
- CN109376663A CN109376663A CN201811268507.XA CN201811268507A CN109376663A CN 109376663 A CN109376663 A CN 109376663A CN 201811268507 A CN201811268507 A CN 201811268507A CN 109376663 A CN109376663 A CN 109376663A
- Authority
- CN
- China
- Prior art keywords
- human body
- obtains
- model
- human
- neural network
- 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
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/103—Static body considered as a whole, e.g. static pedestrian or occupant recognition
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
This application discloses a kind of human posture recognition methods, comprising: carries out human skeleton feature extraction processing to the human body image of preset quantity, obtains training set data；Neural metwork training is carried out using the training set data, obtains gesture recognition model；Testing and debugging processing is carried out to the gesture recognition model according to the test set data of acquisition, obtains optimum attitude identification model；Human body image to be identified is identified according to the optimum attitude identification model, obtains recognition result.It by reducing characteristic amount using framework characteristic as characteristic, and is identified using neural network, improves the efficiency of human body attitude estimation.Disclosed herein as well is a kind of human body attitude identifying system, server and computer readable storage mediums, have the above beneficial effect.
Description
Technical field
This application involves field of computer technology, in particular to a kind of human posture recognition method, human body attitude identification system
System, server and computer readable storage medium.
Background technique
With the continuous development of computer vision technique, human body attitude estimation has been supplied in intelligent video monitoring, intelligent family
The fields such as residence, motion analysis, human-computer interaction and medical rehabilitation.Wherein, human body attitude estimation is mainly relevant by human body attitude
Image data is as input data, then is estimated by related algorithm input data to obtain the probable value of human body attitude,
It is exactly the estimated value of human body attitude.
Currently, the method for feature extraction is carried out in order to retain effective human action information to image data, it is dynamic to human body
The dimension for making to extract data is more and more.Although increasing the information content of action description, guarantee the accuracy of posture judgement,
More and more characteristic dimensions, can make redundant computation amount more and more, reduce speed when data calculate, reduce data
The real-time of calculating.Also, human action information is judged generally by the method that mode compares in the prior art, is sentenced
Disconnected accuracy rate is lower, as a result inaccurate, and human body attitude estimation procedure is caused to obtain the inefficiency of correct result.
Therefore, how to improve the efficiency of human body attitude estimation is the Important Problems of those skilled in the art's concern.
Summary of the invention
The purpose of the application is to provide a kind of human posture recognition method, human body attitude identifying system, server and meter
Calculation machine readable storage medium storing program for executing by reducing characteristic amount using framework characteristic as characteristic, and uses neural network
It is identified, improves the efficiency of human body attitude estimation.
In order to solve the above technical problems, the application provides a kind of human posture recognition method, comprising:
Human skeleton feature extraction processing is carried out to the human body image of preset quantity, obtains training set data；
Neural metwork training is carried out using the training set data, obtains gesture recognition model；
Testing and debugging processing is carried out to the gesture recognition model according to the test set data of acquisition, obtains optimum attitude knowledge
Other model；
Human body image to be identified is identified according to the optimum attitude identification model, obtains recognition result.
Optionally, human skeleton feature extraction processing is carried out to the human body image of preset quantity, obtains training set data, wrapped
It includes:
Median filtering is carried out to the human body image of the preset quantity, obtains multiple filtering human body images；
The extraction of human skeleton data is carried out to all filtering human body images, obtains multiple human skeleton pose presentations；
All human body skeleton pose images are normalized, processing is obtained into image data as the training set
Data.
Optionally, neural metwork training is carried out using the training set data, obtains gesture recognition model, comprising:
The training set data is input in convolutional neural networks and long Memory Neural Networks in short-term, is constructed to obtain
Deep neural network model；
Loss function is calculated to the deep neural network model according to propagated forward algorithm, obtains loss function value；
Judge whether the deep neural network model is convergence state according to the loss function value；
If so, using the deep neural network model as the gesture recognition model；
If it is not, the parameter of the deep neural network model is then adjusted according to back-propagation algorithm, until depth mind
It is convergence state through network model, using the deep neural network model as the gesture recognition model.
Optionally, testing and debugging processing is carried out to the gesture recognition model according to the test set data of acquisition, obtained most
Yogci state identification model, comprising:
The gesture recognition model is tested according to the test set data, obtains test result；
The gesture recognition model is adjusted according to the test result, obtains optimum attitude identification model.
Optionally, human body image to be identified is identified according to the optimum attitude identification model, obtains recognition result,
Include:
The human body image to be identified is subjected to human skeleton feature extraction processing, obtains human skeleton image to be identified；
The human skeleton image to be identified is input to convolutional neural networks to handle, obtains characteristic spectrum；
The characteristic spectrum is input to the length, and Memory Neural Networks are handled in short-term, obtain long short-term memory nerve
Web results；
By the length, Memory Neural Networks result is input to Softmax classifier and handles in short-term, obtains multiple generics
Probability；
Using all generic probability as the recognition result.
The application also provides a kind of human body attitude identifying system, includes:
Training data obtains module, carries out human skeleton feature extraction processing for the human body image to preset quantity, obtains
To training set data；
Neural metwork training module obtains gesture recognition for carrying out neural metwork training using the training set data
Model；
Neural network test module, for carrying out test tune to the gesture recognition model according to the test set data of acquisition
Whole processing obtains optimum attitude identification model；
Identification module is known for being identified according to the optimum attitude identification model to human body image to be identified
Other result.
