CN109635728A - A kind of isomery pedestrian recognition methods again based on asymmetric metric learning - Google Patents
A kind of isomery pedestrian recognition methods again based on asymmetric metric learning Download PDFInfo
- Publication number
- CN109635728A CN109635728A CN201811515924.XA CN201811515924A CN109635728A CN 109635728 A CN109635728 A CN 109635728A CN 201811515924 A CN201811515924 A CN 201811515924A CN 109635728 A CN109635728 A CN 109635728A
- Authority
- CN
- China
- Prior art keywords
- feature
- different modalities
- pedestrian
- isomery
- depth characteristic
- 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
- 238000000034 method Methods 0.000 title claims abstract description 33
- 238000005259 measurement Methods 0.000 claims abstract description 12
- 230000000644 propagated effect Effects 0.000 claims abstract 3
- 238000012549 training Methods 0.000 claims description 19
- 238000009826 distribution Methods 0.000 claims description 13
- 238000000605 extraction Methods 0.000 claims description 13
- 230000006870 function Effects 0.000 claims description 12
- 239000000284 extract Substances 0.000 claims description 5
- 238000005457 optimization Methods 0.000 claims description 5
- 230000004913 activation Effects 0.000 claims description 3
- 238000012512 characterization method Methods 0.000 claims description 3
- 230000009467 reduction Effects 0.000 claims description 3
- 238000012544 monitoring process Methods 0.000 description 7
- 238000005516 engineering process Methods 0.000 description 5
- 230000004069 differentiation Effects 0.000 description 3
- 238000003860 storage Methods 0.000 description 3
- 230000008569 process Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 230000006399 behavior Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000005520 cutting process Methods 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 239000004744 fabric Substances 0.000 description 1
- 230000008571 general function Effects 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000002265 prevention Effects 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
- G06V20/53—Recognition of crowd images, e.g. recognition of crowd congestion
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/10—Character recognition
- G06V30/19—Recognition using electronic means
- G06V30/192—Recognition using electronic means using simultaneous comparisons or correlations of the image signals with a plurality of references
- G06V30/194—References adjustable by an adaptive method, e.g. learning
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Databases & Information Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
Abstract
The isomery pedestrian recognition methods again based on asymmetric metric learning that the invention discloses a kind of, depth characteristic under different modalities is carried out asymmetric measurement by this method, step is: different modalities depth characteristic being projected to the communal space respectively using the sparse self-encoding encoder of two not shared parameters, global restriction is introduced simultaneously and local restriction goes to constrain the distance between different modalities depth characteristic, is reduced the inter- object distance between different modalities feature and is increased between class distance；The constraint result of global restriction and local restriction is propagated backward in trained network as supervisory signals and is used to correct parameters.The present invention makes network ignore modal information as much as possible and focuses more on identity information by mode gap between diminution different modalities, to improve pedestrian's feature representation power and pedestrian's matching accuracy.
Description
Technical field
The present invention relates to computer vision fields, more particularly, to a kind of isomery row based on asymmetric metric learning
People's recognition methods again.
Background technique
With the fast development of modern society, urban population density is higher and higher, and safety problem also increasingly causes people
Attention.In order to prevent and avoid the generation of security incident in time, a large amount of monitoring camera is mounted and applies in public field
Institute.In face of the monitoring data of complicated monitoring network and magnanimity, how to realize and automatically analyze and interpret the monitoring of the multiple-camera ken
The information that system provides, generation and the good social security of maintenance to crime prevention have positive facilitation.Therefore, pedestrian
A hot research content for having become computer vision field is identified again.
