CN110097029A - Identity identifying method based on Highway network multi-angle of view Gait Recognition - Google Patents
Identity identifying method based on Highway network multi-angle of view Gait Recognition Download PDFInfo
- Publication number
- CN110097029A CN110097029A CN201910398430.6A CN201910398430A CN110097029A CN 110097029 A CN110097029 A CN 110097029A CN 201910398430 A CN201910398430 A CN 201910398430A CN 110097029 A CN110097029 A CN 110097029A
- Authority
- CN
- China
- Prior art keywords
- people
- gait
- sample
- visual angle
- layer
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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 identity identifying method based on Highway network multi-angle of view Gait Recognition, feature-rich this problem can not be extracted by mainly solving the multi-angle of view gait recognition method based on convolutional neural networks in the prior art.Specific step is as follows by the present invention: (1) constructing Highway network (2) and construct gait energy diagram；(3) training set, verifying collection and test set are constructed；(4) Highway network is tested in training Highway network (5).Present invention introduces the Highway network structures that can distinguish two picture similitudes, can extract relatively rich visual angle gait feature.
Description
Technical field
The invention belongs to field of computer technology, further relate to one of computer vision recognition technology field base
In the identity identifying method of the multi-angle of view Gait Recognition of Highway network.The present invention can be used for from body gait energy diagram mentioning
Relatively rich visual angle gait feature is taken, and human body authentication is carried out according to extracted multi-angle of view gait feature.
Background technique
Gait Recognition technology is that a kind of biotechnology of identification is carried out according to the posture that people in video sequence walks.By
In Gait Recognition have the characteristics that non-infringement property, remote identity and be difficult to it is hiding, therefore Gait Recognition country it is public
The fields such as safety, financial security, authentication, video monitoring have a wide range of applications.Multi-angle of view Gait Recognition is by all
The influence of such as leg speed size, condition of wearing the clothes and knapsack state complex environment factor, and extracted from gait energy diagram more
Visual angle gait feature has a significant impact the performance of multi-angle of view Gait Recognition.And Highway network can be set by network structure
Meter, excitation function and learning rules provide particular task based on the feature-rich learnt to obtain abstract characteristics abundant
Deduction.So the identity identifying method based on Highway network multi-angle of view Gait Recognition is relative to traditional multi-angle of view gait
The identity identifying method of identification has significant performance advantage.
A kind of patent document " gait knowledge based on deep learning of scientific and technological (Beijing) Co., Ltd of milky way water droplet in its application
It is proposed in other method " (number of patent application: 201410587758X, application publication number: CN104299012A) a kind of based on depth
The gait recognition method of study.This method describes gait sequence using gait energy diagram, is instructed by depth convolutional neural networks
Practice Matching Model, to match the identity of Gait Recognition people.The training process of this method are as follows: to marked good identity be related to it is more
The training gait video sequence at a visual angle extracts gait energy diagram, repeats to choose any two of them to based on convolutional neural networks
Matching Model be trained until model restrain；The identification process of this method are as follows: to single-view gait video to be identified and
Registration gait video sequence extracts gait energy diagram respectively, utilizes based on convolutional neural networks trained in training process
The gait energy diagram and the registered each gait energy diagram of gait video sequence of single-view gait video to be identified are calculated with model
Similarity, the size according to similarity carries out identity prediction, and exports recognition result.This method can obtain preferable more views
Angle Gait Recognition performance.But the shortcoming that this method still has is, when human body walking visual angle change range is larger,
The extracted multi-angle of view gait feature characterization ability of this method is inadequate, and the multi-angle of view gait feature learnt is limited, influences regard more
Angle Gait Recognition accuracy rate.
Inst. of Computing Techn. Academia Sinica is (special in the patent document " gait identification method and system " of its application
Sharp application number: CN201710136803.3, application publication number: CN107016346A) in a kind of gait recognition method is provided.The party
Method extracts behavioural characteristic, position feature and gait feature from the acceleration information that the intelligent terminal carried by user obtains；Benefit
User's current behavior is identified according to the behavioural characteristic with preparatory trained Activity recognition model；Using in advance it is trained
Position identification model identifies the present bit of the intelligent terminal according to the position feature and the user's current behavior identified
It sets；Using preparatory trained Gait Recognition model according to the gait feature, the user's current behavior and the intelligence that are identified
Can the current location of terminal identify the identity of user.This method improves the knowledge of gait identity by being layered progressive identification method
Other accuracy rate and robustness, and do not need to limit the position and direction that related sensor is put, it is very flexibly and square
Just it uses.But the shortcoming that this method still has is, adds since this method needs user to carry intelligent terminal to obtain
Speed data extracts behavioural characteristic, position feature and the gait feature of user using the acceleration information of acquisition, this is weakened
The non-infringement of Approach for Gait Classification, and this method is by preparatory trained Activity recognition model, position identification model and gait
The feature of extraction, can not be dissolved into a model and realize Gait Recognition end to end by three model compositions of identification model.
Summary of the invention
It is a kind of based on Highway network mostly view it is an object of the invention in view of the above shortcomings of the prior art, propose
The identity identifying method of angle Gait Recognition, it is inadequate for solving the extracted multi-angle of view gait feature characterization ability of existing method,
The problem that the multi-angle of view gait feature learnt is limited and causes multi-angle of view Gait Recognition accuracy rate lower.