Optionally, the training data obtains module, comprising:
Filter unit carries out median filtering for the human body image to the preset quantity, obtains multiple filtering human figures
Picture；
Skeleton data acquiring unit obtains multiple for carrying out the extraction of human skeleton data to all filtering human body images
Human skeleton pose presentation；
Processing is obtained figure for all human body skeleton pose images to be normalized by normalized unit
As data are as the training set data.
Optionally, the neural metwork training module, comprising:
Network struction unit, for the training set data to be input to convolutional neural networks and long short-term memory nerve net
In network, constructed to obtain deep neural network model；
Loss function computing unit, for calculating loss letter to the deep neural network model according to propagated forward algorithm
Number, obtains loss function value；
Judging unit is restrained, for judging whether the deep neural network model is convergence according to the loss function value
State；
Identification model acquiring unit is used for when the deep neural network model is convergence state, by the depth mind
Through network model as the gesture recognition model；
Model parameter adjusts unit, for being passed according to reversed when the deep neural network model is not convergence state
The parameter that algorithm adjusts the deep neural network model is broadcast, until the deep neural network model is convergence state, by institute
Deep neural network model is stated as the gesture recognition model.
The application also provides a kind of server, comprising:
Memory, for storing computer program；
Processor, the step of human posture recognition method as described above is realized when for executing the computer program.
The application also provides a kind of computer readable storage medium, and calculating is stored on the computer readable storage medium
The step of machine program, the computer program realizes human posture recognition method as described above when being executed by processor.
A kind of human posture recognition method provided herein, comprising: human body is carried out to the human body image of preset quantity
Framework characteristic extraction process, obtains training set data；Neural metwork training is carried out using the training set data, obtains posture knowledge
Other model；Testing and debugging processing is carried out to the gesture recognition model according to the test set data of acquisition, obtains optimum attitude knowledge
Other model；Human body image to be identified is identified according to the optimum attitude identification model, obtains recognition result.
Characteristic by using human skeleton feature as neural metwork training effectively reduces the number of characteristic
According to amount, reduce calculation amount, accelerate computational efficiency, and framework characteristic can completely reflect the characteristic of human body attitude,
It ensure that the accuracy of identification process, while being identified using the gesture recognition model that neural network model training obtains, mentioned
The high accuracy rate and precision of gesture recognition, also improves the Generalization Capability of identification model, increases human body attitude identification process
In robustness, reduce block, the influence of background, illumination, multi-angle of view and multi-angle to recognition result, improve human body appearance
The reliability of state identification.
The application also provides a kind of human body attitude identifying system, server and computer readable storage medium, have with
Upper beneficial effect, details are not described herein.
Detailed description of the invention
In order to illustrate the technical solutions in the embodiments of the present application or in the prior art more clearly, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this
The embodiment of application for those of ordinary skill in the art without creative efforts, can also basis
The attached drawing of offer obtains other attached drawings.
Fig. 1 is a kind of flow chart of human posture recognition method provided by the embodiment of the present application；
Fig. 2 is the process of human skeleton feature extracting method in human posture recognition method provided by the embodiment of the present application
Figure；
Fig. 3 is the flow chart of neural network training method in human posture recognition method provided by the embodiment of the present application；
Fig. 4 is a kind of structural schematic diagram of human body attitude identifying system provided by the embodiment of the present application.
Specific embodiment
The core of the application is to provide a kind of human posture recognition method, human body attitude identifying system, server and meter
Calculation machine readable storage medium storing program for executing by reducing characteristic amount using framework characteristic as characteristic, and uses neural network
It is identified, improves the efficiency of human body attitude estimation.
To keep the purposes, technical schemes and advantages of the embodiment of the present application clearer, below in conjunction with the embodiment of the present application
In attached drawing, the technical scheme in the embodiment of the application is clearly and completely described, it is clear that described embodiment is
Some embodiments of the present application, instead of all the embodiments.Based on the embodiment in the application, those of ordinary skill in the art
Every other embodiment obtained without making creative work, shall fall in the protection scope of this application.
In current techniques, to image data carry out feature extraction method in order to retain effective human action information, it is right
The dimension that human action extracts data is more and more.Although increasing the information content of action description, guarantee the accurate of posture judgement
Property, but more and more characteristic dimensions, redundant computation amount can be made more and more, reduce speed when data calculate,
Reduce the real-time that data calculate.Also, in the prior art generally by mode comparison method to human action information into
Row judgement, the accuracy rate of judgement is lower, as a result inaccurate, and human body attitude estimation procedure is caused to obtain the low efficiency of correct result
Under.
Therefore, the embodiment of the present application provides a kind of human posture recognition method, by using human skeleton feature as mind
Characteristic through network training effectively reduces the data volume of characteristic, reduces calculation amount, accelerates computational efficiency,
And framework characteristic can completely reflect the characteristic of human body attitude, ensure that the accuracy of identification process, while using mind
The gesture recognition model obtained through network model training is identified, is improved the accuracy rate and precision of gesture recognition, is also improved
The Generalization Capability of identification model, increases the robustness in human body attitude identification process, reduces and blocks, is background, illumination, more
The influence of visual angle and multi-angle to recognition result improves the reliability of human body attitude identification.
Referring to FIG. 1, Fig. 1 is a kind of flow chart of human posture recognition method provided by the embodiment of the present application.