Pedestrian identifies the crucial composition portion of (person re-identification) as video monitoring research field again
Point, the purpose is to monitor in other camera kens of network to some target pedestrian appeared in the monitoring camera ken
A large amount of pedestrians in quickly and accurately this target pedestrian is identified.Pedestrian again identification technology application can greatly reduce
Artificial participation in video monitoring is realized to the pedestrian and its behavior quick in monitor video and is accurately analyzed.Currently, main
Recognition methods is mainly that the external appearance characteristic for passing through extraction pedestrian and color characteristic (RGB feature) go to match to the pedestrian of stream again,
These methods can be considered as RGB-RGB single mode pedestrian matching.However, these methods have a strong hypothesis: assuming that
Clothes of the same pedestrian when different cameras occur remain unchanged as much as possible, can be regarded as pedestrian in short-term and know again
Not.Therefore, when pedestrian clothes occur conspicuousness variation or when the color characteristic of pedestrian become not using when, these
The performance of method will sharply decline, because color characteristic at this time is more the interference shown to model, largely
On the different pedestrians for wearing same color clothes can be misjudged to be same a group traveling together.So in recent years to overcome in extreme condition
The shortcomings that lower color characteristic fails, the data for introducing other mode go to make up the deficiency of RGB data, such as infrared data (IR number
According to), RGB-IR cross-module state pedestrian can be regarded as and match (isomery pedestrian matching), isomery pedestrian identifies that maximum challenge exists again
Mode gap between how reducing different modalities.Currently, there is scholar to propose to be gone to reduce different moulds with the method for depth zero padding
Mode gap between state, but the recognition result of this method is inaccurate, and is not able to satisfy application request still.
Summary of the invention
The purpose of the present invention is to overcome the shortcomings of the existing technology with it is insufficient, provide a kind of based on asymmetric metric learning
Isomery pedestrian recognition methods again, this method can overcome under extreme condition color characteristic failure and isomery pedestrian, and identification accuracy is not again
High disadvantage makes network ignore modal information as much as possible and focus more on identity by mode gap between diminution different modalities
Information, to improve pedestrian's feature representation power and pedestrian's matching accuracy.
The purpose of the present invention is realized by the following technical solution: a kind of isomery pedestrian based on asymmetric metric learning is again
Recognition methods, comprising steps of
During training pattern, the pedestrian image under both modalities which is inputted, extracts depth characteristic respectively；
Depth characteristic under different modalities is subjected to asymmetric measurement, step is: using the sparse of two not shared parameters
Different modalities depth characteristic is projected to the communal space respectively by self-encoding encoder, while introducing global restriction and local restriction goes to constrain
Distance between different modalities depth characteristic reduces the inter- object distance between different modalities feature and increases between class distance；It will be global
The constraint result of constraint and local restriction propagates backward in trained network as supervisory signals for correcting parameters；
The loss of global characteristics and local feature is calculated, according to depth characteristic with overall situation loss, local losses and non-right
Claim global restriction and the sum of local restriction in measurement to reach to be minimised as target and go optimization training pattern.
The present invention through the above steps, as long as giving the pedestrian of any both modalities which recognition training data again, so that it may instruct
Isomery pedestrian identification model again is practised, has the advantages that precision is high, fireballing to isomery pedestrian matching.
Preferably, the depth characteristic of image under different modalities is extracted, step is:
Firstly, using the ResNet50 disaggregated model of the pre-training on ImageNet data set as core network, will lead
It is divided into three branches after dry network；
Then, from top to bottom, each branch extracts the high-level characteristic of disaggregated model, and by its horizontal homogeneous piecemeal；
Then, each branch is obtained into several global characteristics and the part of fixed size by the operation of pondization and dimensionality reduction
Feature；
Finally, above-mentioned global characteristics and local feature are stitched together in order, the depth characteristic of input picture is obtained,
That is the complete characterization expression of pedestrian.