Realizing the thinking of the object of the invention is, because Method of Gait Feature Extraction precision has directly multi-angle of view Gait Recognition accuracy rate
It influences, and Highway network can learn the feature of different levels, increase the pumping of feature by splicing the feature of different levels
As property and rich, the precision of feature extraction is improved, to improve the accuracy of multi-angle of view Gait Recognition.The present invention is as out
Point is sent out, a kind of identity identifying method based on Highway network multi-angle of view Gait Recognition is devised.First to all gait videos
Gait energy diagram is synthesized, gait energy diagram construction multi-angle of view gait energy diagram training set, verifying collection and test set are utilized.In training
Stage obtains by training set gait energy diagram to being sent into designed Highway network structure in batches and represents two gaits
The probability of energy diagram similarity, and then using the loss of the probability calculation model obtained, more using back-propagation algorithm finally
The weight of new network determines whether training terminates according to the size of training set accuracy rate and verifying collection accuracy rate.In test rank
Section, test set is sent in trained Highway, the identity of each tester is determined using k nearest neighbor algorithm, is finally obtained
The accuracy rate of test set under each viewing angle.
The specific steps of the present invention are as follows:
(1) Highway network is constructed:
(1a) builds a dimeric characteristic extracting module；Wherein, the structure of the first part is successively are as follows:
Input layer → the first convolutional layer；The structure of second part is successively are as follows: first normalization layer → second convolutional layer → second batch is returned
One changes layer → third convolutional layer → third batch normalization layer → Volume Four lamination；The first part composes in series with second part
Characteristic extracting module；
(1b) builds a feature learning module, and structure is successively are as follows: the 4th batch of normalization layer → first partial response is returned
One, which changes layer → first pond layer → the 5th convolutional layer → five crowd, normalizes layer → second local acknowledgement normalization layer → second pond
Change layer → the 6th full articulamentum → output layer of random deactivating layer → the first of convolutional layer → the first；
Characteristic extracting module and feature learning module are composed in series Highway network by (1c)；
The parameter that (1d) is arranged each layer in Highway network architecture is as follows:
2 are set by the input layer characteristic pattern sum of Highway network model, characteristic pattern is sized to 126 × 126；
It will be in the first convolutional layer, the second convolutional layer, third convolutional layer, Volume Four lamination, the 5th convolutional layer, the 6th convolutional layer
The sum of convolution filter is respectively set to 16,16,16,16,64,256, and the size of convolution filter therein is respectively set
It is 7 × 7,1 × 1,3 × 3,1 × 1,7 × 7,7 × 7, convolution step-length is disposed as 1 pixel；
First is normalized into layer, second batch normalization layer, third batch normalization layer, the 4th batch of normalization layer, the 5th batch
The port number of normalization layer is respectively set to 16,16,16,16,64；
2 × 2 are set by the area size of the first pond layer and the second pond layer, pond step-length is set as 2 pixels；
0.5 is set by the reservation probability of the first random deactivating layer；
First partial is responded into normalization layer, the parameter of the second local acknowledgement normalization layer is disposed as: α=10-4, β=
0.75, γ=2, k=5；
(2) gait energy diagram is constructed:
(2a) carries out background modeling to 124 body gait videos of input, obtains everyone all body gait wheels
Wide foreground image；
(2b) utilizes gait energy diagram synthetic method, synthesizes everyone gait energy diagram under each viewing angle；
(3) training set, verifying collection and test set are constructed:
From the gait energy diagram of 124 people of input, using the methods of sampling, the gait energy diagram of preceding 50 people is chosen,
Training set is formed, the gait energy diagram of intermediate 24 people is chosen, composition verifying collection chooses the gait energy diagram of last 50 people,
Form test set；
(4) training Highway network:
Training set is input in Highway network by (4a), carries out identity prediction using k nearest neighbor algorithm, and using accurately
Rate calculation formula calculates training set accuracy rate；
Verifying collection is input in Highway network by (4b), carries out identity prediction using k nearest neighbor algorithm, and using accurately
Rate calculation formula calculates verifying collection accuracy rate；
(4c) judges whether verifying collection accuracy rate is less than training set accuracy rate, if so, executing step (4a), otherwise, executes
Step (4d)；
(4d) obtains trained Highway network；
(5) Highway network is tested:
Test set is input in trained Highway network by (5a), using k nearest neighbor algorithm, carries out identity prediction；
(5b) utilizes test set accuracy rate calculation formula, calculates test set under each viewing angle accurate after identity prediction
Rate.
The present invention compared with prior art, has the advantage that
First, since the present invention extracts the feature of different levels by building Highway network, increase the abstract of feature
Property and rich, improves the precision of feature extraction.The prior art is overcome when human body walking visual angle change range is larger, is mentioned
The multi-angle of view gait feature characterization ability taken is inadequate, the limited problem of the multi-angle of view gait feature learnt, so that the present invention has
The advantages of having the feature for extracting abundant gait, improving the accuracy rate of multi-angle of view Gait Recognition.
Second, since the present invention utilizes Highway module directly to predict the similitude of two gait energy diagrams, overcome existing
There is technology sub-module to extract feature, allow the extraction feature that the present invention is implicit, realizes multi-angle of view gait end to end
Identification.
Detailed description of the invention
Fig. 1 is flow chart of the invention；
Fig. 2 is the accuracy rate line chart of test set under each viewing angle in emulation experiment of the present invention；
Fig. 3 is the Highway schematic network structure constructed in emulation experiment of the present invention.
Specific embodiment
With reference to the accompanying drawing, the present invention will be further described.
Referring to attached drawing 1, the specific steps realized to the present invention are further described.
Step 1, Highway network is constructed.
Build a dimeric characteristic extracting module；Wherein, the structure of the first part is successively are as follows: input
Layer → the first convolutional layer；The structure of second part is successively are as follows: first normalization layer → second convolutional layer → second batch normalization
Layer → third convolutional layer → third batch normalization layer → Volume Four lamination；The first part and second part compose in series feature
Extraction module.