This method may include:
S101 carries out human skeleton feature extraction processing to the human body image of preset quantity, obtains training set data；
This step is intended to carry out human skeleton feature extraction processing to human body image, obtains training set data.
During general neural metwork training, also needs to carry out feature extraction to original data, be closed
Suitable characteristic.Currently, during for human body attitude identification, the profile information of human body is mainly got, although obtaining
Taken complete information that can carry out relatively accurate posture judgement, but the data volume of profile information is larger, data dimension compared with
More, redundant computation amount is increasing, and the speed of training process can be made lower, equally also will affect the speed of identification process, unfavorable
In the real-time of identification process.
Therefore, human skeleton feature is mainly extracted in original human body image in this step, by multiple groups human body bone
Frame characteristic is as training set data.Since human skeleton is made of multiple bones with certain length, that is, only
The beginning and end for describing the bone can be obtained by the position of the direction or the bone of the bone in space, it is possible to reduce
The case where carrying out the data volume of feature description, and can completely showing human body attitude by framework information, guarantees feature
The accuracy of data.So selecting human skeleton characteristic as training data in the present embodiment, reduction is trained and knows
Other data volume improves the speed of training and identification in the case where keeping accuracy.
Wherein, characteristic relevant to skeleton mainly is got for human skeleton feature extraction processing, specifically,
Since the mode of description skeleton is different, the characteristic of available whole skeletons, the also spy of available part main skeleton
Data are levied, no matter special using any skeleton any one the framework characteristic acquisition methods that can also be provided using the prior art are
The method for levying data acquisition, the data volume and data dimension of the characteristic finally obtained can be reduced.
Wherein, the number of preset quantity will affect the data volume of training set data, if the effect of the more training of quantity
Can be better, the time of negligible amounts training will be shorter, and therefore, preset quantity can be 1000 quantity less in this way,
It can be 50,000, suitable quantity can also be selected according to actual applicable cases, be not specifically limited herein.
S102 carries out neural metwork training using training set data, obtains gesture recognition model；
On the basis of this step S101, this step be intended to use more than training set data carry out neural metwork training,
Obtain gesture recognition model.
Neural network training method conducted in this step can be general neural network training method, can also be
Any one neural network training method provided in the prior art is also possible to the instruction of neural network provided by following embodiment
Practice method, it is seen then that neural network training method is not unique, is not specifically limited herein.
Wherein, the gesture recognition model obtained is the model judged the characteristic that human body image extracts, can
To export appearance probability of state.For example, being mentioned now with a specific human body image according to the human skeleton feature in previous step
It takes processing to carry out feature extraction, obtains framework characteristic, then carry out identification by the gesture recognition model to export accordingly
Generic probability, if the posture classification of setting has station, sits, lies, walking, jumping this five kinds, then every kind of appearance probability of state is just exported,
The probability 80% stood, the probability 10% of seat, the probability 5% lain, the probability 5% walked, the probability 20% of jump.
S103 carries out testing and debugging processing to gesture recognition model according to the test set data of acquisition, obtains optimum attitude
Identification model；
On the basis of step S102, this step is intended to the gesture recognition mould obtained using test set data to previous step
Type optimizes, that is, testing and debugging processing, is specifically first tested, is adjusted further according to test result, obtained most
Yogci state identification model.
Wherein, test set data are the characteristic data sets of known posture, thus can by known posture and identification posture into
Row comparison, to be adjusted according to comparing result.The data volume of test set data will affect that testing and debugging is handled as a result, more
More test results are more accurate, and fewer test speed is faster, can be selected according to the actual situation.
S104 identifies human body image to be identified according to optimum attitude identification model, obtains recognition result.
On the basis of step S103, this step is intended for human body attitude identification, that is, is identified using optimum attitude
Model identifies human body image to be identified, obtains recognition result.Wherein, including to human body image to be identified human body bone is carried out
Frame feature extraction processing is obtained framework characteristic to be identified, then is carried out using optimum attitude identification model to framework characteristic to be identified
Identification, obtains recognition result.
Wherein, it carries out knowing any one recognition methods that method for distinguishing can be provided using the prior art, can also use
The recognition methods that following optinal plan provides.
Optionally, the S103 in this step may include:
Step 1, gesture recognition model is tested according to test set data, obtains test result；
Step 2, gesture recognition model is adjusted according to test result, obtains optimum attitude identification model.
This optinal plan mainly introduces a kind of specific testing and debugging processing method, first according to test set data to posture
Identification model is tested, and test result is obtained, and is adjusted further according to test result to gesture recognition model, is obtained most Yogci
State identification model.
Specifically, first passing through camera collects 300 human body images, classifies to these human body images, indicate it
Specific human body attitude, such as stand, sit, lie, walk, jump.Then it pre-processes, extracts pair to all people's body image
The characteristic answered, then these characteristics are identified by gesture recognition model, obtains recognition result, then by the identification
As a result it is compared with known classification, determines that model is that correctly, then output is corresponding right for the identification of which image
Than as a result, as test result.Finally gesture recognition model is adjusted according to the test result, obtains optimum attitude identification
Model.Wherein, it is adjusted the parameter for usually adjusting neural network.
Wherein, multiple human body images can be tested simultaneously, is adjusted according to test result, it can also will be more
Human body image obtains test result in batches and is adjusted again, it is seen then that the execution sequence of the specific steps in this optinal plan has
A variety of sequences and a variety of numbers can be selected according to practical situation, be not specifically limited herein.