For the mode gap reduced between isomery pedestrian's data, it is non-right that the present invention carries out the depth characteristic under different modalities
Claim measurement, step is:
Firstly, the depth characteristic of extraction is divided into FBAnd FRTwo groups, Wherein B, R respectively represent RGB mode, IR mode,Indicate i-th of depth characteristic vector；
Then, by two groups of feature FBAnd FR, respectively by the sparse self-encoding encoder SAE of two not shared parameters, each is dilute
It dredges self-encoding encoder to be made of two full articulamentums, respectively as encoder E and decoder D, encoder E is responsible for different modalities spy
Sign projects to the communal space, and decoder D is responsible for being remapped to the feature of coding and the consistent sky of input feature vector space size
Between；
Then, building reconstruct loss, is denoted as lr, for constraining the output and input of SAE, it is consistent it as far as possible:
lr=| | fB, DB(EB(fB))||2+||fR, DR(ER(fR))||2；
fB, fR, EB, ER, DB, DRRespectively represent feature, the encoder and decoder of mode B and mode R；
Finally, introducing global restriction in the communal space and being used to constrain the gap between different modalities feature distribution, introduce part
Constraint is believed for reducing the inter- object distance between different modalities feature and increasing between class distance using above-mentioned constraint result as supervision
Number backpropagation returns training pattern and removes amendment parameters.
Further, global restriction is used to constrain the gap between different modalities feature distribution, is denoted as lglobal=W (EB(fB),
ER(fR))2, wherein W meets for two any given distribution X=N (mX, CX) and Y=N (mr, CY), m, C respectively represent X distribution and Y
The mean value and variance of distribution, have
Further, local restriction is used to reduce the inter- object distance between different modalities feature and increases between class distance, note
For llocal=(max (d (f, p))-min (d (f, n))+α), p ∈ A (f), n ∈ B (f), A (f), B (f) are respectively represented and feature f
There is the characteristic set of common identity information and different identity information, d () represents the Euclidean distance between two features, and α is to use
In control positive sample to and negative sample to the hyper parameter of spacing.
Further, in order to keep the communal space more effective, a sparse loss l is constructedsparseRemove the defeated of constraint hidden layer
Out, lsparse=| | EB(fB)||1+||ER(fR)||1。
Further, each sparse self-encoding encoder is made of two full articulamentums with ReLU activation primitive.
Preferably, to the depth characteristic extracted, the loss for calculating global characteristics is gone with Triplet Loss function, is used
Softmax function goes to calculate the loss of local feature, with the reconstruct damage in global loss, local losses and asymmetric measurement
It loses, sparse loss, the sum of global restriction and local restriction reach and be minimised as target and go optimization training pattern.To improve model
Feature representation power.
Compared with the prior art, the invention has the following advantages and beneficial effects:
1, the big, pedestrian of the invention for the modal information gap between different modalities in existing isomery pedestrian again identification mission
With the big disadvantage of difficulty, propose that asymmetric measurement is carried out to the depth characteristic under different modalities mutually goes together under different modalities to reduce
The characteristic distance in the human world may be used on going in arbitrary feature extraction network providing supervision message for training network and realize that end is arrived
The training at end effectively improves the quality of pedestrian's feature extraction and accelerates network convergence.
2, different loss function coorinated trainings are taken with having differentiation for different features in the present invention, compared to single damage
Function training is lost, the present invention is purposive to make network ignore modal information as much as possible and focus more on pedestrian's identity information, from
And the more complete feature representation of pedestrian is obtained, therefore it is better than existing method from far away in precision.
3, the global and local thought combined is employed herein to go to extract pedestrian's feature, compared to single features,
The present invention has obtained the more complete feature representation of pedestrian, therefore increases in precision.
Detailed description of the invention
Fig. 1 is the general function frame diagram of the present embodiment method.
Specific embodiment
The attached figures are only used for illustrative purposes and cannot be understood as limitating the patent；To those skilled in the art,
The omitting of some known structures and their instructions in the attached drawings are understandable.With reference to the accompanying drawings and examples to of the invention
Technical solution is described further.
As shown in Figure 1, a kind of isomery pedestrian recognition methods again based on asymmetric metric learning of the present embodiment, mainly includes
Three feature extraction, asymmetric measurement, classification steps, are below specifically described each step.
