CN110163175A - A kind of gait recognition method and system based on improvement VGG-16 network - Google Patents
A kind of gait recognition method and system based on improvement VGG-16 network Download PDFInfo
- Publication number
- CN110163175A CN110163175A CN201910449120.2A CN201910449120A CN110163175A CN 110163175 A CN110163175 A CN 110163175A CN 201910449120 A CN201910449120 A CN 201910449120A CN 110163175 A CN110163175 A CN 110163175A
- Authority
- CN
- China
- Prior art keywords
- network
- obj
- recognition
- gait
- module
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of based on the gait recognition method and system that improve VGG-16 network, comprising steps of S1, handling original pedestrian's gait data collection based on improving VGG-16 network, extracts static nature and behavioral characteristics；S2, the training static nature and behavioral characteristics obtain static identification network, Dynamic Recognition network, Fusion Features module and comprehensive identification module；S3, the gait in image or video is identified based on the static identification network, Dynamic Recognition network, Fusion Features module and comprehensive identification module.The present invention has the following advantages compared with prior art: 1. present invention combine convolutional neural networks to carry out pedestrian's Gait Recognition and the accuracy and real-time to pedestrian's Gait Recognition are substantially improved using the outstanding ability in feature extraction of convolutional neural networks and computing capability.2. the static nature and behavioral characteristics when the present invention walks in conjunction with pedestrian carry out Gait Recognition, the negative effect as caused by angle, clothing and distance can be effectively removed.
Description
Technical field
The invention belongs to the object detection fields of image procossing, and in particular to Gait Recognition, specifically, i.e., based on improvement
Pedestrian's gait recognition method of VGG-16 network.
Background technique
Gait Recognition is a kind of biometrics identification technology, the gesture recognition row that can be walked according to pedestrian in video sequence
Personal part；Compared with the biological identification technologies such as traditional fingerprint, face, iris, there is contactless identification, be easy to hiding, acquire
Advantages, the object recognition task being particularly suitable under remote scene such as facilitate.Gait Recognition always is computer vision heat
One of the research direction of door.It is excellent in recent years due to deep learning in field of image processing, the pedestrian based on deep learning
Gait Recognition also flourishes.
Algorithm for gait recognition, which can be mainly divided into, to be statisticallyd analyze based on " non-model " and based on " model learning " two major classes.Base
It is usually to pass through statistics human motion timing, empty template, speed when establishing in the Algorithm for gait recognition of non-modeling statistics analysis
Or the data such as area change, Gait Recognition is carried out using motion state is described based on the method for shape features of statistics, so
Gait Recognition based on non-model is also known as based on appearance or based on the Algorithm for gait recognition at visual angle.Step based on model learning
State recognizer usually learns in parameter information using in model and expresses the characteristic energy of gait image, recycles this
A little feature vector training classifiers, reach Gait Recognition purpose with this.
As convolutional neural networks and deep learning are extracting the outstanding representation in feature, conventional method manual extraction is utilized
The drawbacks of Gait Recognition feature, is more obvious, no matter being all lacking in accuracy or validity.Convolutional neural networks are
A kind of common deep learning frame, with application of the deep learning in terms of image procossing and pattern-recognition, convolutional Neural net
The research and application of network are also increasingly valued by people.
For example, the patent of invention of Publication No. CN107958221A discloses a kind of human body fortune based on convolutional neural networks
Dynamic gait classification method, comprising the following steps: step 1, acquire 3-axis acceleration and three axis of the human body under different motion gait
Angular velocity information obtains the sample data of six parameters, for the sample data of each parameter, calculates its step under each gait
State statistics feature, and obtain the gait statistics feature difference of each parameter；It chooses gait statistics feature difference and is greater than setting
The parameter of threshold value is sensitive parameter, is inputted using the data of sensitive parameter as convolutional neural networks；Using human motion gait as
Convolutional neural networks output；Step 2, the sample data for the sensitive parameter establishing convolutional neural networks, and being obtained using step 1 with
And gait is trained convolutional neural networks, obtains trained convolutional neural networks；Step 3, sensitive parameter is acquired in real time
Data and be input in trained convolutional neural networks, obtained output result is current human motion gait, real
Existing human motion gait classification.
However, the existing Gait Recognition based on convolutional neural networks is all whole using all gait features as one
Body has that accuracy of identification caused by single features is low without distinguishing behavioral characteristics and static nature.
Therefore in view of the drawbacks of the prior art, how to realize that high-precision pedestrian's Gait Recognition is that this field is urgently to be resolved
Problem.
Summary of the invention
It is a kind of based on the gait knowledge for improving VGG-16 network the purpose of the present invention is in view of the drawbacks of the prior art, providing
Other method and system.Static nature extraction and behavioral characteristics extraction in pedestrian's Gait Recognition are carried out using convolutional neural networks,
It realizes the feature extraction of high accuracy and validity, and the feature extracted is respectively used to pedestrian's Gait Recognition；Utilize matching
Fusion method based on Adding law in layer multi-biological characteristic blending algorithm realizes Fusion Features, make full use of behavioral characteristics and
Static nature realizes complete pedestrian's Gait Recognition.