Optionally, the S104 in this step may include:
Step 1, human body image to be identified is subjected to human skeleton feature extraction processing, obtains human skeleton figure to be identified
Picture；
Step 2, human skeleton image to be identified is input to convolutional neural networks to handle, obtains characteristic spectrum；
Step 3, characteristic spectrum is input to long Memory Neural Networks in short-term to handle, obtains long short-term memory nerve net
Network result；
Step 4, long Memory Neural Networks result in short-term is input to Softmax classifier to handle, obtains multiple classes
Belong to probability；
Step 5, using all generic probability as recognition result.
Wherein, the convolution kernel size in convolutional neural networks is respectively 3 × 3, and down-sampling layer is reduced using maximum pond
Convolution exports dimension, to handle in image redundant data, while can reduce motion frequency variation and cause ground to accuracy of identification
It influences, pond area size is 2 × 2.Each convolutional layer is followed by a pond layer in convolutional neural networks, and pond technology can improve
The translation invariance of algorithm, all convolution, the setting of the step-length of pond operation are all 1.
By place in CNN (convolutional neural networks) treated characteristic spectrum input LSTM (long Memory Neural Networks in short-term)
Reason, to prevent long-time data from inputting excessive the problem of causing gradient to be exploded.
Full articulamentum is set as 1 layer in this optinal plan, neuron number 1024, and using feed forward type neural network
Connection type.Classify layer using Softmax classifier, is normalized by Softmax layers, determine that the generic of human skeleton is general
Rate.
Wherein, Softmax function is as follows:
Wherein, p (Ck) indicate generic probability；ok、oiRespectively indicate the kth and i-th of element of full articulamentum output；C is indicated
Human body attitude classification sum.
The all hyperbolic tangent function tanh of the activation primitive of convolutional neural networks in this optinal plan.
Formula is as follows:
Wherein, x is input variable.
To sum up, characteristic of the present embodiment by using human skeleton feature as neural metwork training, is effectively reduced
The data volume of characteristic, reduces calculation amount, accelerates computational efficiency, and framework characteristic can completely reflect people
The characteristic of body posture ensure that the accuracy of identification process, while the gesture recognition mould obtained using neural network model training
Type is identified, is improved the accuracy rate and precision of gesture recognition, is also improved the Generalization Capability of identification model, increase human body
Robustness during gesture recognition, reduce block, the influence of background, illumination, multi-angle of view and multi-angle to recognition result,
Improve the reliability of human body attitude identification.
On the basis of a upper embodiment, the present embodiment mainly introduces a kind of extracting method of human skeleton feature.Upper one
Human skeleton feature extracting method conducted in embodiment can be using any one feature extraction provided in the prior art
Method can use the extracting method of human skeleton feature provided in this embodiment, other parts and a upper embodiment substantially phase
Together, same section can refer to a upper embodiment, and this will not be repeated here.
Referring to FIG. 2, Fig. 2 is human skeleton feature extraction in human posture recognition method provided by the embodiment of the present application
The flow chart of method.
This method may include:
S201 carries out median filtering to the human body image of preset quantity, obtains multiple filtering human body images；
This step is mainly to carry out median filtering to the human body image of preset quantity first, obtains multiple filtering human figures
Picture.The human body image for obtaining preset quantity is original image, and original image exists due to the difference of acquiring way and acquisition modes
Different degrees of clutter noise point and noise fritter, noise therein can impact the accuracy of feature extraction.Therefore, originally
Median filtering is carried out to multiple human body images of acquisition in step, to remove the clutter noise point and noise fritter in image,
Reduce noise data amount.
Wherein, the method for the median filtering of progress can select 3 × 3 filtering templates to be handled.
S202 carries out the extraction of human skeleton data to all filtering human body images, obtains multiple human skeleton pose presentations；
On the basis of previous step, this step is intended to obtaining all filtering human body images and carry out human skeleton data to mention
It takes, obtains human skeleton pose presentation.This step is exactly to analyze the human body in filtering human body image, obtains corresponding people
Body skeleton data is the equal of that corresponding skeleton pose image is partitioned into from human body image.It is conceivable that having compared people
The image of external shape wherein the data volume for describing skeleton pose also can be fewer, therefore passes through this due to being skeleton pose image
The characteristic of sample carries out the calculation amount that neural metwork training can be significantly reduced data operation, reduces computing redundancy, improves
Speed.
Wherein, any one human skeleton data that the method that human skeleton data are extracted can be provided using the prior art
The method of extraction, is not specifically limited herein.
All human body skeleton pose images are normalized in S203, and processing is obtained image data as training
Collect data.
On the basis of step S202, this step is intended to that all people's body skeleton pose image is normalized,
Multiple images data are obtained, then using all image datas as training set data.
Wherein, the figure that image data is mainly all adjusted to unified size specification is specifically normalized
Data, wherein size specification can select suitable size specification according to practical situation, can also set size specification to
256×256。
It is less that data volume can be extracted through this embodiment, and the less characteristic of computing redundancy amount improves subsequent step
Calculating speed in rapid.
On the basis of all of above embodiment, the present embodiment mainly does one specifically to the method for neural metwork training
Illustrate, the corresponding portion of other parts and all of above embodiment is substantially the same, and same section can refer to all of above implementation
Example, this will not be repeated here.