One, feature extraction
Feature extraction is carried out to pedestrian, step is:
Firstly, using the ResNet50 disaggregated model of the pre-training on ImageNet data set as core network, will lead
It is divided into three branches after dry network；
Then, from top to bottom, each branch extracts the high-level characteristic of disaggregated model and is 1 piece, 2 by its horizontal homogeneous cutting
Block and 3 pieces；
Then, by each branch by the operation of pondization and dimensionality reduction obtain fixed size be 256 dimension 3 global characteristics and
5 local features；
Finally, the feature that 8 256 are tieed up is stitched together in order to obtain the feature of one 2048 dimension, as defeated
Enter the depth characteristic of image, i.e. the complete characterization expression of pedestrian.Features described above is used for subsequent asymmetric measurement and classification.
The global and local thought combined is employed herein to go to extract pedestrian's feature, compared to single features, originally
Invention has obtained the more complete feature representation of pedestrian, therefore increases in precision.
Two, asymmetric measurement
The step is that the feature under different modalities is projected and rebuild by different projection matrixes, to reduce isomery
Mode gap between pedestrian's data, step is:
Firstly, for the feature of extraction(B/R respectively represents RGB
Mode/IR mode), the feature of extraction is divided by F by mode selectorBAnd FRTwo groups；
Then, by two groups of feature FBAnd FR, respectively by the sparse self-encoding encoder SAE of two not shared parameters, each is dilute
It dredges self-encoding encoder to be made of two full articulamentums with ReLU activation primitive, respectively as encoder E and decoder D, coding
Device E is responsible for different modalities Projection Character to the communal space, and decoder D is responsible for for the feature of coding being remapped to special with input
The consistent space of space size is levied, then building reconstruct loss, is denoted as lr, for constraining the output and input of SAE, it is made to the greatest extent may be used
It is able to maintain consistent.lr=| | fB, DB(EB(fB))||2+||fR, DR(ER(fR))||2, fB, fR, EB, ER, DB, DRRespectively represent mode B
Feature, encoder and decoder with mode R.
Meanwhile in order to keep the communal space more effective, a sparse loss l is constructedsparseThe output of constraint hidden layer is gone,
lsparse=| | EB(fB)||1+||ER(fR)||1；
Finally, introducing global restriction l in the communal spaceglobalWith local restriction llocalIt goes between constraint different modalities feature
Distance, and other modules for returning network as supervisory signals backpropagation go the parameter of amendment characteristic extraction step, make spy
Sign, which is extracted, to be ignored modal information as far as possible and pays close attention to pedestrian's identity information, to improve the expression of characteristics of image.
Global restriction is used to constrain the gap between different modalities feature distribution, is denoted as lglobal=W (EB(fB), ER(fR))2,
Wherein W meets for two any given distribution X=N (mX, CX) and Y=N (mY, CY), m, C respectively represent X distribution and Y points
The mean value and variance of cloth, have Part
Constraint is denoted as l for reducing the inter- object distance between different modalities feature and increasing between class distancelocal=(max (d (f, p))-min
(d (f, n))+α), p ∈ A (f), n ∈ B (f), A (f)/B (f), which is respectively represented, has common identity information and different identity with feature f
The characteristic set of information, d () represent the Euclidean distance between two features, α be for control positive sample to and negative sample pair
The hyper parameter of spacing.
In the present invention using two not shared parameter sparse self-encoding encoder respectively by different modalities Projection Character to share
Space.Global restriction l is introduced simultaneouslyglobalWith local restriction llocalThe distance between constraint different modalities feature is gone, different modalities are made
Inter- object distance between feature reduces and between class distance increases, and effectively makes to ignore modal information in characteristic extraction procedure as far as possible
And id information is paid close attention to, to improve the expression of characteristics of image.
Three, classify
The step is to take different loss coorinated trainings while input feature vector is had to differentiation, to carry out effectively about to pedestrian's feature
Beam, step are: for the depth characteristic extracted from characteristic extraction step, being gone to calculate 3 global spies with Triplet Loss function
The loss for calculating five local features is gone in the loss of sign with Softmax function, then minimizes global loss, office by joint
Optimization mould is removed in the sum of reconstruct loss, sparse loss, global restriction and local restriction in portion's loss and asymmetric metric module
Type improves aspect of model expressiveness.
Different loss function coorinated trainings are taken with having differentiation for different features in the present invention, compared to individual losses
Function training, the present invention is purposive to make network ignore modal information as much as possible and focus more on pedestrian's identity information, thus
The more complete feature representation of pedestrian is obtained, therefore is better than existing method from far away in precision.
The experimental results showed that the present invention knows the Rank1 on data set SYSU-MM01 in current maximum cross-module state pedestrian again
It is promoted respectively from 24.43% and 26.92% to 66.26% and 66.7% with mAP, there is very big performance to mention compared with other methods
It rises.
It can implement the technology that the present invention describes by various means.For example, these technologies may be implemented in hardware, consolidate
In part, software or combinations thereof.For hardware embodiments, processing module may be implemented in one or more specific integrated circuits
(ASIC), digital signal processor (DSP), programmable logic device (PLD), field-programmable logic gate array (FPGA), place
Manage device, controller, microcontroller, electronic device, other electronic units for being designed to execute function described in the invention or
In a combination thereof.
It, can be with the module of execution functions described herein (for example, process, step for firmware and/or Software implementations
Suddenly, process etc.) implement the technology.Firmware and/or software code are storable in memory and are executed by processor.Storage
Device may be implemented in processor or outside processor.
Those of ordinary skill in the art will appreciate that: realize that all or part of the steps of above method embodiment can pass through
The relevant hardware of program instruction is completed, and program above-mentioned can store in a computer-readable storage medium, the program
When being executed, step including the steps of the foregoing method embodiments is executed；And storage medium above-mentioned includes: ROM, RAM, magnetic disk or light
The various media that can store program code such as disk.
Obviously, the above embodiment of the present invention be only to clearly illustrate example of the present invention, and not be pair
The restriction of embodiments of the present invention.For those of ordinary skill in the art, may be used also on the basis of the above description
To make other variations or changes in different ways.There is no necessity and possibility to exhaust all the enbodiments.It is all this
Made any modifications, equivalent replacements, and improvements etc., should be included in the claims in the present invention within the spirit and principle of invention
Protection scope within.
Claims (8)
1. a kind of isomery pedestrian recognition methods again based on asymmetric metric learning, which is characterized in that comprising steps of
During training pattern, the pedestrian image under both modalities which is inputted, extracts depth characteristic respectively；
Depth characteristic under different modalities is subjected to asymmetric measurement, step is: using the sparse self-editing of two not shared parameters
Different modalities depth characteristic is projected to the communal space respectively by code device, while introducing global restriction and local restriction goes constraint different
Distance between mode depth characteristic reduces the inter- object distance between different modalities feature and increases between class distance；By global restriction
It is propagated backward in trained network as supervisory signals for correcting parameters with the constraint result of local restriction；
The loss of global characteristics and local feature is calculated, according to depth characteristic with overall situation loss, local losses and asymmetric degree
The sum of global restriction and local restriction in amount reach be minimised as target go optimization training pattern.
2. the isomery pedestrian recognition methods again according to claim 1 based on asymmetric metric learning, which is characterized in that mention
The depth characteristic of image under different modalities is taken, step is:
Firstly, using the ResNet50 disaggregated model of the pre-training on ImageNet data set as core network, by backbone network
It is divided into three branches after network；
Then, from top to bottom, each branch extracts the high-level characteristic of disaggregated model, and by its horizontal homogeneous piecemeal；
Then, each branch is obtained into several global characteristics and the part spy of fixed size by the operation of pondization and dimensionality reduction
Sign；
Finally, above-mentioned global characteristics and local feature are stitched together in order, the depth characteristic of input picture is obtained, at once
The complete characterization of people is expressed.