In order to achieve the goal above, the invention adopts the following technical scheme:
A kind of gait recognition method based on improvement VGG-16 network, comprising steps of
S1, original pedestrian's gait data collection is handled based on improvement VGG-16 network, extracts static nature and dynamic
Feature；
S2, the training static nature and behavioral characteristics obtain static identification network, Dynamic Recognition network, Fusion Features
Module and comprehensive identification module；
S3, figure is identified based on the static identification network, Dynamic Recognition network, Fusion Features module and comprehensive identification module
Gait in picture or video.
Further, the improvement VGG-16 network specifically:
Using VGG-16 network as basic network, residual error network module is added.
Further, the step S2 specifically:
Static state identification network carries out Gait Recognition using the static nature；Dynamic Recognition network utilizes the behavioral characteristics
Carry out Gait Recognition.
Further, the step S3 specifically:
S3.1, input original image or video identify network and the Dynamic Recognition network to the static state；The static state
Identification network and the Dynamic Recognition network carry out pedestrian's Gait Recognition respectively, and export objective degrees of confidence Scl, Tcl and identification
As a result Sobj, Tobj are to the Fusion Features module；
S3.2, Fusion Features module utilize the fusion side based on Adding law in matching layer multi-biological characteristic blending algorithm
Method realizes Fusion Features, and objective degrees of confidence Scl, Tcl and recognition result Sobj, Tobj are carried out Fusion Features, obtains fusion mesh
Mark confidence level Fcl；Scl, Tcl, Fcl, Sobj and Tobj are input to the comprehensive identification module；
S3.3, comprehensive identification module are based on Scl, Tcl, Fcl, Sobj and Tobj and are integrated, and export final recognition result
Fobj。
Further, the step S3.3 specifically:
When Sobj and Tobj is identical, recognition result Fobj is equal to Sobj, if Fcl is greater than objective degrees of confidence threshold value Ocl-1,
Recognition result Fobj is exported, it is on the contrary then do not export recognition result；
When Sobj with Tobj difference, if Scl, Tcl are all larger than Ocl-2, and Scl > Tcl, output recognition result Fobj etc.
In Sobj, otherwise output recognition result Fobj is equal to Tobj；If Scl, Tcl are respectively less than Ocl-2, recognition result is not exported；If Scl
And Tcl one of them be greater than Ocl-2, another is less than Ocl-2, then Fobj is equal to the confidence level belonging network greater than Ocl-2
Recognition result exports Fobj.
A kind of Gait Recognition system based on improvement VGG-16 network, comprising:
Characteristic extracting module is extracted for being handled based on improvement VGG-16 network original pedestrian's gait data collection
Static nature and behavioral characteristics；
Training module obtains static identification network, Dynamic Recognition net for training the static nature and behavioral characteristics
Network, Fusion Features module and comprehensive identification module；
Identification module, for based on the static identification network, Dynamic Recognition network, Fusion Features module and comprehensive identification
Module identifies the gait in image or video.
Further, the improvement VGG-16 network specifically:
Using VGG-16 network as basic network, residual error network module is added.
Further, the training module includes:
Static state identification network carries out Gait Recognition using the static nature；Dynamic Recognition network utilizes the behavioral characteristics
Carry out Gait Recognition.
Further, the identification module includes:
Static state identification network obtains objective degrees of confidence for identifying to original image Picture or video Video
Scl and recognition result Sobj；
Dynamic Recognition network obtains objective degrees of confidence for identifying to original image Picture or video Video
Tcl and recognition result Tobj；
Fusion Features module, for utilizing the fusion side based on Adding law in matching layer multi-biological characteristic blending algorithm
Method realizes Fusion Features, and objective degrees of confidence Scl, Tcl and recognition result Sobj, Tobj are carried out Fusion Features, obtains fusion mesh
Mark confidence level Fcl；Scl, Tcl, Fcl, Sobj and Tobj are input to comprehensive identification module；
Comprehensive identification module exports final recognition result for being integrated based on Scl, Tcl, Fcl, Sobj and Tobj
Fobj。
Further, the comprehensive identification module specifically includes:
When Sobj and Tobj is identical, recognition result Fobj is equal to Sobj, if Fcl is greater than objective degrees of confidence threshold value Ocl-1,
Recognition result Fobj is exported, it is on the contrary then do not export recognition result；
When Sobj with Tobj difference, if Scl, Tcl are all larger than Ocl-2, and Scl > Tcl, output recognition result Fobj etc.
In Sobj, otherwise output recognition result Fobj is equal to Tobj；If Scl, Tcl are respectively less than Ocl-2, recognition result is not exported；If Scl
And Tcl one of them be greater than Ocl-2, another is less than Ocl-2, then Fobj is equal to the confidence level belonging network greater than Ocl-2
Recognition result exports Fobj.
The present invention has the following advantages compared with prior art:
1. the present invention is obtained in more profound feature by increasing extraction of the residual error network module raising to feature
Hold, to improve the accuracy of pedestrian's Gait Recognition.