Referring to FIG. 3, Fig. 3 is neural network training method in human posture recognition method provided by the embodiment of the present application
Flow chart.
This method may include:
Training set data is input in convolutional neural networks and long Memory Neural Networks in short-term, construct by S301
To deep neural network model；
S302 calculates loss function to deep neural network model according to propagated forward algorithm, obtains loss function value；
S303 judges whether deep neural network model is convergence state according to loss function value；
S304, if so, using deep neural network model as gesture recognition model；
S305, if it is not, the parameter of deep neural network model is then adjusted according to back-propagation algorithm, until depth nerve net
Network model is convergence state, using deep neural network model as gesture recognition model.
Wherein, for loss function using entropy function is intersected, formula is as follows:
Wherein, p (x) is sample label, indicates true sample distribution state, and it is to pass through number that q (x), which is estimating for model,
According to the estimated probability being calculated, the distribution of model is indicated.
Backpropagation (BP Backpropagation) algorithm regulating networks parameter is used in the present embodiment, using boarding steps
It spends descent method (SGD Stochastic gradient descent) and carries out parameter optimization, learning rate when training is
0.001, weight attenuation coefficient is 0.0001.
It jumps out and follows it can be found that the step in the present embodiment is that circulation carries out by the S303 to S305 in this implementation
The judgement of ring is to judge whether deep neural network model is convergence state.If being unsatisfactory for S303 to continue to execute S301, directly
To S303 is met, S304 is executed, gesture recognition model is obtained.
The embodiment of the present application provides neural network training method, the gesture recognition obtained by neural network model training
Model is identified, is improved the accuracy rate and precision of gesture recognition, is also improved the Generalization Capability of identification model, increase people
Robustness during body gesture recognition, reduce block, background, illumination, multi-angle of view and multi-angle be to the shadow of recognition result
It rings, improves the reliability of human body attitude identification.
A kind of human body attitude identifying system provided by the embodiments of the present application is introduced below, a kind of people described below
Body gesture recognition system can correspond to each other reference with a kind of above-described human posture recognition method.
Referring to FIG. 4, Fig. 4 is a kind of structural schematic diagram of human body attitude identifying system provided by the embodiment of the present application.
The system may include:
Training data obtains module 100, carries out human skeleton feature extraction processing for the human body image to preset quantity,
Obtain training set data；
Neural metwork training module 200 obtains gesture recognition mould for carrying out neural metwork training using training set data
Type；
Neural network test module 300, for carrying out test tune to gesture recognition model according to the test set data of acquisition
Whole processing obtains optimum attitude identification model；
Identification module 400 is identified for being identified according to optimum attitude identification model to human body image to be identified
As a result.
Optionally, which obtains module 100, may include:
Filter unit carries out median filtering for the human body image to preset quantity, obtains multiple filtering human body images；
Skeleton data acquiring unit obtains multiple for carrying out the extraction of human skeleton data to all filtering human body images
Human skeleton pose presentation；
Processing is obtained figure for all human body skeleton pose images to be normalized by normalized unit
As data are as training set data.
Optionally, the neural metwork training module 200 may include:
Network struction unit, for training set data to be input to convolutional neural networks and long Memory Neural Networks in short-term
In, it is constructed to obtain deep neural network model；
Loss function computing unit, for calculating loss function to deep neural network model according to propagated forward algorithm,
Obtain loss function value；
Judging unit is restrained, for judging whether deep neural network model is convergence state according to loss function value；
Identification model acquiring unit is used for when deep neural network model is convergence state, by deep neural network mould
Type is as gesture recognition model；
Model parameter adjusts unit, for being calculated according to backpropagation when deep neural network model is not convergence state
Method adjusts the parameter of deep neural network model, until deep neural network model is convergence state, by deep neural network mould
Type is as gesture recognition model.
The embodiment of the present application can also provide a kind of server, comprising:
Memory, for storing computer program；
Processor realizes human posture recognition method as described above in Example when for executing the computer program
The step of.
The embodiment of the present application can also provide a kind of computer readable storage medium, which is characterized in that the computer can
It reads to be stored with computer program on storage medium, be realized when the computer program is executed by processor as described above in Example
Human posture recognition method the step of.
The computer readable storage medium may include: USB flash disk, mobile hard disk, read-only memory (Read-Only
Memory, ROM), random access memory (RandomAccess Memory, RAM), magnetic or disk etc. is various to deposit
Store up the medium of program code.
Each embodiment is described in a progressive manner in specification, the highlights of each of the examples are with other realities
The difference of example is applied, the same or similar parts in each embodiment may refer to each other.For device disclosed in embodiment
Speech, since it is corresponded to the methods disclosed in the examples, so being described relatively simple, related place is referring to method part illustration
?.
Professional further appreciates that, unit described in conjunction with the examples disclosed in the embodiments of the present disclosure
And algorithm steps, can be realized with electronic hardware, computer software, or a combination of the two, in order to clearly demonstrate hardware and
The interchangeability of software generally describes each exemplary composition and step according to function in the above description.These
Function is implemented in hardware or software actually, the specific application and design constraint depending on technical solution.Profession
Technical staff can use different methods to achieve the described function each specific application, but this realization is not answered
Think beyond scope of the present application.