3. the isomery pedestrian recognition methods again according to claim 1 based on asymmetric metric learning, which is characterized in that will
Depth characteristic under different modalities carries out asymmetric measurement, and step is:
Firstly, the depth characteristic of extraction is divided into FBAnd FRTwo groups,
Wherein B, R respectively represent RGB mode, IR mode, fi mIndicate i-th of depth characteristic vector；
Then, by two groups of feature FBAnd FR, respectively by the sparse self-encoding encoder SAE of two not shared parameters, each it is sparse from
Encoder is made of two full articulamentums, and respectively as encoder E and decoder D, encoder E is responsible for throwing different modalities feature
To the communal space, decoder D is responsible for being remapped to the feature of coding and input feature vector space size consistent space shadow；
Then, building reconstruct loss, is denoted as lr, for constraining the output and input of SAE, it is consistent it as far as possible:
lr=| | fB, DB(EB(fB))||2+||fR, DR(ER(fR))||2；
fB, fR, EB, ER, DB, DRRespectively represent feature, the encoder and decoder of mode B and mode R；
Finally, introducing global restriction in the communal space and being used to constrain the gap between different modalities feature distribution, introduce local restriction
For reducing the inter- object distance between different modalities feature and increasing between class distance, and above-mentioned constraint result is anti-as supervisory signals
Amendment parameters are removed to training pattern is propagated back to.
4. the isomery pedestrian recognition methods again according to claim 3 based on asymmetric metric learning, which is characterized in that complete
Office's constraint is denoted as l for constraining the gap between different modalities feature distributionglobal=W (EB(fB), ER(fR))2, wherein W meet pair
In two any given distribution X=N (mX, CX) and Y=N (mY, CY), m, C respectively represent mean value and the side of X distribution and Y distribution
Difference has
5. the isomery pedestrian recognition methods again according to claim 3 based on asymmetric metric learning, which is characterized in that office
Portion's constraint is denoted as l for reducing the inter- object distance and increase between class distance that different modalities feature is askedlocal=(max (d (f, p))-
Min (d (f, n))+α), p ∈ A (f), n ∈ B (f), A (f), B (f), which are respectively represented, has common identity information and different bodies from feature f
The characteristic set of part information, d () represent the Euclidean distance between two features, α be for control positive sample to and negative sample
To the hyper parameter of spacing.
6. the isomery pedestrian recognition methods again according to claim 3 based on asymmetric metric learning, which is characterized in that structure
Build a sparse loss lsparseRemove the output of constraint hidden layer, lsparse=| | EB(fB)||1+||ER(fR)||1。
7. the isomery pedestrian recognition methods again according to claim 3 based on asymmetric metric learning, which is characterized in that every
One sparse self-encoding encoder is made of two full articulamentums with ReLU activation primitive.
8. the isomery pedestrian recognition methods again according to claim 1 based on asymmetric metric learning, which is characterized in that right
The depth characteristic extracted goes the loss for calculating global characteristics with Triplet Loss function, goes to calculating office with Softmax function
The loss of portion's feature, with reconstruct loss, the sparse loss, global restriction in global loss, local losses and asymmetric measurement
Reach with the sum of local restriction and is minimised as target and goes optimization training pattern.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811515924.XA CN109635728B (en) | 2018-12-12 | 2018-12-12 | Heterogeneous pedestrian re-identification method based on asymmetric metric learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811515924.XA CN109635728B (en) | 2018-12-12 | 2018-12-12 | Heterogeneous pedestrian re-identification method based on asymmetric metric learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109635728A true CN109635728A (en) | 2019-04-16 |