A feature learning module is built, structure is successively are as follows: the 4th batch of normalization layer → first partial response normalization
Layer → the first pond layer → the 5th convolutional layer → five crowd normalizes layer → second local acknowledgement and normalizes layer → second pond layer
→ the six full articulamentum → output layer of random deactivating layer → the first of convolutional layer → the first.
Characteristic extracting module and feature learning module are composed in series into Highway network.
It is as follows that each layer in Highway network architecture of parameter is set:
2 are set by the input layer characteristic pattern sum of Highway network model, characteristic pattern is sized to 126 × 126.
It will be in the first convolutional layer, the second convolutional layer, third convolutional layer, Volume Four lamination, the 5th convolutional layer, the 6th convolutional layer
The sum of convolution filter is respectively set to 16,16,16,16,64,256, and the size of convolution filter therein is respectively set
It is 7 × 7,1 × 1,3 × 3,1 × 1,7 × 7,7 × 7, convolution step-length is disposed as 1 pixel.
First is normalized into layer, second batch normalization layer, third batch normalization layer, the 4th batch of normalization layer, the 5th batch
The port number of normalization layer is respectively set to 16,16,16,16,64.
2 × 2 are set by the area size of the first pond layer and the second pond layer, pond step-length is set as 2 pixels.
0.5 is set by the reservation probability of the first random deactivating layer.
First partial is responded into normalization layer, the parameter of the second local acknowledgement normalization layer is disposed as: α=10-4, β=
0.75, γ=2, k=5.
Step 2, gait energy diagram is constructed.
Background modeling is carried out to 124 body gait videos of input, before obtaining everyone all body gait profiles
Scape image.
The background modeling refers to, the prospect frame of every frame in everyone body gait video is subtracted each other with background frames,
The gait profile foreground image of every frame in the body gait video of the people is obtained, by the gait profile foreground image of all frames of the people
Constitute the body gait profile foreground image of the people.
Using gait energy diagram synthetic method, everyone gait energy diagram under each viewing angle is synthesized.
The step of described gait energy diagram synthetic method, is as follows:
Step 1 selects the people of a non-selected mistake from 124 people.
Step 2 selects the visual angle of a unselected mistake from 11 visual angles of the people.
Step 3 selects a unselected mistake from all gait profile foreground images under visual angle selected by it of choosing
Position sums to the pixel value at the selected location for all gait profile foreground images under visual angle selected by it of choosing.
Step 4, by the selected location for all gait profile foreground images chosen under visual angle selected by it labeled as
Selection.
Step 5, all whether all positions for all gait profile foreground images for judging to choose under visual angle selected by it
It is marked as selected, if so, the summation gait profile foreground image for obtaining choosing under visual angle selected by it, executes this step
Step 6 otherwise execute the step 3 of this step.
Step 6 selects a unselected mistake from the summation gait profile foreground image under the visual angle chosen selected by it
Position, by the pixel value of the selected location of the summation gait profile foreground image under the visual angle chosen selected by it divided by institute
The sum for all gait profile foreground images chosen under visual angle selected by it.
The selected location of summation gait profile foreground image under the visual angle chosen selected by it is labeled as by step 7
It is selected.
Step 8, all whether all positions in the summation gait profile foreground image for judging to choose under visual angle selected by it
Labeled as selected, if so, obtaining the gait energy diagram chosen under visual angle selected by it, this step step 9 is executed, otherwise,
Execute the step 6 of this step.
Step 9, by the selected visual angle chosen labeled as selected, if all visual angles chosen are collectively labeled as having selected
It selects, then will choose labeled as selected.
Step 10 executes the step 11 of this step if proprietary all visual angles are collectively labeled as selected, and otherwise, executes
The step 1 of this step.
Step 11, gait energy diagram synthesis of the owner under all visual angles finish.
Step 3, training set, verifying collection and test set are constructed.
From the gait energy diagram of 124 people of input, using the methods of sampling, the gait energy diagram of preceding 50 people is chosen,
Training set is formed, the gait energy diagram of intermediate 24 people is chosen, composition verifying collection chooses the gait energy diagram of last 50 people,
Form test set.
Steps are as follows for the methods of sampling:
Step 1 randomly chooses the gait energy diagram of a people from the gait energy diagram of selected x people, forms training set
When, x refers to preceding 50 people, and when composition verifying collects, x refers to intermediate 25 people, and when forming test set, x refers to last 50 people, from
Two visual angles are randomly choosed in selected gait energy diagram, and corresponding two gait energy diagrams in two visual angles are spliced into one according to channel
A positive sample sets 1 for the label of the positive sample.
Step 2 randomly chooses the gait energy diagram of two people, from every selected from the gait energy diagram of selected x people
A visual angle is randomly choosed in gait energy diagram, and the corresponding gait energy diagram in each visual angle is spliced into a negative sample according to channel
This, sets 0 for the label of the negative sample.
Step 3 constructs 64 positive samples using method identical with the step 1 of this step, utilizes this step step 2 phase
Same method constructs 64 negative samples.
64 positive samples and 64 negative samples are formed a collection of training data by step 4.
Step 5 constructs 200000 batches of training datas, by 200000 using this step step 1 to the identical method of step 4
It criticizes training data and forms training set.
Step 6 constructs positive sample of the x people under all visual angles using method identical with the step 1 of this step and step 2
This and negative sample, when x refers to intermediate 25 people, positive sample and negative sample composition verifying collection under all visual angles, when x refers to most
Afterwards when 50 people, positive sample and negative sample under all visual angles form test set.
Step 4, training Highway network.
(4.1) training set is input in Highway network, identity prediction is carried out using k nearest neighbor algorithm, and using accurately
Rate calculation formula calculates training set accuracy rate.