The step of method described in conjunction with the examples disclosed in this document or algorithm, can directly be held with hardware, processor
The combination of capable software module or the two is implemented.Software module can be placed in random access memory (RAM), memory, read-only deposit
Reservoir (ROM), electrically programmable ROM, electrically erasable ROM, register, hard disk, moveable magnetic disc, CD-ROM or technology
In any other form of storage medium well known in field.
Above to a kind of human posture recognition method provided herein, human body attitude identifying system, server and
Computer readable storage medium is described in detail.Principle and embodiment of the specific case to the application used herein
It is expounded, the description of the example is only used to help understand the method for the present application and its core ideas.It should be pointed out that
For those skilled in the art, under the premise of not departing from the application principle, can also to the application into
Row some improvements and modifications, these improvement and modification are also fallen into the protection scope of the claim of this application.
Claims (10)
1. a kind of human posture recognition method characterized by comprising
Human skeleton feature extraction processing is carried out to the human body image of preset quantity, obtains training set data；
Neural metwork training is carried out using the training set data, obtains gesture recognition model；
Testing and debugging processing is carried out to the gesture recognition model according to the test set data of acquisition, obtains optimum attitude identification mould
Type；
Human body image to be identified is identified according to the optimum attitude identification model, obtains recognition result.
2. human posture recognition method according to claim 1, which is characterized in that carried out to the human body image of preset quantity
Human skeleton feature extraction processing, obtains training set data, comprising:
Median filtering is carried out to the human body image of the preset quantity, obtains multiple filtering human body images；
The extraction of human skeleton data is carried out to all filtering human body images, obtains multiple human skeleton pose presentations；
All human body skeleton pose images are normalized, processing is obtained into image data as the training set number
According to.
3. human posture recognition method according to claim 1, which is characterized in that carry out mind using the training set data
Through network training, gesture recognition model is obtained, comprising:
The training set data is input in convolutional neural networks and long Memory Neural Networks in short-term, is constructed to obtain depth
Neural network model；
Loss function is calculated to the deep neural network model according to propagated forward algorithm, obtains loss function value；
Judge whether the deep neural network model is convergence state according to the loss function value；
If so, using the deep neural network model as the gesture recognition model；
If it is not, the parameter of the deep neural network model is then adjusted according to back-propagation algorithm, until the depth nerve net
Network model is convergence state, using the deep neural network model as the gesture recognition model.
4. human posture recognition method according to claim 1, which is characterized in that according to the test set data of acquisition to institute
It states gesture recognition model and carries out testing and debugging processing, obtain optimum attitude identification model, comprising:
The gesture recognition model is tested according to the test set data, obtains test result；
The gesture recognition model is adjusted according to the test result, obtains optimum attitude identification model.
5. human posture recognition method according to claim 1, which is characterized in that according to the optimum attitude identification model
Human body image to be identified is identified, recognition result is obtained, comprising:
The human body image to be identified is subjected to human skeleton feature extraction processing, obtains human skeleton image to be identified；
The human skeleton image to be identified is input to convolutional neural networks to handle, obtains characteristic spectrum；
The characteristic spectrum is input to the length, and Memory Neural Networks are handled in short-term, obtain long Memory Neural Networks in short-term
As a result；
By the length, Memory Neural Networks result is input to Softmax classifier and handles in short-term, obtains multiple generic probability；
Using all generic probability as the recognition result.
6. a kind of human body attitude identifying system characterized by comprising
Training data obtains module, carries out human skeleton feature extraction processing for the human body image to preset quantity, is instructed
Practice collection data；
Neural metwork training module obtains gesture recognition model for carrying out neural metwork training using the training set data；
Neural network test module, for being carried out at testing and debugging according to the test set data of acquisition to the gesture recognition model
Reason, obtains optimum attitude identification model；
Identification module obtains identification knot for identifying according to the optimum attitude identification model to human body image to be identified
Fruit.
7. human body attitude identifying system according to claim 6, which is characterized in that the training data obtains module, packet
It includes:
Filter unit carries out median filtering for the human body image to the preset quantity, obtains multiple filtering human body images；
Skeleton data acquiring unit obtains multiple human bodies for carrying out the extraction of human skeleton data to all filtering human body images
Skeleton pose image；
Processing is obtained picture number for all human body skeleton pose images to be normalized by normalized unit
According to as the training set data.
8. human body attitude identifying system according to claim 6, which is characterized in that the neural metwork training module, packet
It includes:
Network struction unit, for the training set data to be input to convolutional neural networks and long Memory Neural Networks in short-term
In, it is constructed to obtain deep neural network model；
Loss function computing unit, for calculating loss function to the deep neural network model according to propagated forward algorithm,
Obtain loss function value；
Judging unit is restrained, for judging whether the deep neural network model is convergence shape according to the loss function value
State；
Identification model acquiring unit is used for when the deep neural network model is convergence state, by the depth nerve net
Network model is as the gesture recognition model；
Model parameter adjusts unit, for being calculated according to backpropagation when the deep neural network model is not convergence state
Method adjusts the parameter of the deep neural network model, until the deep neural network model is convergence state, by the depth
Neural network model is spent as the gesture recognition model.
9. a kind of server characterized by comprising
Memory, for storing computer program；
Processor realizes that human body attitude described in any one of claim 1 to 5 such as identifies when for executing the computer program
The step of method.