| CN109635728B CN109635728B (en) | 2020-10-13 |
Family
ID=66073049
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811515924.XA Active CN109635728B (en) | 2018-12-12 | 2018-12-12 | Heterogeneous pedestrian re-identification method based on asymmetric metric learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109635728B (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111079939A (en) * | 2019-11-28 | 2020-04-28 | 支付宝(杭州)信息技术有限公司 | Machine learning model feature screening method and device based on data privacy protection |
| CN112001438A (en) * | 2020-08-19 | 2020-11-27 | 四川大学 | Multi-mode data clustering method for automatically selecting clustering number |
| CN112016523A (en) * | 2020-09-25 | 2020-12-01 | 北京百度网讯科技有限公司 | Cross-modal face recognition method, device, equipment and storage medium |
| CN112016401A (en) * | 2020-08-04 | 2020-12-01 | 杰创智能科技股份有限公司 | Cross-modal-based pedestrian re-identification method and device |
| CN112200093A (en) * | 2020-10-13 | 2021-01-08 | 北京邮电大学 | Pedestrian re-identification method based on uncertainty estimation |
| CN112241682A (en) * | 2020-09-14 | 2021-01-19 | 同济大学 | End-to-end pedestrian searching method based on blocking and multi-layer information fusion |
| CN113128441A (en) * | 2021-04-28 | 2021-07-16 | 安徽大学 | System and method for identifying vehicle weight by embedding structure of attribute and state guidance |
| CN114550208A (en) * | 2022-02-10 | 2022-05-27 | 南通大学 | Cross-modal pedestrian re-identification method based on global level and local level combined constraint |
| WO2022147977A1 (en) * | 2021-01-05 | 2022-07-14 | 山东交通学院 | Vehicle re-identification method and system based on depth feature and sparse metric projection |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106778921A (en) * | 2017-02-15 | 2017-05-31 | 张烜 | Personnel based on deep learning encoding model recognition methods again |
| CN108345860A (en) * | 2018-02-24 | 2018-07-31 | 江苏测联空间大数据应用研究中心有限公司 | Personnel based on deep learning and learning distance metric recognition methods again |
| CN108960127A (en) * | 2018-06-29 | 2018-12-07 | 厦门大学 | Pedestrian's recognition methods again is blocked based on the study of adaptive depth measure |
-
2018
- 2018-12-12 CN CN201811515924.XA patent/CN109635728B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106778921A (en) * | 2017-02-15 | 2017-05-31 | 张烜 | Personnel based on deep learning encoding model recognition methods again |
| CN108345860A (en) * | 2018-02-24 | 2018-07-31 | 江苏测联空间大数据应用研究中心有限公司 | Personnel based on deep learning and learning distance metric recognition methods again |
| CN108960127A (en) * | 2018-06-29 | 2018-12-07 | 厦门大学 | Pedestrian's recognition methods again is blocked based on the study of adaptive depth measure |
Non-Patent Citations (2)
| Title |
|---|
| MANG YE等: "Visible Thermal Person Re-Identification via Dual-Constrained Top-Ranking", 《PROCEEDINGS OF THE TWENTY-SEVENTH INTERNATIONAL JOINT CONFERENCE ON ARTIFICIAL INTELLIGENCE》 * |
| SATEESH PEDAGADI等: "Local Fisher Discriminant Analysis for Pedestrian Re-identification", 《26TH IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111079939A (en) * | 2019-11-28 | 2020-04-28 | 支付宝(杭州)信息技术有限公司 | Machine learning model feature screening method and device based on data privacy protection |
| CN111079939B (en) * | 2019-11-28 | 2021-04-20 | 支付宝(杭州)信息技术有限公司 | Machine learning model feature screening method and device based on data privacy protection |
| CN112016401A (en) * | 2020-08-04 | 2020-12-01 | 杰创智能科技股份有限公司 | Cross-modal-based pedestrian re-identification method and device |
| CN112016401B (en) * | 2020-08-04 | 2024-05-17 | 杰创智能科技股份有限公司 | Cross-mode pedestrian re-identification method and device |
| CN112001438A (en) * | 2020-08-19 | 2020-11-27 | 四川大学 | Multi-mode data clustering method for automatically selecting clustering number |
| CN112001438B (en) * | 2020-08-19 | 2023-01-10 | 四川大学 | Multi-mode data clustering method for automatically selecting clustering number |