It is as follows that the k nearest neighbor algorithm carries out the step of identity prediction:
Step 1 chooses a non-selected people as people to be identified from 50 people of training set.
Step 2 is matched with the sample of 50 people in training set two-by-two respectively with the sample of people to be identified, obtain with it is to be identified
The related combined sample of people.
Combined sample related with people to be identified is input in Highway network by step 3, and output has with people to be identified
The similarity vector of the combined sample of pass.
Step 4 arranges the element in similarity vector in descending order, and first 3 are extracted from the similarity vector after sequence
Element, and obtain identity corresponding to this 3 elements.
The identity of people to be identified is predicted as the identity that frequency of occurrence is most in this 3 identity by step 5.
Step 6, by selected people to be identified labeled as selected.
Step 7 executes the step 8 of this step if owner selects to finish, and otherwise, executes the step 1 of this step.
Step 8, identity prediction finish.
The accuracy rate calculation formula is as follows:
Wherein, P indicates the accuracy rate of training set, and T indicates the total sample number of training set, and ∑ indicates sum operation, and t indicates instruction
Practice collection sample serial number, ItIt indicates the indicator function of t-th of sample in training set, indicates t-th of sample in training set in step (4a)
This indicator function, when the identity prediction of t-th of sample in training set is correct, It=1, otherwise, It=0.
(4.2) verifying collection is input in Highway network, identity prediction is carried out using k nearest neighbor algorithm, and using accurately
Rate calculation formula calculates verifying collection accuracy rate.
It is as follows that the k nearest neighbor algorithm carries out the step of identity prediction:
Step 1 chooses a non-selected people as people to be identified from 25 people of verifying collection.
Step 2, with the sample of people to be identified respectively with verifying concentrate 25 people sample match two-by-two, obtain with it is to be identified
The related combined sample of people.
Combined sample related with people to be identified is input in Highway network by step 3, and output has with people to be identified
The similarity vector of the combined sample of pass.
Step 4 arranges the element in similarity vector in descending order, and first 3 are extracted from the similarity vector after sequence
Element, and obtain identity corresponding to this 3 elements.
The identity of people to be identified is predicted as the identity that frequency of occurrence is most in this 3 identity by step 5.
Step 6, by selected people to be identified labeled as selected.
Step 7 executes the step 8 of this step if owner selects to finish, and otherwise, executes the step 1 of this step.
Step 8, identity prediction finish.
The accuracy rate calculation formula is as follows:
Wherein, P indicates the accuracy rate of verifying collection, indicates the total sample number of verifying collection, and ∑ indicates sum operation, and t expression is tested
Card collection sample serial number, ItIndicate that the indicator function of t-th of sample is concentrated in verifying, when the identity prediction of t-th of sample is concentrated in verifying
When correct, It=1, otherwise, It=0.
(4.3) judge whether verifying collection accuracy rate is less than training set accuracy rate, if so, (4.1) of this step are executed, it is no
Then, (4.4) of this step are executed.
(4.4) trained Highway network is obtained.
Step 5, Highway network is tested.
Test set is input in trained Highway network, using k nearest neighbor algorithm, carries out identity prediction.
It is as follows that the k nearest neighbor algorithm carries out the step of identity prediction:
Step 1 chooses a non-selected people as people to be identified from 50 people of test set.
Step 2 is matched with the sample of 50 people in test set two-by-two respectively with the sample of people to be identified, obtain with it is to be identified
The related combined sample of people.
Combined sample related with people to be identified is input in Highway network by step 3, and output has with people to be identified
The similarity vector of the combined sample of pass.
Step 4 arranges the element in similarity vector in descending order, and first 3 are extracted from the similarity vector after sequence
Element, and obtain identity corresponding to this 3 elements.
The identity of people to be identified is predicted as the identity that frequency of occurrence is most in this 3 identity by step 5.
Step 6, by selected people to be identified labeled as selected.
Step 7 executes the step 8 of this step if owner selects to finish, and otherwise, executes the step 1 of this step.
Step 8, identity prediction finish.
Using test set accuracy rate calculation formula, the accuracy rate of test set under each viewing angle after identity prediction is calculated.
The accuracy rate calculation formula is as follows:
Wherein, AlIndicate that test set is accurate under first of visual angle, the serial number at 11 visual angles of l expression, l=1,2,3,4,
5,6,7,8,9,10,11, l be the corresponding visual angle of 1 expression be 0 °, and l is that the corresponding visual angle of 2 expressions is 18 °, and l is 3 to indicate corresponding
Visual angle is 36 °, and l is that the corresponding visual angle of 4 expressions is 54 °, and l is that the corresponding visual angle of 5 expressions is 72 °, and l is the corresponding visual angle of 6 expressions
It is 90 °, l is that the corresponding visual angle of 7 expressions is 108 °, and l is that the corresponding visual angle of 8 expressions is 126 °, and l is that the corresponding visual angle of 9 expressions is
144 °, l is that the corresponding visual angle of 10 expressions is 162 °, and l is that the corresponding visual angle of 11 expressions is 180 °, NlIndicate test set at the visual angle l
Under total sample number, ∑ indicates sum operation, and nl indicates the serial number of sample of the test set under the visual angle l, InlIndicate test set in l
The indicator function of n-th of sample under visual angle, when the prediction of the identity of n-th sample of the test set under the visual angle l is correct, Inl=
1, otherwise Inl=0.
Effect of the invention is described further below with reference to emulation experiment.
1, emulation experiment condition:
The hardware platform of emulation experiment of the invention are as follows: processor is Intel Core i7-8700K CPU, and dominant frequency is
3.70GHz inside saves as 64GB.