2. the present invention is in order to make full use of the feature extracted, in the training process by the feature of shallow hierarchy and profound level
Feature is merged, and each feature obtained in training process is made full use of, to obtain optimal network model.
3. the present invention increases Fusion Features module to make full use of the feature that can be extracted in pedestrian's gait.It compares
In carrying out pedestrian's Gait Recognition using single static nature or single behavioral characteristics, Fusion Features module can be made full use of
All information obtained promote accuracy of identification to make up the possible deficiency of single features.
Detailed description of the invention
Fig. 1 is a kind of gait recognition method flow chart based on improvement VGG-16 network that embodiment one provides；
Fig. 2 is residual error network module (Residual) structure chart；
Fig. 3 is the structure chart of static identification network and Dynamic Recognition network；
Fig. 4 embodiment one provides a kind of based on the Gait Recognition system construction drawing for improving VGG-16 network.
Specific embodiment
Illustrate embodiments of the present invention below by way of specific specific example, those skilled in the art can be by this specification
Other advantages and efficacy of the present invention can be easily understood for disclosed content.The present invention can also pass through in addition different specific realities
The mode of applying is embodied or practiced, the various details in this specification can also based on different viewpoints and application, without departing from
Various modifications or alterations are carried out under spirit of the invention.It should be noted that in the absence of conflict, following embodiment and implementation
Feature in example can be combined with each other.
It should be noted that illustrating the basic structure that only the invention is illustrated in a schematic way provided in following embodiment
Think, only shown in schema then with related component in the present invention rather than component count, shape and size when according to actual implementation
Draw, when actual implementation kenel, quantity and the ratio of each component can arbitrarily change for one kind, and its assembly layout kenel
It is likely more complexity.
The present invention will be further explained below with reference to the attached drawings and specific examples, but not as the limitation of the invention.
Embodiment one
As shown in Figure 1, the present embodiment proposes a kind of gait recognition method based on improvement VGG-16 network, comprising:
S1, original pedestrian's gait data collection is handled based on improvement VGG-16 network, extracts static nature and dynamic
Feature；
Specifically, the present invention handles original pedestrian's gait data collection, extracts each data for obtaining meeting method needs, whole
Manage into new gait data collection.The research of gait Recognition technology originates from the 1990s earliest.Southern Florida is state big
The CASIA database of the USF database and Institute of Automation Research of CAS's publication of learning (FSU) publication is generally acknowledged in the world
The fairly perfect most popular gait data library in gait data library and Research on Gait Recognition field.But due to CASIA number
Untreated base video image data is only provided according to library, so needing to be further processed just obtain meeting the present invention
It is required that gait data collection.On the basis of existing CASIA database, processing obtains can be used for improving the present invention present invention thus
The VOC-CASIA data set of algorithm training.The present invention to the processing of original pedestrian's gait data collection include extract static nature and
Behavioral characteristics.
VGG full name is that Visual Geometry Group belongs to Scientific Engineering system, Oxford University, issued it is some column with
VGG beginning convolutional network model, can apply recognition of face, in terms of, respectively from VGG16~VGG19.
The original intention of VGG research convolutional network depth is intended to make convolutional network depth is how to influence large-scale image classification and identification clear
Precision and accuracy rate, be initially that be known as very deep convolutional network full name be (GG-Very-Deep-16CNN) to VGG-16,
VGG is deepening the network number of plies simultaneously in order to avoid parameter is excessive, all uses the small convolution kernel of 3x3, convolutional layer step-length in all layers
It is arranged to 1.The input of VGG is arranged to the RGB image of 224x244 size, calculates on training set image all images
RGB mean value, then using image as incoming VGG convolutional network is inputted, using the filter of 3x3 or 1x1, convolution step-length is consolidated
Fixed 1.VGG is complete, and articulamentum haves three layers, can be least from VGG11~VGG19 according to convolutional layer+full articulamentum total number difference
VGG11 has 8 convolutional layers and 3 full articulamentums, and most VGG19 has 16 convolutional layer+3 full articulamentums, furthermore VGG network
Be not in each convolutional layer followed by a upper pond layer, or total 5 pond layers, be distributed in different convolutional layers it
Under.
VGG-16, input layer 224*224*3, by two layers of identical convolution, convolution filter is 3*3, stride 1,
Filter number is 64, then passes through one layer of pooling.Then according to identical mode, make wide and height smaller and smaller, and port number
By increasing again, until 512.Finally with two layers of identical full connection plus a softmax.The performance and VGG-19 difference of VGG-16 is not
More, therefore, the present invention uses VGG-16.
In feature extraction, carrying out feature extraction using convolutional layer has very outstanding performance.But with the network number of plies
Deepen, carrying out feature extraction using single convolutional layer will appear the case where network is degenerated.Therefore, as shown in Fig. 2, net of the invention
Network deepens the depth of basic network by addition residual error network module using VGG-16 network as basic network, improves feature and mentions
The function of taking and identify.The essence of residual error network is to be changed to even add process by the way that the company of feature extraction is multiplied process, is being mentioned
Further feature extraction is carried out on the basis of the feature got, so that the purpose that network is infinitely deepened on realization theory, avoids
There is network degenerate case.1 residual error network module is connected in series by 3 different convolutional layers of core scale, between convolutional layer
Between be inserted with batch normalization layer and amendment linear unit.