| CN112241682A (en) * | 2020-09-14 | 2021-01-19 | 同济大学 | End-to-end pedestrian searching method based on blocking and multi-layer information fusion |
| CN112241682B (en) * | 2020-09-14 | 2022-05-10 | 同济大学 | End-to-end pedestrian searching method based on blocking and multi-layer information fusion |
| CN112016523A (en) * | 2020-09-25 | 2020-12-01 | 北京百度网讯科技有限公司 | Cross-modal face recognition method, device, equipment and storage medium |
| CN112016523B (en) * | 2020-09-25 | 2023-08-29 | 北京百度网讯科技有限公司 | Cross-modal face recognition method, device, equipment and storage medium |
| CN112200093A (en) * | 2020-10-13 | 2021-01-08 | 北京邮电大学 | Pedestrian re-identification method based on uncertainty estimation |
| WO2022147977A1 (en) * | 2021-01-05 | 2022-07-14 | 山东交通学院 | Vehicle re-identification method and system based on depth feature and sparse metric projection |
| CN113128441A (en) * | 2021-04-28 | 2021-07-16 | 安徽大学 | System and method for identifying vehicle weight by embedding structure of attribute and state guidance |
| CN113128441B (en) * | 2021-04-28 | 2022-10-14 | 安徽大学 | System and method for identifying vehicle weight by embedding structure of attribute and state guidance |
| CN114550208A (en) * | 2022-02-10 | 2022-05-27 | 南通大学 | Cross-modal pedestrian re-identification method based on global level and local level combined constraint |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109635728B (en) | 2020-10-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109635728A (en) | A kind of isomery pedestrian recognition methods again based on asymmetric metric learning | |
| CN110781838B (en) | Multi-mode track prediction method for pedestrians in complex scene | |
| CN109472298B (en) | Deep bidirectional feature pyramid enhanced network for small-scale target detection | |
| CN110909690B (en) | Method for detecting occluded face image based on region generation | |
| CN109800710B (en) | Pedestrian re-identification system and method | |
| Yang et al. | Single image haze removal via region detection network | |
| CN108537824B (en) | Feature map enhanced network structure optimization method based on alternating deconvolution and convolution | |
| Huang et al. | A depth extraction method based on motion and geometry for 2D to 3D conversion | |
| CN109376696A (en) | Method, apparatus, computer equipment and the storage medium of video actions classification | |
| CN113887349A (en) | Road area image identification method based on image and point cloud fusion network | |
| CN111160295A (en) | Video pedestrian re-identification method based on region guidance and space-time attention | |
| CN116343330A (en) | Abnormal behavior identification method for infrared-visible light image fusion | |
| CN112686276A (en) | Flame detection method based on improved RetinaNet network | |
| CN112801019B (en) | Method and system for eliminating re-identification deviation of unsupervised vehicle based on synthetic data | |
| CN112766040B (en) | Method, device, apparatus and readable storage medium for detecting residual bait | |
| CN110022422A (en) | A kind of sequence of frames of video generation method based on intensive connection network | |
| Tang et al. | Smoking behavior detection based on improved YOLOv5s algorithm | |
| CN116342953A (en) | Dual-mode target detection model and method based on residual shrinkage attention network | |
| Lee et al. | Pedestrian detection based on deep fusion network using feature correlation | |
| CN114627269A (en) | Virtual reality security protection monitoring platform based on degree of depth learning target detection | |
| Ma et al. | Object-level proposals | |
| CN110321801B (en) | Clothing changing pedestrian re-identification method and system based on self-coding network | |
| Cong et al. | Enhancing nerf akin to enhancing llms: Generalizable nerf transformer with mixture-of-view-experts | |
| CN117036770A (en) | Detection model training and target detection method and system based on cascade attention | |
| Hu et al. | Vehicle detection technology based on cascading classifiers of multi-feature integration |
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 | ||
| OL01 | Intention to license declared |