The software platform of emulation experiment of the invention are as follows: Windows7 operating system, python2.7 and Pytorch frame.
Emulation experiment of the present invention, should using gait data library Dataset B disclosed in Institute of Automation Research of CAS
Database is collected in January, 2005, including (usual terms wears overcoat to 124 pedestrians, carries package under three kinds of dressing states
Condition) and 11 visual angles under (0 °, 18 °, 36 °, 54 °, 72 °, 90 °, 108 °, 126 °, 144 °, 162 °, 180 °) walking video.
2, emulation content and interpretation of result:
Emulation experiment of the present invention synthesizes the gait energy diagram of 124 pedestrians, and the gait energy diagram for choosing preceding 50 people is constituted
Training set, the gait energy diagram of intermediate 24 people constitute verifying collection, and the gait energy diagram of last 50 people constitutes test set.
Emulation experiment of the present invention is using the multi-angle of view Gait Recognition of method of the invention and a prior art based on CNN
Method carries out identity prediction to the pedestrian that input test is concentrated respectively, calculates the predictablity rate of test set under each viewing angle.
In emulation experiment, the multi-angle of view gait recognition method based on CNN of use refers to, Wu Z, Huang Y, Wang
L et al. is in " A Comprehensive Study on Cross-View Gait Based Human Identification
With Deep CNNs, IEEE Trans.Pattern Analysis&Machine Intelligence., vol.39, no.8,
Across the viewpoint classification method based on CNN proposed in pp.209-226,2016 ".In former paper, dividing across visual angle based on CNN
Class method not only can be used for across visual angle Gait Recognition but also can be used to multi-angle of view Gait Recognition, and across visual angle gait is known in the paper
The best network structure of other effect is LB, so the network structure of the multi-angle of view classification method of the present invention based on CNN
Refer specifically to the LB network structure in the multi-angle of view gait recognition method based on CNN.
In the training stage for the Highway network that emulation experiment of the present invention constructs the present invention, training set is inputted first
Into the Highway network structure built, the weight of network is updated using back-propagation algorithm, is obtained trained
Highway network.Secondly, calculating the accuracy rate of training set, then verifying collection is input to trained Highway network, and count
Calculate the accuracy rate of verifying collection.Training is iterated to Highway network, until verifying collection accuracy rate is accurate more than or equal to training set
Rate obtains final trained Highway network, terminates training.
In the test phase for the Highway network that emulation experiment of the present invention constructs the present invention, test set is sent to instruction
In the Highway network perfected, the accuracy rate of test set under each viewing angle is calculated, the accuracy rate under each visual angle is depicted as
Broken line, as shown in Figure 2.
It is indicated using method of the invention, respectively under 11 visual angles of test set in Fig. 2 with the dotted line that dot indicates
The broken line of multi-angle of view discrimination.The multi-angle of view based on CNN point using the prior art is indicated with the dotted line that five-pointed star indicates in Fig. 2
Class method, the respectively broken line of the multi-angle of view discrimination under 11 visual angles of test set.Abscissa in Fig. 2 indicates visual angle model
It encloses, takes 0 °, 18 °, 36 °, 54 °, 72 °, 90 °, 108 °, 126 °, 144 °, 162 °, 180 ° respectively.Ordinate in Fig. 2 is test
Collect corresponding multi-angle of view discrimination under each viewing angle.
It can be seen in fig. 2 that the discrimination line chart of the method for the present invention is in the prior art when visual angle is between 0 ° to 180 °
The multi-angle of view classification method based on CNN top, when illustrating that visual angle is between 0 ° to 180 °, the discrimination of the method for the present invention is wanted
Higher than the discrimination of the multi-angle of view classification method based on CNN, it was demonstrated that network structure designed by the present invention can be extracted abundant
Gait feature improves the discrimination of multi-angle of view gait recognition method.
The above emulation experiment shows: using the Highway network built, can extract the gait feature of different levels, increase
Add the abstractness of feature and rich, improves the precision of feature extraction, improve the accuracy rate of multi-angle of view Gait Recognition, overcome existing
There is technology when human body walking visual angle change range is larger, extracted multi-angle of view gait feature characterization ability is inadequate, and study is arrived
The limited problem of multi-angle of view gait feature, be a kind of very useful multi-angle of view gait recognition method.And the method for the present invention
The similitude that two gait energy diagrams are directly predicted using Highway module is overcome prior art sub-module and extracts spy
Sign, allows the extraction feature that the present invention is implicit, realizes multi-angle of view Gait Recognition end to end.