S2, the training static nature and behavioral characteristics obtain static identification network, Dynamic Recognition network, Fusion Features
Module and comprehensive identification module；
Specifically, the training that pedestrian's Gait Recognition network is carried out using Pytorch deep learning frame, obtains network mould
Type.Wherein, network model includes 2 sub-networks, Fusion Features module and comprehensive identification module.
2 sub-networks include static identification network S and Dynamic Recognition network T.Static state identification network utilizes pedestrian's walking
Static nature carries out Gait Recognition；Dynamic Recognition network carries out Gait Recognition using the behavioral characteristics that pedestrian walks.Static state identification
Network S and Dynamic Recognition network T is to be based on identical network structure, but utilize static nature and behavioral characteristics in the training process
It is trained to obtain respectively.Preceding 5 convolutional layers for retaining VGG-16 network, are revised as 2 for the 6th layer and the 7th layer full articulamentum
Residual error network, is revised as residual error network for the 8th layer of convolutional layer, and the 9th layer, the 10th layer and 11th layer convolutional layer are all revised as 1 × 1
Convolutional layer removes the 12nd layer to 16 layers convolutional layer.Network concrete outcome is as shown in Figure 3.
2 sub-networks are subjected to information input that Gait Recognition obtains to Fusion Features module respectively, carry out Fusion Features,
To improve final Gait Recognition accuracy.The information input that Fusion Features module is obtained is to comprehensive identification module, and output is most
Whole pedestrian's Gait Recognition result.
S3, figure is identified based on the static identification network, Dynamic Recognition network, Fusion Features module and comprehensive identification module
Gait in picture or video.
Based on trained pedestrian's Gait Recognition network model, input original image or video, image after being identified or
Video.Specifically:
S3.1, input original image Picture or video Video identify network S and Dynamic Recognition network T to static state；It is quiet
State identification network S and Dynamic Recognition network T carries out pedestrian's Gait Recognition respectively, and exports objective degrees of confidence Scl、TclIt is tied with identification
Fruit Sobj、TobjTo Fusion Features module.
Static state identification network S exports the objective degrees of confidence S identified using static natureclWith recognition result Sobj, dynamic
Identification network T exports the objective degrees of confidence T identified using behavioral characteristicsclWith recognition result Tobj。
S3.2, Fusion Features module utilize the fusion side based on Adding law in matching layer multi-biological characteristic blending algorithm
Method realizes Fusion Features, by objective degrees of confidence Scl、TclWith recognition result Sobj、TobjFusion Features are carried out, fusion target is obtained and sets
Reliability Fcl；By Scl、Tcl、Fcl、SobjAnd TobjIt is input to comprehensive identification module.
Fusion Features module is real using the fusion method based on Adding law in matching layer multi-biological characteristic blending algorithm
Existing Fusion Features.By the objective degrees of confidence S of static identification network S outputclWith the objective degrees of confidence of Dynamic Recognition network T output
TclAddition obtains fusion objective degrees of confidence Fcl, by Scl、Tcl、Fcl、SobjAnd TobjIt is input to comprehensive identification module.
S3.3, comprehensive identification module are based on Scl、Tcl、Fcl、SobjAnd TobjIt is integrated, exports final recognition result Fobj。
Specifically, as the recognition result S of static identification network S outputobjWith the recognition result of Dynamic Recognition network T output
TobjWhen identical, recognition result FobjEqual to SobjIf merging objective degrees of confidence FclGreater than objective degrees of confidence threshold value Ocl-1, output knowledge
Other result Fobj, on the contrary then do not export recognition result；As the recognition result S of static identification network S outputobjWith Dynamic Recognition network
The recognition result T of T outputobjWhen different, by objective degrees of confidence SclWith objective degrees of confidence TclRespectively with objective degrees of confidence threshold value Ocl-2
It is compared, if being all larger than Ocl-2If Scl>Tcl, export recognition result FobjEqual to Sobj, export recognition result Fobj, on the contrary then defeated
Recognition result F outobjEqual to Tobj, export recognition result Fobj；If respectively less than Ocl-2, do not export recognition result；If one of mesh
It marks confidence level and is greater than Ocl-2, another is less than Ocl-2, then FobjEqual to greater than Ocl-2Confidence level belonging network recognition result, it is defeated
F outobj。
Embodiment two
As shown in figure 4, the present embodiment proposes a kind of Gait Recognition system based on improvement VGG-16 network, comprising:
Characteristic extracting module is extracted for being handled based on improvement VGG-16 network original pedestrian's gait data collection
Static nature and behavioral characteristics；
Specifically, the present invention handles original pedestrian's gait data collection, extracts each data for obtaining meeting method needs, whole
Manage into new gait data collection.The research of gait Recognition technology originates from the 1990s earliest.Southern Florida is state big
The CASIA database of the USF database and Institute of Automation Research of CAS's publication of learning (FSU) publication is generally acknowledged in the world
The fairly perfect most popular gait data library in gait data library and Research on Gait Recognition field.But due to CASIA number
Untreated base video image data is only provided according to library, so needing to be further processed just obtain meeting the present invention
It is required that gait data collection.On the basis of existing CASIA database, processing obtains can be used for improving the present invention present invention thus
The VOC-CASIA data set of algorithm training.The present invention to the processing of original pedestrian's gait data collection include extract static nature and
Behavioral characteristics.
VGG full name is that Visual Geometry Group belongs to Scientific Engineering system, Oxford University, issued it is some column with
VGG beginning convolutional network model, can apply recognition of face, in terms of, respectively from VGG16~VGG19.
The original intention of VGG research convolutional network depth is intended to make convolutional network depth is how to influence large-scale image classification and identification clear
Precision and accuracy rate, be initially that be known as very deep convolutional network full name be (GG-Very-Deep-16CNN) to VGG-16,
VGG is deepening the network number of plies simultaneously in order to avoid parameter is excessive, all uses the small convolution kernel of 3x3, convolutional layer step-length in all layers
It is arranged to 1.The input of VGG is arranged to the RGB image of 224x244 size, calculates on training set image all images
RGB mean value, then using image as incoming VGG convolutional network is inputted, using the filter of 3x3 or 1x1, convolution step-length is consolidated
Fixed 1.VGG is complete, and articulamentum haves three layers, can be least from VGG11~VGG19 according to convolutional layer+full articulamentum total number difference
VGG11 has 8 convolutional layers and 3 full articulamentums, and most VGG19 has 16 convolutional layer+3 full articulamentums, furthermore VGG network
Be not in each convolutional layer followed by a upper pond layer, or total 5 pond layers, be distributed in different convolutional layers it
Under.
VGG-16, input layer 224*224*3, by two layers of identical convolution, convolution filter is 3*3, stride 1,
Filter number is 64, then passes through one layer of pooling.Then according to identical mode, make wide and height smaller and smaller, and port number
By increasing again, until 512.Finally with two layers of identical full connection plus a softmax.The performance and VGG-19 difference of VGG-16 is not
More, therefore, the present invention uses VGG-16.
In feature extraction, carrying out feature extraction using convolutional layer has very outstanding performance.But with the network number of plies
Deepen, carrying out feature extraction using single convolutional layer will appear the case where network is degenerated.Therefore, as shown in Fig. 2, net of the invention
Network deepens the depth of basic network by addition residual error network module using VGG-16 network as basic network, improves feature and mentions
The function of taking and identify.The essence of residual error network is to be changed to even add process by the way that the company of feature extraction is multiplied process, is being mentioned
Further feature extraction is carried out on the basis of the feature got, so that the purpose that network is infinitely deepened on realization theory, avoids
There is network degenerate case.1 residual error network module is connected in series by 3 different convolutional layers of core scale, between convolutional layer
Between be inserted with batch normalization layer and amendment linear unit.
Training module obtains static identification network, Dynamic Recognition net for training the static nature and behavioral characteristics
Network, Fusion Features module and comprehensive identification module；
Specifically, the training that pedestrian's Gait Recognition network is carried out using Pytorch deep learning frame, obtains network mould
Type.Wherein, network model includes 2 sub-networks, Fusion Features module and comprehensive identification module.
2 sub-networks include static identification network S and Dynamic Recognition network T.Static state identification network utilizes pedestrian's walking
Static nature carries out Gait Recognition；Dynamic Recognition network carries out Gait Recognition using the behavioral characteristics that pedestrian walks.Static state identification
Network S and Dynamic Recognition network T is to be based on identical network structure, but utilize static nature and behavioral characteristics in the training process
It is trained to obtain respectively.Preceding 5 convolutional layers for retaining VGG-16 network, are revised as 2 for the 6th layer and the 7th layer full articulamentum
Residual error network, is revised as residual error network for the 8th layer of convolutional layer, and the 9th layer, the 10th layer and 11th layer convolutional layer are all revised as 1 × 1
Convolutional layer removes the 12nd layer to 16 layers convolutional layer.Network concrete outcome is as shown in Figure 3.
2 sub-networks are subjected to information input that Gait Recognition obtains to Fusion Features module respectively, carry out Fusion Features,
To improve final Gait Recognition accuracy.The information input that Fusion Features module is obtained is to comprehensive identification module, and output is most
Whole pedestrian's Gait Recognition result.
Identification module, for based on the static identification network, Dynamic Recognition network, Fusion Features module and comprehensive identification
Module identifies the gait in image or video.