10. a kind of computer readable storage medium, which is characterized in that be stored with computer on the computer readable storage medium
Program realizes such as human body attitude identification side described in any one of claim 1 to 5 when the computer program is executed by processor
The step of method.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811268507.XA CN109376663A (en) | 2018-10-29 | 2018-10-29 | A kind of human posture recognition method and relevant apparatus |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811268507.XA CN109376663A (en) | 2018-10-29 | 2018-10-29 | A kind of human posture recognition method and relevant apparatus |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109376663A true CN109376663A (en) | 2019-02-22 |
Family
ID=65390435
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811268507.XA Pending CN109376663A (en) | 2018-10-29 | 2018-10-29 | A kind of human posture recognition method and relevant apparatus |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109376663A (en) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110135304A (en) * | 2019-04-30 | 2019-08-16 | 北京地平线机器人技术研发有限公司 | Human body method for recognizing position and attitude and device |
| CN110310350A (en) * | 2019-06-24 | 2019-10-08 | 清华大学 | Action prediction generation method and device based on animation |
| CN110490109A (en) * | 2019-08-09 | 2019-11-22 | 郑州大学 | A kind of online human body recovery action identification method based on monocular vision |
| CN110728209A (en) * | 2019-09-24 | 2020-01-24 | 腾讯科技（深圳）有限公司 | Gesture recognition method and device, electronic equipment and storage medium |
| CN110826401A (en) * | 2019-09-26 | 2020-02-21 | 广州视觉风科技有限公司 | Human body limb language identification method and system |
| CN111723729A (en) * | 2020-06-18 | 2020-09-29 | 成都颜禾曦科技有限公司 | Intelligent identification method for dog posture and behavior of surveillance video based on knowledge graph |
| CN112364922A (en) * | 2020-11-13 | 2021-02-12 | 苏州浪潮智能科技有限公司 | Method and system for predicting human skeleton motion in machine room environment |
| CN112733722A (en) * | 2021-01-11 | 2021-04-30 | 深圳力维智联技术有限公司 | Gesture recognition method, device and system and computer readable storage medium |
| CN113111939A (en) * | 2021-04-12 | 2021-07-13 | 中国人民解放军海军航空大学航空作战勤务学院 | Aircraft flight action identification method and device |
| CN113688790A (en) * | 2021-09-22 | 2021-11-23 | 武汉工程大学 | Human body action early warning method and system based on image recognition |
| CN113705542A (en) * | 2021-10-27 | 2021-11-26 | 北京理工大学 | Pedestrian behavior state identification method and system |
| WO2024036825A1 (en) * | 2022-08-16 | 2024-02-22 | 深圳先进技术研究院 | Attitude processing method, apparatus and system, and storage medium |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104615983A (en) * | 2015-01-28 | 2015-05-13 | 中国科学院自动化研究所 | Behavior identification method based on recurrent neural network and human skeleton movement sequences |
| CN105912991A (en) * | 2016-04-05 | 2016-08-31 | 湖南大学 | Behavior identification method based on 3D point cloud and key bone nodes |
| CN105930767A (en) * | 2016-04-06 | 2016-09-07 | 南京华捷艾米软件科技有限公司 | Human body skeleton-based action recognition method |
| CN106897670A (en) * | 2017-01-19 | 2017-06-27 | 南京邮电大学 | A kind of express delivery violence sorting recognition methods based on computer vision |
| CN107492121A (en) * | 2017-07-03 | 2017-12-19 | 广州新节奏智能科技股份有限公司 | A kind of two-dimension human body bone independent positioning method of monocular depth video |
| CN108537145A (en) * | 2018-03-21 | 2018-09-14 | 东北电力大学 | Human bodys' response method based on space-time skeleton character and depth belief network |
-
2018
- 2018-10-29 CN CN201811268507.XA patent/CN109376663A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104615983A (en) * | 2015-01-28 | 2015-05-13 | 中国科学院自动化研究所 | Behavior identification method based on recurrent neural network and human skeleton movement sequences |
| CN105912991A (en) * | 2016-04-05 | 2016-08-31 | 湖南大学 | Behavior identification method based on 3D point cloud and key bone nodes |
| CN105930767A (en) * | 2016-04-06 | 2016-09-07 | 南京华捷艾米软件科技有限公司 | Human body skeleton-based action recognition method |
| CN106897670A (en) * | 2017-01-19 | 2017-06-27 | 南京邮电大学 | A kind of express delivery violence sorting recognition methods based on computer vision |
| CN107492121A (en) * | 2017-07-03 | 2017-12-19 | 广州新节奏智能科技股份有限公司 | A kind of two-dimension human body bone independent positioning method of monocular depth video |
| CN108537145A (en) * | 2018-03-21 | 2018-09-14 | 东北电力大学 | Human bodys' response method based on space-time skeleton character and depth belief network |
Non-Patent Citations (8)
| Title |
|---|
| JEFF DONAHUE等: "Long-term Recurrent Convolutional Networks for Visual Recognition and Description", 《ARXIV:1411.4389V4 [CS.CV] 》 * |
| WEI SHEN等: "DeepSkeleton: Learning Multi-Task Scale-Associated Deep Side Outputs for Object Skeleton Extraction in Natural Images", 《IEEE TRANSACTIONS ON IMAGE PROCESSING》 * |
| 刘鹏等: "基于深度学习的园林智能浇灌系统", 《湖北汽车工业学院学报》 * |
| 吕耀坤: "基于卷积神经网络的实景交通标志识别", 《物联网技术》 * |
| 秦阳等: "3D CNNs与LSTMs在行为识别中的组合及其应用", 《测控技术》 * |
| 董军: "《"心迹"的计算——隐性知识的人工智能途径》", 31 December 2016 * |
| 郑攀海等: "基于TensorFlow的卷积神经网络的研究与实现", 《电子技术与工程》 * |
| 黄丁: "基于骨架特征的人体姿态识别研究", 《万方数据知识服务平台》 * |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110135304A (en) * | 2019-04-30 | 2019-08-16 | 北京地平线机器人技术研发有限公司 | Human body method for recognizing position and attitude and device |
| CN110310350A (en) * | 2019-06-24 | 2019-10-08 | 清华大学 | Action prediction generation method and device based on animation |
| CN110490109B (en) * | 2019-08-09 | 2022-03-25 | 郑州大学 | Monocular vision-based online human body rehabilitation action recognition method |
| CN110490109A (en) * | 2019-08-09 | 2019-11-22 | 郑州大学 | A kind of online human body recovery action identification method based on monocular vision |
| CN110728209A (en) * | 2019-09-24 | 2020-01-24 | 腾讯科技（深圳）有限公司 | Gesture recognition method and device, electronic equipment and storage medium |
| CN110728209B (en) * | 2019-09-24 | 2023-08-08 | 腾讯科技（深圳）有限公司 | Gesture recognition method and device, electronic equipment and storage medium |
| CN110826401A (en) * | 2019-09-26 | 2020-02-21 | 广州视觉风科技有限公司 | Human body limb language identification method and system |
| CN110826401B (en) * | 2019-09-26 | 2023-12-26 | 广州视觉风科技有限公司 | Human body limb language identification method and system |
| CN111723729B (en) * | 2020-06-18 | 2022-08-05 | 四川千图禾科技有限公司 | Intelligent identification method for dog posture and behavior of surveillance video based on knowledge graph |
| CN111723729A (en) * | 2020-06-18 | 2020-09-29 | 成都颜禾曦科技有限公司 | Intelligent identification method for dog posture and behavior of surveillance video based on knowledge graph |
| CN112364922B (en) * | 2020-11-13 | 2023-01-10 | 苏州浪潮智能科技有限公司 | Method and system for predicting human skeleton motion in machine room environment |
| CN112364922A (en) * | 2020-11-13 | 2021-02-12 | 苏州浪潮智能科技有限公司 | Method and system for predicting human skeleton motion in machine room environment |
| CN112733722A (en) * | 2021-01-11 | 2021-04-30 | 深圳力维智联技术有限公司 | Gesture recognition method, device and system and computer readable storage medium |
| CN113111939A (en) * | 2021-04-12 | 2021-07-13 | 中国人民解放军海军航空大学航空作战勤务学院 | Aircraft flight action identification method and device |
| CN113688790A (en) * | 2021-09-22 | 2021-11-23 | 武汉工程大学 | Human body action early warning method and system based on image recognition |
| CN113705542A (en) * | 2021-10-27 | 2021-11-26 | 北京理工大学 | Pedestrian behavior state identification method and system |
| WO2024036825A1 (en) * | 2022-08-16 | 2024-02-22 | 深圳先进技术研究院 | Attitude processing method, apparatus and system, and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109376663A (en) | A kind of human posture recognition method and relevant apparatus | |
| CN109948647B (en) | Electrocardiogram classification method and system based on depth residual error network | |
| CN106778854B (en) | Behavior identification method based on trajectory and convolutional neural network feature extraction | |
| CN106709461B (en) | Activity recognition method and device based on video | |
| Liu et al. | Learning spatio-temporal representations for action recognition: A genetic programming approach | |
| US20220004744A1 (en) | Human posture detection method and apparatus, device and storage medium | |
| CN109215013A (en) | Automatic stone age prediction technique, system, computer equipment and storage medium | |
| CN109919928A (en) | Detection method, device and the storage medium of medical image | |
| CN109785928A (en) | Diagnosis and treatment proposal recommending method, device and storage medium | |
| CN116051574A (en) | Semi-supervised segmentation model construction and image analysis method, device and system | |
| CN110084313A (en) | A method of generating object detection model | |
| CN110580445A (en) | Face key point detection method based on GIoU and weighted NMS improvement | |
| CN111401106B (en) | Behavior identification method, device and equipment | |
| CN110543895B (en) | Image classification method based on VGGNet and ResNet | |
| CN111832516A (en) | Video behavior identification method based on unsupervised video representation learning | |
| CN109583331B (en) | Deep learning-based accurate positioning method for positions of wrist vein and mouth of person | |
| CN112836602B (en) | Behavior recognition method, device, equipment and medium based on space-time feature fusion | |
| CN109284761A (en) | A kind of image characteristic extracting method, device, equipment and readable storage medium storing program for executing | |
| CN113569805A (en) | Action recognition method and device, electronic equipment and storage medium | |
| CN109766873A (en) | pedestrian re-identification method based on hybrid deformable convolution | |
| CN111428761A (en) | Image feature visualization method, image feature visualization device and electronic equipment | |
| CN112949654A (en) | Image detection method and related device and equipment | |
| CN111429414B (en) | Artificial intelligence-based focus image sample determination method and related device | |
| WO2021179198A1 (en) | Image feature visualization method, image feature visualization apparatus, and electronic device | |
| CN109284700A (en) | The method of multiple Face datections, storage medium, equipment and system in image |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190222 |