Claims (7)
1. a kind of identity identifying method based on Highway network multi-angle of view Gait Recognition, which is characterized in that construct and train
The human body multi-angle of view gait feature of Highway network, extraction compares two picture similitudes, utilizes trained Highway net
Network carries out authentication, and this method specific steps include the following:
(1) Highway network is constructed:
(1a) builds a dimeric characteristic extracting module；Wherein, the structure of the first part is successively are as follows: input
Layer → the first convolutional layer；The structure of second part is successively are as follows: first normalization layer → second convolutional layer → second batch normalization
Layer → third convolutional layer → third batch normalization layer → Volume Four lamination；The first part and second part compose in series feature
Extraction module；
(1b) builds a feature learning module, and structure is successively are as follows: the 4th batch of normalization layer → first partial response normalization
Layer → the first pond layer → the 5th convolutional layer → five crowd normalizes layer → second local acknowledgement and normalizes layer → second pond layer
→ the six full articulamentum → output layer of random deactivating layer → the first of convolutional layer → the first；
Characteristic extracting module and feature learning module are composed in series Highway network by (1c)；
The parameter that (1d) is arranged each layer in Highway network architecture is as follows:
2 are set by the input layer characteristic pattern sum of Highway network model, characteristic pattern is sized to 126 × 126；
By convolution in the first convolutional layer, the second convolutional layer, third convolutional layer, Volume Four lamination, the 5th convolutional layer, the 6th convolutional layer
The sum of filter is respectively set to 16,16,16,16,64,256, and the size of convolution filter therein is respectively set to 7
× 7,1 × 1,3 × 3,1 × 1,7 × 7,7 × 7, convolution step-length is disposed as 1 pixel；
First is normalized into layer, second batch normalization layer, third batch normalization layer, the 4th batch of normalization layer, the 5th batch of normalizing
The port number for changing layer is respectively set to 16,16,16,16,64；
2 × 2 are set by the area size of the first pond layer and the second pond layer, pond step-length is set as 2 pixels；
0.5 is set by the reservation probability of the first random deactivating layer；
First partial is responded into normalization layer, the parameter of the second local acknowledgement normalization layer is disposed as: α=10-4, β=0.75,
γ=2, k=5；
(2) gait energy diagram is constructed:
(2a) carries out background modeling to 124 body gait videos of input, before obtaining everyone all body gait profiles
Scape image；
(2b) utilizes gait energy diagram synthetic method, synthesizes everyone gait energy diagram under each viewing angle；
(3) training set, verifying collection and test set are constructed:
From the gait energy diagram of 124 people of input, using the methods of sampling, the gait energy diagram of preceding 50 people, composition are chosen
Training set, chooses the gait energy diagram of intermediate 24 people, and composition verifying collection chooses the gait energy diagram of last 50 people, composition
Test set；
(4) training Highway network:
Training set is input in Highway network by (4a), carries out identity prediction using k nearest neighbor algorithm, and utilize accuracy rate meter
Formula is calculated, training set accuracy rate is calculated；
Verifying collection is input in Highway network by (4b), carries out identity prediction using k nearest neighbor algorithm, and utilize accuracy rate meter
Formula is calculated, verifying collection accuracy rate is calculated；
(4c) judges whether verifying collection accuracy rate is less than training set accuracy rate, if so, executing step (4a), otherwise, executes step
(4d)；
(4d) obtains trained Highway network；
(5) Highway network is tested:
Test set is input in trained Highway network by (5a), using k nearest neighbor algorithm, carries out identity prediction；
(5b) utilizes test set accuracy rate calculation formula, calculates the accuracy rate of test set under each viewing angle after identity prediction.
2. the identity identifying method according to claim 1 based on Highway network multi-angle of view Gait Recognition, feature exist
In background modeling described in step (2a) refers to, by the prospect frame and background frames of every frame in everyone body gait video
Subtract each other, the gait profile foreground image of every frame in the body gait video of the people is obtained, before the gait profile of all frames of the people
The body gait profile foreground image of scape image construction the people.
3. the identity identifying method according to claim 1 based on Highway network multi-angle of view Gait Recognition, feature exist
It is as follows in, the gait energy diagram synthetic method described in step (2b) the step of:
The first step selects the people of a non-selected mistake from 124 people；
Second step selects the visual angle of a unselected mistake from 11 visual angles of the people；
Third step selects the position of a unselected mistake from all gait profile foreground images under visual angle selected by it of choosing
It sets, sums to the pixel value at the selected location for all gait profile foreground images under visual angle selected by it of choosing；
The selected location for all gait profile foreground images chosen under visual angle selected by it is labeled as having selected by the 4th step
It selects；
Whether all marked all positions of 5th step, all gait profile foreground images under visual angle selected by it that judge to choose
It is denoted as selected, if so, the summation gait profile foreground image for obtaining choosing under visual angle selected by it, executes the 6th step, it is no
Then, third step is executed；
6th step selects the position of a unselected mistake from the summation gait profile foreground image under the visual angle chosen selected by it
It sets, by the pixel value of the selected location of the summation gait profile foreground image under the visual angle chosen selected by it divided by being chosen
The sum of all gait profile foreground images under visual angle selected by it；
The selected location of summation gait profile foreground image under the visual angle chosen selected by it is labeled as having selected by the 7th step
It selects；
8th step, whether all positions all mark in the summation gait profile foreground image under visual angle selected by it that judges to choose
It is selected, if so, obtaining the gait energy diagram chosen under visual angle selected by it, executes the 9th step, otherwise, executes the 6th
Step；
9th step, by the selected visual angle chosen labeled as selected, if all visual angles chosen are collectively labeled as selected,
It will choose labeled as selected；
Tenth step executes the 11st step if proprietary all visual angles are collectively labeled as selected, and otherwise, executes the first step；
11st step, gait energy diagram synthesis of the owner under all visual angles finish.
4. the identity identifying method according to claim 1 based on Highway network multi-angle of view Gait Recognition, feature exist
In steps are as follows for the methods of sampling described in step (3):
The first step randomly chooses the gait energy diagram of a people from the gait energy diagram of selected x people, when forming training set, x
Refer to preceding 50 people, when composition verifying collection, x refers to intermediate 25 people, and when forming test set, x refers to last 50 people, from selected
Two visual angles are randomly choosed in gait energy diagram, and corresponding two gait energy diagrams in two visual angles are being spliced into one just according to channel
Sample sets 1 for the label of the positive sample；
Second step randomly chooses the gait energy diagram of two people, from selected every gait from the gait energy diagram of selected x people
A visual angle is randomly choosed in energy diagram, and the corresponding gait energy diagram in each visual angle is spliced into a negative sample according to channel, it will
The label of the negative sample is set as 0；
Third step constructs 64 positive samples using the identical method of the first step, constructs 64 using the identical method of second step and bears
Sample；
64 positive samples and 64 negative samples are formed a collection of training data by the 4th step；
5th step constructs 200000 batches of training datas using the first step to the identical method of the 4th step, by 200000 batches of trained numbers
According to composition training set；
6th step constructs positive sample and negative sample of the x people under all visual angles using the first step and the identical method of second step,
When x refers to intermediate 25 people, positive sample and negative sample composition verifying collection under all visual angles, when x refers to last 50 people,
Positive sample and negative sample under all visual angles form test set.