Based on trained pedestrian's Gait Recognition network model, input original image or video, image after being identified or
Video.It specifically includes:
Static state identification network obtains objective degrees of confidence for identifying to original image Picture or video Video
SclWith recognition result Sobj；
Dynamic Recognition network obtains objective degrees of confidence for identifying to original image Picture or video Video
TclWith recognition result Tobj；
Static state identification network S exports the objective degrees of confidence S identified using static natureclWith recognition result Sobj, dynamic
Identification network T exports the objective degrees of confidence T identified using behavioral characteristicsclWith recognition result Tobj。
Fusion Features module, for utilizing the fusion side based on Adding law in matching layer multi-biological characteristic blending algorithm
Method realizes Fusion Features, by objective degrees of confidence Scl、TclWith recognition result Sobj、TobjFusion Features are carried out, fusion target is obtained and sets
Reliability Fcl；By Scl、Tcl、Fcl、SobjAnd TobjIt is input to comprehensive identification module.
Fusion Features module is real using the fusion method based on Adding law in matching layer multi-biological characteristic blending algorithm
Existing Fusion Features.By the objective degrees of confidence S of static identification network S outputclWith the objective degrees of confidence of Dynamic Recognition network T output
TclAddition obtains fusion objective degrees of confidence Fcl, by Scl、Tcl、Fcl、SobjAnd TobjIt is input to comprehensive identification module.
Comprehensive identification module, for being based on Scl、Tcl、Fcl、SobjAnd TobjIt is integrated, exports final recognition result Fobj。
Specifically, as the recognition result S of static identification network S outputobjWith the recognition result of Dynamic Recognition network T output
TobjWhen identical, recognition result FobjEqual to SobjIf merging objective degrees of confidence FclGreater than objective degrees of confidence threshold value Ocl-1, output knowledge
Other result Fobj, on the contrary then do not export recognition result；As the recognition result S of static identification network S outputobjWith Dynamic Recognition network
The recognition result T of T outputobjWhen different, by objective degrees of confidence SclWith objective degrees of confidence TclRespectively with objective degrees of confidence threshold value Ocl-2
It is compared, if being all larger than Ocl-2If Scl>Tcl, export recognition result FobjEqual to Sobj, export recognition result Fobj, on the contrary then defeated
Recognition result F outobjEqual to Tobj, export recognition result Fobj；If respectively less than Ocl-2, do not export recognition result；If one of mesh
It marks confidence level and is greater than Ocl-2, another is less than Ocl-2, then FobjEqual to greater than Ocl-2Confidence level belonging network recognition result, it is defeated
F outobj。
The invention discloses a kind of based on the pedestrian's gait recognition method for improving VGG-16 network, is related to the knowledge of pedestrian's gait
Not, so that the accuracy and real-time of Gait Recognition have more is obviously improved.Its realization process is: utilizing the Chinese Academy of Sciences
Automation research CASIA database make to obtain and meet the data set that the present invention needs, carried out respectively using data set quiet
State identification network training and Dynamic Recognition network training obtain the sub-network model of pedestrian's Gait Recognition network, more using matching layer
Fusion method based on Adding law in biological characteristic blending algorithm carries out Fusion Features, make full use of static identification network and
The output of Dynamic Recognition network is as a result, realize pedestrian's Gait Recognition.This method more effective solution is due to feature unicity
With caused by angle, clothing and distance negatively affect cause accuracy of identification lower problem, while utilize convolutional Neural net
Network and deep learning improve the identification and extraction of feature, realize pedestrian's Gait Recognition problem of high accuracy and real-time.
Note that the above is only a better embodiment of the present invention and the applied technical principle.It will be appreciated by those skilled in the art that
The invention is not limited to the specific embodiments described herein, be able to carry out for a person skilled in the art it is various it is apparent variation,
It readjusts and substitutes without departing from protection scope of the present invention.Therefore, although being carried out by above embodiments to the present invention
It is described in further detail, but the present invention is not limited to the above embodiments only, without departing from the inventive concept, also
It may include more other equivalent embodiments, and the scope of the invention is determined by the scope of the appended claims.
Claims (10)
1. a kind of based on the gait recognition method for improving VGG-16 network, which is characterized in that comprising steps of
S1, original pedestrian's gait data collection is handled based on improvement VGG-16 network, extracts static nature and behavioral characteristics；
S2, the training static nature and behavioral characteristics obtain static identification network, Dynamic Recognition network, Fusion Features module
With comprehensive identification module；
S3, based on the static identification network, Dynamic Recognition network, Fusion Features module and comprehensive identification module identification image or
Gait in video.
2. gait recognition method according to claim 1, which is characterized in that the improvement VGG-16 network specifically:
Using VGG-16 network as basic network, residual error network module is added.
3. gait recognition method according to claim 2, which is characterized in that the step S2 specifically: static state identification net
Network carries out Gait Recognition using the static nature；Dynamic Recognition network carries out Gait Recognition using the behavioral characteristics.
4. gait recognition method according to claim 2, which is characterized in that the step S3 specifically:
S3.1, input original image or video identify network and the Dynamic Recognition network to the static state；The static identification
Network and the Dynamic Recognition network carry out pedestrian's Gait Recognition respectively, and export objective degrees of confidence Scl、TclAnd recognition result
Sobj、TobjTo the Fusion Features module；
S3.2, Fusion Features module are real using the fusion method based on Adding law in matching layer multi-biological characteristic blending algorithm
Existing Fusion Features, by objective degrees of confidence Scl、TclWith recognition result Sobj、TobjFusion Features are carried out, fusion objective degrees of confidence is obtained
Fcl；By Scl、Tcl、Fcl、SobjAnd TobjIt is input to the comprehensive identification module；
S3.3, comprehensive identification module are based on Scl、Tcl、Fcl、SobjAnd TobjIt is integrated, exports final recognition result Fobj。
5. gait recognition method according to claim 4, which is characterized in that the step S3.3 specifically:
Work as SobjAnd TobjWhen identical, recognition result FobjEqual to SobjIf FclGreater than objective degrees of confidence threshold value Ocl-1, output identification knot
Fruit Fobj, on the contrary then do not export recognition result；
Work as SobjAnd TobjWhen different, if Scl、TclIt is all larger than Ocl-2, and Scl>Tcl, export recognition result FobjEqual to Sobj, otherwise it is defeated
Recognition result F outobjEqual to Tobj；If Scl、TclRespectively less than Ocl-2, do not export recognition result；If SclAnd TclOne of them is greater than
Ocl-2, another is less than Ocl-2, then FobjEqual to greater than Ocl-2Confidence level belonging network recognition result, export Fobj。
6. a kind of based on the Gait Recognition system for improving VGG-16 network characterized by comprising
Characteristic extracting module is extracted static for being handled based on improvement VGG-16 network original pedestrian's gait data collection
Feature and behavioral characteristics；
Training module obtains static identification network, Dynamic Recognition network, spy for training the static nature and behavioral characteristics
Levy Fusion Module and comprehensive identification module；
Identification module, for based on the static identification network, Dynamic Recognition network, Fusion Features module and comprehensive identification module
Identify the gait in image or video.
7. Gait Recognition system according to claim 6, which is characterized in that the improvement VGG-16 network specifically:
Using VGG-16 network as basic network, residual error network module is added.
8. Gait Recognition system according to claim 7, which is characterized in that the training module includes: static identification net
Network carries out Gait Recognition using the static nature；Dynamic Recognition network carries out Gait Recognition using the behavioral characteristics.
9. Gait Recognition system according to claim 7, which is characterized in that the identification module includes:
Static state identification network obtains objective degrees of confidence S for identifying to original image Picture or video VideoclWith
Recognition result Sobj；
Dynamic Recognition network obtains objective degrees of confidence T for identifying to original image Picture or video VideoclWith
Recognition result Tobj；
Fusion Features module, for real using the fusion method based on Adding law in matching layer multi-biological characteristic blending algorithm
Existing Fusion Features, by objective degrees of confidence Scl、TclWith recognition result Sobj、TobjFusion Features are carried out, fusion objective degrees of confidence is obtained
Fcl；By Scl、Tcl、Fcl、SobjAnd TobjIt is input to comprehensive identification module；
Comprehensive identification module, for being based on Scl、Tcl、Fcl、SobjAnd TobjIt is integrated, exports final recognition result Fobj。
10. Gait Recognition system according to claim 9, which is characterized in that the comprehensive identification module specifically includes:
Work as SobjAnd TobjWhen identical, recognition result FobjEqual to SobjIf FclGreater than objective degrees of confidence threshold value Ocl-1, output identification knot
Fruit Fobj, on the contrary then do not export recognition result；
Work as SobjAnd TobjWhen different, if Scl、TclIt is all larger than Ocl-2, and Scl>Tcl, export recognition result FobjEqual to Sobj, otherwise it is defeated
Recognition result F outobjEqual to Tobj；If Scl、TclRespectively less than Ocl-2, do not export recognition result；If SclAnd TclOne of them is greater than
Ocl-2, another is less than Ocl-2, then FobjEqual to greater than Ocl-2Confidence level belonging network recognition result, export Fobj。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910449120.2A CN110163175A (en) | 2019-05-28 | 2019-05-28 | A kind of gait recognition method and system based on improvement VGG-16 network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910449120.2A CN110163175A (en) | 2019-05-28 | 2019-05-28 | A kind of gait recognition method and system based on improvement VGG-16 network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110163175A true CN110163175A (en) | 2019-08-23 |
Family
ID=67629115
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910449120.2A Pending CN110163175A (en) | 2019-05-28 | 2019-05-28 | A kind of gait recognition method and system based on improvement VGG-16 network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110163175A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110969087A (en) * | 2019-10-31 | 2020-04-07 | 浙江省北大信息技术高等研究院 | Gait recognition method and system |
| CN112507803A (en) * | 2020-11-16 | 2021-03-16 | 北京理工大学 | Gait recognition method based on double-flow network |
| CN112734911A (en) * | 2021-01-07 | 2021-04-30 | 北京联合大学 | Single image three-dimensional face reconstruction method and system based on convolutional neural network |
| CN114012742A (en) * | 2022-01-05 | 2022-02-08 | 北京动思创新科技有限公司 | Control system of hip joint power assisting device |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103049758A (en) * | 2012-12-10 | 2013-04-17 | 北京工业大学 | Method for realizing remote authentication by fusing gait flow images (GFI) and head and shoulder procrustes mean shapes (HS-PMS) |
| CN108491794A (en) * | 2018-03-22 | 2018-09-04 | 腾讯科技（深圳）有限公司 | The method and apparatus of face recognition |
| CN108683724A (en) * | 2018-05-11 | 2018-10-19 | 江苏舜天全圣特科技有限公司 | A kind of intelligence children's safety and gait health monitoring system |
| US20180366228A1 (en) * | 2015-11-24 | 2018-12-20 | International Business Machines Corporation | Performing a health analysis using a smart floor mat |
| CN109446991A (en) * | 2018-10-30 | 2019-03-08 | 北京交通大学 | Gait recognition method based on global and local Fusion Features |
-
2019
- 2019-05-28 CN CN201910449120.2A patent/CN110163175A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103049758A (en) * | 2012-12-10 | 2013-04-17 | 北京工业大学 | Method for realizing remote authentication by fusing gait flow images (GFI) and head and shoulder procrustes mean shapes (HS-PMS) |
| US20180366228A1 (en) * | 2015-11-24 | 2018-12-20 | International Business Machines Corporation | Performing a health analysis using a smart floor mat |
| CN108491794A (en) * | 2018-03-22 | 2018-09-04 | 腾讯科技（深圳）有限公司 | The method and apparatus of face recognition |
| CN108683724A (en) * | 2018-05-11 | 2018-10-19 | 江苏舜天全圣特科技有限公司 | A kind of intelligence children's safety and gait health monitoring system |
| CN109446991A (en) * | 2018-10-30 | 2019-03-08 | 北京交通大学 | Gait recognition method based on global and local Fusion Features |
Non-Patent Citations (1)
| Title |
|---|
| 陈峰: "基于步态的目标识别技术的研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110969087A (en) * | 2019-10-31 | 2020-04-07 | 浙江省北大信息技术高等研究院 | Gait recognition method and system |
| CN110969087B (en) * | 2019-10-31 | 2023-11-21 | 杭州未名信科科技有限公司 | Gait recognition method and system |
| CN112507803A (en) * | 2020-11-16 | 2021-03-16 | 北京理工大学 | Gait recognition method based on double-flow network |
| CN112734911A (en) * | 2021-01-07 | 2021-04-30 | 北京联合大学 | Single image three-dimensional face reconstruction method and system based on convolutional neural network |
| CN114012742A (en) * | 2022-01-05 | 2022-02-08 | 北京动思创新科技有限公司 | Control system of hip joint power assisting device |
| CN114012742B (en) * | 2022-01-05 | 2022-03-29 | 北京动思创新科技有限公司 | Control system of hip joint power assisting device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Liu et al. | Research on the use of YOLOv5 object detection algorithm in mask wearing recognition | |
| CN110163175A (en) | A kind of gait recognition method and system based on improvement VGG-16 network | |
| CN104537393B (en) | A kind of traffic sign recognition method based on multiresolution convolutional neural networks | |
| CN106203395B (en) | Face attribute recognition method based on multitask deep learning | |
| CN107844795B (en) | Convolutional neural network feature extraction method based on principal component analysis | |
| CN111325111A (en) | Pedestrian re-identification method integrating inverse attention and multi-scale deep supervision | |
| Qu et al. | A fast face recognition system based on deep learning | |
| Sincan et al. | Using motion history images with 3d convolutional networks in isolated sign language recognition | |
| CN107292250A (en) | A kind of gait recognition method based on deep neural network | |
| CN107092894A (en) | A kind of motor behavior recognition methods based on LSTM models | |
| CN106909938B (en) | Visual angle independence behavior identification method based on deep learning network | |
| CN106778664A (en) | The dividing method and its device of iris region in a kind of iris image | |
| CN109447053A (en) | A kind of face identification method based on dual limitation attention neural network model | |
| CN108985231B (en) | Palm vein feature extraction method based on multi-scale convolution kernel | |
| CN106897673A (en) | A kind of recognition methods again of the pedestrian based on retinex algorithms and convolutional neural networks | |
| CN108960288B (en) | Three-dimensional model classification method and system based on convolutional neural network | |
| CN108564040B (en) | Fingerprint activity detection method based on deep convolution characteristics | |
| CN114299559A (en) | Finger vein identification method based on lightweight fusion global and local feature network | |
| Xu et al. | Research on ResNet101 network chemical reagent label image classification based on transfer learning | |
| CN110135277B (en) | Human behavior recognition method based on convolutional neural network | |
| Harish et al. | A convolution neural network engine for sclera recognition | |
| CN107818299A (en) | Face recognition algorithms based on fusion HOG features and depth belief network | |
| CN112597873A (en) | Dual-channel facial expression recognition method based on deep learning | |
| CN111985332A (en) | Gait recognition method for improving loss function based on deep learning | |
| Wang et al. | Research on gesture image recognition method based on transfer learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190823 |
|
| RJ01 | Rejection of invention patent application after publication |