5. the identity identifying method according to claim 1 based on Highway network multi-angle of view Gait Recognition, feature exist
It is as follows that the k nearest neighbor algorithm described in, step (4a), step (4b), step (5a) carries out the step of identity prediction:
The first step chooses a non-selected people as people to be identified, and step (4a) is middle to be indicated to select from 50 people of training set
Take a non-selected people as people to be identified, indicate to choose from 25 people of verifying collection in step (4b) one it is non-selected
For people as people to be identified, step (5a) is middle to indicate to choose a non-selected people from 50 people of test set as to be identified
People；
Second step is matched with owner's sample two-by-two respectively with the sample of people to be identified, and owner's sample indicates in step (4a)
The sample of 50 people in training set, owner's sample indicates that the sample of 25 people, step (5a) institute are concentrated in verifying in step (4b)
Someone's sample indicates the sample of 50 people in test set, obtains combined sample related with people to be identified；
Combined sample related with people to be identified is input in Highway network by third step, is exported related with people to be identified
The similarity vector of combined sample；
4th step arranges the element in similarity vector in descending order, and preceding 3 members are extracted from the similarity vector after sequence
Element, and obtain identity corresponding to this 3 elements；
The identity of people to be identified is predicted as the identity that frequency of occurrence is most in this 3 identity by the 5th step；
6th step, by selected people to be identified labeled as selected；
7th step executes the 8th step if owner selects to finish, and otherwise, executes the first step；
8th step, identity prediction finish.
6. the identity identifying method according to claim 1 based on Highway network multi-angle of view Gait Recognition, feature exist
In accuracy rate calculation formula described in step (4a), step (4b) is as follows:
Wherein, P indicates accuracy rate, and the accuracy rate of training set is indicated in step (4a), the accurate of verifying collection is indicated in step (4b)
Rate, T indicate total sample number, and the total sample number of training set is indicated in step (4a), indicate that the sample of verifying collection is total in step (4b)
Number, ∑ indicate sum operation, and t indicates sample serial number, and the sample serial number of training set is indicated in step (4a), is indicated in step (4b)
Verify the sample serial number of collection, ItIt indicates the indicator function of t-th of sample, t-th of sample is indicated in training set in step (4a)
Indicator function, when the identity prediction of t-th of sample in training set is correct, It=1, otherwise, It=0, step (4b) is middle to be indicated to survey
The indicator function of t-th of sample is concentrated in examination, when the identity prediction of t-th of sample in test set is correct, It=1, otherwise, It=
0。
7. the identity identifying method according to claim 1 based on Highway network multi-angle of view Gait Recognition, feature exist
In accuracy rate calculation formula described in step (5b) is as follows:
Wherein, AlIndicate that test set is accurate under first of visual angle, the serial number at 11 visual angles of l expression, l=1,2,3,4,5,6,7,
8,9,10,11, l be the corresponding visual angle of 1 expression be 0 °, and l is that the corresponding visual angle of 2 expressions is 18 °, and l is that the corresponding visual angle of 3 expressions is
36 °, l is that the corresponding visual angle of 4 expressions is 54 °, and l is that the corresponding visual angle of 5 expressions is 72 °, and l is that the corresponding visual angle of 6 expressions is 90 °, l
Indicate that corresponding visual angle is 108 ° for 7, l is that the corresponding visual angle of 8 expressions is 126 °, and l is that the corresponding visual angle of 9 expressions is 144 °, and l is
10 indicate that corresponding visual angle is 162 °, and l is that the corresponding visual angle of 11 expressions is 180 °, NlIndicate sample of the test set under the visual angle l
Sum, ∑ indicate sum operation, and nl indicates the serial number of sample of the test set under the visual angle l, InlIndicate test set under the visual angle l
The indicator function of n-th of sample, when the prediction of the identity of n-th sample of the test set under the visual angle l is correct, Inl=1, otherwise
Inl=0.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910398430.6A CN110097029B (en) | 2019-05-14 | 2019-05-14 | Identity authentication method based on high way network multi-view gait recognition |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910398430.6A CN110097029B (en) | 2019-05-14 | 2019-05-14 | Identity authentication method based on high way network multi-view gait recognition |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110097029A true CN110097029A (en) | 2019-08-06 |
| CN110097029B CN110097029B (en) | 2022-12-06 |
Family
ID=67447990
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910398430.6A Active CN110097029B (en) | 2019-05-14 | 2019-05-14 | Identity authentication method based on high way network multi-view gait recognition |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110097029B (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112163487A (en) * | 2020-09-21 | 2021-01-01 | 浙江师范大学 | Gait identity recognition method based on improved space-time gait energy diagram |
| CN113065468A (en) * | 2021-04-02 | 2021-07-02 | 西安电子科技大学 | Gait authentication method based on user coordinate system and GRU network |
| CN113095265A (en) * | 2021-04-21 | 2021-07-09 | 西安电子科技大学 | Fungal target detection method based on feature fusion and attention |
| CN113111797A (en) * | 2021-04-19 | 2021-07-13 | 杭州电子科技大学 | Cross-view gait recognition method combining self-encoder and view transformation model |
| CN114360058A (en) * | 2021-12-31 | 2022-04-15 | 北京航空航天大学 | Cross-visual angle gait recognition method based on walking visual angle prediction |
| CN114840834A (en) * | 2022-04-14 | 2022-08-02 | 浙江大学 | Implicit identity authentication method based on gait characteristics |
| CN114998995A (en) * | 2022-06-13 | 2022-09-02 | 西安电子科技大学 | Cross-view-angle gait recognition method based on metric learning and space-time double-flow network |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2016065534A1 (en) * | 2014-10-28 | 2016-05-06 | 中国科学院自动化研究所 | Deep learning-based gait recognition method |
| WO2017133009A1 (en) * | 2016-02-04 | 2017-08-10 | 广州新节奏智能科技有限公司 | Method for positioning human joint using depth image of convolutional neural network |
| CN108460340A (en) * | 2018-02-05 | 2018-08-28 | 北京工业大学 | A kind of gait recognition method based on the dense convolutional neural networks of 3D |
-
2019
- 2019-05-14 CN CN201910398430.6A patent/CN110097029B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2016065534A1 (en) * | 2014-10-28 | 2016-05-06 | 中国科学院自动化研究所 | Deep learning-based gait recognition method |
| WO2017133009A1 (en) * | 2016-02-04 | 2017-08-10 | 广州新节奏智能科技有限公司 | Method for positioning human joint using depth image of convolutional neural network |
| CN108460340A (en) * | 2018-02-05 | 2018-08-28 | 北京工业大学 | A kind of gait recognition method based on the dense convolutional neural networks of 3D |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112163487A (en) * | 2020-09-21 | 2021-01-01 | 浙江师范大学 | Gait identity recognition method based on improved space-time gait energy diagram |
| CN113065468A (en) * | 2021-04-02 | 2021-07-02 | 西安电子科技大学 | Gait authentication method based on user coordinate system and GRU network |
| CN113065468B (en) * | 2021-04-02 | 2023-04-07 | 西安电子科技大学 | Gait authentication method based on user coordinate system and GRU network |
| CN113111797A (en) * | 2021-04-19 | 2021-07-13 | 杭州电子科技大学 | Cross-view gait recognition method combining self-encoder and view transformation model |
| CN113111797B (en) * | 2021-04-19 | 2024-02-13 | 杭州电子科技大学 | Cross-view gait recognition method combining self-encoder and view transformation model |
| CN113095265A (en) * | 2021-04-21 | 2021-07-09 | 西安电子科技大学 | Fungal target detection method based on feature fusion and attention |
| CN114360058A (en) * | 2021-12-31 | 2022-04-15 | 北京航空航天大学 | Cross-visual angle gait recognition method based on walking visual angle prediction |
| CN114360058B (en) * | 2021-12-31 | 2024-05-28 | 北京航空航天大学 | Cross-view gait recognition method based on walking view prediction |
| CN114840834A (en) * | 2022-04-14 | 2022-08-02 | 浙江大学 | Implicit identity authentication method based on gait characteristics |
| CN114840834B (en) * | 2022-04-14 | 2024-06-11 | 浙江大学 | Implicit identity authentication method based on gait characteristics |
| CN114998995A (en) * | 2022-06-13 | 2022-09-02 | 西安电子科技大学 | Cross-view-angle gait recognition method based on metric learning and space-time double-flow network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110097029B (en) | 2022-12-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110097029A (en) | Identity identifying method based on Highway network multi-angle of view Gait Recognition | |
| CN111259850B (en) | Pedestrian re-identification method integrating random batch mask and multi-scale representation learning | |
| CN107330396B (en) | Pedestrian re-identification method based on multi-attribute and multi-strategy fusion learning | |
| CN109948425A (en) | A kind of perception of structure is from paying attention to and online example polymerize matched pedestrian's searching method and device | |
| CN104063719B (en) | Pedestrian detection method and device based on depth convolutional network | |
| CN105512680B (en) | A kind of more view SAR image target recognition methods based on deep neural network | |
| CN104881637B (en) | Multimodal information system and its fusion method based on heat transfer agent and target tracking | |
| CN103679158B (en) | Face authentication method and device | |
| CN110119703A (en) | The human motion recognition method of attention mechanism and space-time diagram convolutional neural networks is merged under a kind of security protection scene | |
| CN109543603B (en) | Micro-expression recognition method based on macro-expression recognition migration | |
| CN106909938B (en) | Visual angle independence behavior identification method based on deep learning network | |
| CN106503687A (en) | The monitor video system for identifying figures of fusion face multi-angle feature and its method | |
| CN106548159A (en) | Reticulate pattern facial image recognition method and device based on full convolutional neural networks | |
| CN106372581A (en) | Method for constructing and training human face identification feature extraction network | |
| CN104537393A (en) | Traffic sign recognizing method based on multi-resolution convolution neural networks | |
| Xu et al. | Aligning correlation information for domain adaptation in action recognition | |
| CN110765873A (en) | Facial expression recognition method and device based on expression intensity label distribution | |
| CN109117897A (en) | Image processing method, device and readable storage medium storing program for executing based on convolutional neural networks | |
| CN103186775A (en) | Human body motion recognition method based on mixed descriptor | |
| CN114998934B (en) | Clothes-changing pedestrian re-identification and retrieval method based on multi-mode intelligent perception and fusion | |
| CN104298974A (en) | Human body behavior recognition method based on depth video sequence | |
| CN104966075B (en) | A kind of face identification method and system differentiating feature based on two dimension | |
| CN105893947A (en) | Bi-visual-angle face identification method based on multi-local correlation characteristic learning | |
| CN110210550A (en) | Image fine granularity recognition methods based on integrated study strategy | |
| Shuang et al. | Scale-balanced loss for object detection |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |