CN107330396A - A kind of pedestrian's recognition methods again based on many attributes and many strategy fusion study - Google Patents
A kind of pedestrian's recognition methods again based on many attributes and many strategy fusion study Download PDFInfo
- Publication number
- CN107330396A CN107330396A CN201710504662.6A CN201710504662A CN107330396A CN 107330396 A CN107330396 A CN 107330396A CN 201710504662 A CN201710504662 A CN 201710504662A CN 107330396 A CN107330396 A CN 107330396A
- Authority
- CN
- China
- Prior art keywords
- pedestrian
- mrow
- attribute
- image
- msubsup
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- 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
-
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of pedestrian's recognition methods again based on many attributes and many strategy fusion study, the inventive method online lower training stage, selection is easily determined and pedestrian's attribute with enough discriminations first, pedestrian's attribute identifier is trained on attribute data collection, it is again that identification data collection marks attribute tags to pedestrian again with the attribute identifier, then in conjunction with attribute and pedestrian's identity label, using pedestrian's classification and the novel strategy for constraining contrast verification is merged, pedestrian's identification model again is trained;Inquiry phase on line, with pedestrian, identification model, respectively to query image and storehouse image zooming-out feature, calculates the Euclidean distance between query image feature and each storehouse characteristics of image, obtains closest image, it is believed that be the result that pedestrian recognizes again again.In aspect of performance, feature of the invention has ga s safety degree, achieves good accuracy rate;In terms of efficiency, the pedestrian that the present invention can be in pedestrian's image library represented by quick-searching to query image.
Description
Technical field
The invention belongs to mode identification technology, learned more particularly, to one kind based on many attributes and many strategy fusions
The pedestrian of habit recognition methods again.
Background technology
The attack of terrorism occurred again and again both at home and abroad in recent years is caused greatly to the security of the lives and property of the people of the world
Threat, as global anti-terrorism situation is increasingly severe, national governments are increasing to the input for safeguarding public safety.2005
Year, State Council has given an written reply " safe city " plan of Ministry of Public Security's proposition, and present China has had more than 600 cities and built energetically
" safe city ".Video monitoring system is a ring indispensable during " safe city " is built, and 20,000,000 prisons are had more than at present
Control camera covering and be arranged on national each public place.Each camera in video monitoring system is continuously being produced
Multitude of video data, and the monitoring range of single camera is limited, retrieval will often cross over multiple with following the trail of some specific objective
Camera.When case occurs, if only being analyzed with manpower monitoring content of multiple cameras before and after crime, both took
Effort lacks accuracy again.With the development of the technologies such as machine learning, deep learning, pattern-recognition, newest monitor video intelligence
Can analytical technology can handle monitor video while automatically analyze video in moving target.
Identification technology is important content in monitor video Intellectual Analysis Technology to the pedestrian of view-based access control model again, and main research exists
In the case of giving pedestrian's query image, how itself and the pedestrian image candidate storehouse from different cameras are subjected to feature and carried
Take and matched with analysis, judge which image belongs to same pedestrian with query image in image candidate storehouse.Pedestrian's identification technology again
Had broad application prospects in security protection, criminal investigation, searching fields such as population, human body behavioural analysis of wandering away, to safeguarding public safety
With important social effect and economic implications.Researcher proposed it is a variety of know method for distinguishing again for pedestrian, but very
It can be applied in practice less, because identification technology has lot of challenges to pedestrian again.Due to shooting condition limitation, monitoring camera
Head resolution ratio is different, and place environment is also different from position, and the different pedestrian images captured by same camera are in intensity of illumination, the back of the body
Than more consistent on scape, zoom degree;Due to the particularity of pedestrian itself, the posture walked often changes, self-contained article meeting
Generation is blocked, and also has block from each other, same a group traveling together captured by different cameras in posture, angle, block and respectively have
It is different.
Conventional method concentrates on two aspects, and one is the manual feature for being designed with ability to express, and another is that study has
The distance metric of distinction, but traditional-handwork feature representation is limited in one's ability, and also it is general for the distance metric method of manual feature
Change poor ability.The success that a variety of visual processing methods based on deep learning are obtained, recognizes that problem is provided again to solve pedestrian
Thinking, its performance is better than conventional method, because compare shallow-layer feature, depth characteristic extraction level is deeper, ability to express more
Good, resistance cosmetic variation robustness is stronger, based on convolutional neural networks (Convolutional Neural Network,
CNN feature extraction is both realized in recognition methods to pedestrian) in one network again, and distance metric is realized again, can be solved substantially point
The problem of influence such as resolution, intensity of illumination, zoom degree is caused.Correlative study is in ground zero, because data set size limit,
Contrast verification method is nearly all employed, it is high to loss function design requirement, have after large-scale dataset, there is a small amount of research to adopt
Use sorting technique.Pedestrian verifies and pedestrian's classification respectively has advantage and defect, and pedestrian's checking can make full use of the pass between image
Different pedestrians " pushing away " are obtained farther by system, and same a group traveling together " drawing " is obtained closer to pedestrian's classification can fully excavate the depth of image itself
Layer is semantic, and the two is complementary to a certain extent.It is existing to use the network structure for having multiple subnets, to based on two tuples and triple
Contrast verification network structure studied, each branch road complete validation task while also carry out classification based training, for checking
Task extracts feature more discriminatory, but this network structure when not solving test also must combination image input network and enter
The problem of row is compared, and the input construction difficulty of triple is big, and network effect depends on the quality of training sample unduly.Also adopt
Inputted with paired image, while the contrast verification between image and classification are carried out, but it contrasts the design of loss function excessively
Simply, it is difficult to tackle between feature the problem of distance is excessive to be impacted to whole loss function.It can be seen that, these are with reference to pedestrian point
Class and the method for pedestrian's contrast verification achieve certain effect promoting, but also still have some deficits.
Further, will not be with camera, background, pedestrian itself by adding sex, hair style, garment type, color etc.
Various change and the attribute information changed, can dissolve to a certain extent during pedestrian recognizes again by pedestrian itself posture, block,
The identification difficulty that the change such as angle is brought, Billy has more ga s safety degree with visual signature.In current large-scale data set
Mainly there is three-step approach with reference to attribute, learn jointly with inherent nature with reference to semantic attribute, be more focused on attribute forecast and process
In Optimization Work, and using the correlation joint modeling of many attribute, recognized again while carrying out attribute forecast and pedestrian, but
It is them by the way of mark pedestrian by hand again identification data collection, workload is too big, influences algorithm whole efficiency.
In summary, recognize that field has carried out substantial amounts of research work again in pedestrian at present, but there is table in existing method
The problems such as up to ability, to pedestrian's cosmetic variation poor robustness, structure and complex application, it is impossible to directly apply to monitoring and regard
In frequency intelligent analysis system.Therefore, how to design a kind of can have Shandong to the various influence factors for causing pedestrian's cosmetic variation
Rod, have higher recognition accuracy, and the not pedestrian of crash rate recognition methods again, be in monitor video intelligent analysis system urgently
To be solved the problem of.
The content of the invention
For the disadvantages described above or Improvement requirement of prior art, melted the invention provides one kind based on many attributes and many strategies
Pedestrian's recognition methods again of study is closed, its object is to select to easily determine and have pedestrian's attribute of enough discriminations first,
On attribute data collection train pedestrian's attribute identifier, then with the attribute identifier be pedestrian again identification data collection mark attribute mark
Label, then in conjunction with attribute and pedestrian's identity label, using fusion pedestrian classification and the strategy of novel constraint contrast verification, training
Pedestrian's identification model again;Inquiry phase on line, with pedestrian again identification model respectively to query image and storehouse image zooming-out feature,
Calculate the Euclidean distance between query image feature and each storehouse characteristics of image, obtain closest image, it is believed that be pedestrian again
The result of identification, this method can have robustness to many influence factors of pedestrian's cosmetic variation, have higher identification accurate
True rate, and not crash rate, meet the requirement of monitor video intelligent analysis system Real time identification pedestrian.
To achieve the above object, merged according to one aspect of the present invention there is provided one kind based on many attributes and many strategies
The pedestrian of study recognition methods again, methods described is divided into line lower training stage and on-line checking stage:
The line lower training stage specifically includes following steps:
(1) many attribute are chosen from pedestrian's attribute data collection, and is every kind of Attribute transposition classification, build attribute tags;
Selection is easily determined and the attribute with enough discriminations, including sex, hair length, upper part of the body pattern, upper half
This 6 kinds of body color, lower part of the body pattern, lower part of the body color, and be every kind of Attribute transposition classification;
(2) to pedestrian's Attribute Recognition data set, the convolutional neural networks model of a variety of attribute tags is built with, one is trained
Pedestrian's attribute identifier;
(3) it is that identification data collection marks attribute mark to pedestrian using " ballot method " again with the pedestrian's attribute identifier trained
Label, pedestrian's image that identification data is concentrated again is inputted after convolutional neural networks, and propagated forward calculates the value of each classification layer, wherein
Classification corresponding to the sequence number of greatest member is classification of this image to the attribute;For each attribute, with a group traveling together
Every image has one " ballot paper ", according to predicting the outcome as attribute classification " ballot " for every image, by " number of votes obtained " at most
Classification be used as the final label of the attribute;
(4) to pedestrian identification data collection and the attribute tags marked again, pedestrian's identity label, many attribute marks are built with
The convolutional neural networks model of label, integrated classification and constraint contrast verification, the model that one pedestrian of training recognizes again;
The on-line checking stage comprises the following steps:
S1, with the pedestrian trained identification model again, respectively to query image and the high-level characteristic of storehouse image zooming-out network,
One image is inputted after network, propagated forward calculates the value of last full articulamentum before division, feature needed for being, each
Image can all obtain a 4096 dimensional vector features;
S2, calculate Euclidean distance between query image feature and each storehouse characteristics of image, by obtained distance value from it is small to
Big sequence, the corresponding storehouse image of distance value in the top and query image are that the probability of same a group traveling together is larger, are taken closest
Multiple images be used as inquiry target.
Further, in the step of line lower training stage (2) as pedestrian's attribute identifier convolutional neural networks
The basic structure of model is VGGNet;Last full articulamentum of the convolutional neural networks model splits into 6;Full articulamentum
Classification layer afterwards also has 6, and 6 attribute labels are corresponded to respectively;Respectively there is one after each classification layer using Softmax losses
Classification Loss layer.
Further, in the step of line lower training stage (2) convolutional neural networks of pedestrian's attribute identifier instruction
White silk specifically includes following sub-step:
(21) 50 samples in a training batch are pre-processed;
(22) model crossed using large data collection ImageNet training in advance as network initial parameter, by training sample
Input after network, the value of each layer of propagated forward calculating network, until the Classification Loss layer of each attribute, the power of each Classification Loss value
Heavy phase etc.;
(23) if predetermined totality iterations is not up to, step (24) is continued;If having reached, terminate training;Wherein,
The span of the predetermined overall iterations is 10000 to 50000, preferably 50000;
(24) each layer parameter of network is successively reversely updated using gradient descent algorithm, while the classification for minimizing each attribute is damaged
Lose;Repeat step (21) to (23).
Further, in the step of line lower training stage (4) as pedestrian's identification model again convolutional neural networks
Model basic structure is CaffeNet;Last classification split layer of the convolutional neural networks corresponds to pedestrian's body respectively into 7
Part label and 6 attribute labels;Respectively there is a Classification Loss layer lost using Softmax after each classification layer;Except classification layer
Outside, also 7 dimensionality reduction layers;Respectively there is a constraint contrast verification loss for calculating distance between sample characteristics pair after each dimensionality reduction layer
Layer.
Further, in the step of line lower training stage (4) pedestrian's convolutional neural networks of identification model again instruction
White silk specifically includes following sub-step:
(41) 64 training samples in a training batch are pre-processed;
(42) initial parameter of convolutional neural networks, general are used as using the model that large data collection ImageNet training in advance is crossed
After sample input convolutional neural networks, propagated forward calculates the value of each layer of convolutional neural networks, until pedestrian's identity and each attribute
Classification Loss layer and constraint contrast verification loss layer, the penalty values of different labels and distinct methods have respective weight, pedestrian
The shared weight of identity loss is 3 times of each attribute loss, and weight shared by Classification Loss is constrain contrast verification loss 10 times;
(43) if predetermined totality iterations is not up to, step (44) is continued;If having reached, terminate training;Wherein,
The span of the predetermined overall iterations is 10000 to 50000, preferably 50000;
(44) each layer parameter of network is successively reversely updated using gradient descent algorithm, while minimizing pedestrian's identity and each category
Property Classification Loss and constraint contrast verification loss;Repeat step (41) to (43).
Further, the expression-form of constraint contrast verification loss function is in the step (42):
Wherein, j=0 represents pedestrian's identity, 6 attribute of the correspondence pedestrians of j=1,2 ..., 6;Represent the training batch
In m-th of characteristics of image pairWithWhether same a group traveling together is belonged to or with same alike result, 1 represents it is that 0 represents no;Formula
Section 1 punish with the excessive situation of distance between a group traveling together or the feature pair of same alike result, the distance metric side used here
Formula, wherein,It is L2 norms, that is, Euclidean distance;Section 2 punish the feature of different pedestrians or different attribute to spacing
From too small situation, θ represents boundary threshold parameter, and θ span is 100 to 300, preferably 200, for preventing distance
The excessive feature pair that peels off;Section 3 is bound term, | | | |1It is L1 norms, its value is vectorial every absolute value sum, this
The absolute value of characteristic value each single item is all leveled off to 1 as far as possible, can normalization characteristic, characteristic distance can be made again predictable
Within the scope of;β is weight shared by bound term, β=0.01;
The average constraint contrast verification loss function of the batch is expressed as:
Wherein, M is the feature that can constitute of sample of a training batch to quantity, M=2016.
In general, by the contemplated above technical scheme of the present invention compared with prior art, it is special with following technology
Levy and beneficial effect:
(1) the technical program learns pedestrian's identity information and attribute information simultaneously, and main innovation point is to trained one
Attribute identifier, can be pedestrian's identification data collection mark attribute tags again automatically, save the time;Attribute information is to cosmetic variation
With robustness, for aiding in pedestrian to recognize again, solve to a certain extent by pedestrian itself posture, block, the change such as angle
The identification problem brought.
(2) the technical program proposes a kind of constraint contrast verification loss function, for binding characteristic value, by characteristic distance
Within the specific limits, in one network, checking target and the optimization of class object Simultaneous Iteration can both learn Geng You areas for limitation
The feature of point property, measures the relation between pedestrian again, solve to a certain extent resolution ratio, intensity of illumination, zoom degree etc. by
The problem of shooting condition causes visible change.
(3) the technical program employs a kind of many attributes and shifty solution, in one end to end framework
The Classification Loss, the Classification Loss of the loss of constraint contrast verification and attribute, constraint contrast verification loss of pedestrian's identity are merged, anti-
To the weighting sum that each loss is propagated when propagating.Merging many attribute and a variety of tactful methods can make each several part complementary, enter
One step improves pedestrian's discrimination again.
Brief description of the drawings
Fig. 1 is the flow frame diagram of the inventive method;
Fig. 2 is the convolutional neural networks structure in the inventive method as pedestrian's attribute identifier;
Fig. 3 be the inventive method in as pedestrian's attribute identifier convolutional neural networks training flow chart;
Fig. 4 is the convolutional neural networks structure as pedestrian's identification model again in the inventive method;
Fig. 5 is the training flow chart as pedestrian's convolutional neural networks of identification model again in the inventive method;
Fig. 6 be the method C-CNN based on pedestrian's identities, based on pedestrian's identity constrain contrast verification method V-CNN,
With reference to pedestrian's identities and method CV-CNN, method ATTR-CNN, the base based on pedestrian's attributive classification of constraint contrast verification
Method ATTR+C-CNN, the present invention in pedestrian's identity and attributive classification merges the method learnt based on many attributes and many strategies
ATTR+CV-CNN cumulative matches characteristic curve compares figure.
Embodiment
In order to make the purpose , technical scheme and advantage of the present invention be clearer, it is right below in conjunction with drawings and Examples
The present invention is further elaborated.It should be appreciated that specific embodiment described herein is only to explain the present invention, not
For limiting the present invention.As long as in addition, technical characteristic involved in each embodiment of invention described below that
Not constituting conflict between this can just be mutually combined.
The flow of the inventive method is as shown in Figure 1.Online lower training stage, first selection easily determine and with enough areas
Pedestrian's attribute of indexing, carries out building convolutional neural networks model on rational category division, humanized identification data collection of being expert at,
Train pedestrian's attribute identifier, then with the attribute identifier be pedestrian again identification data collection mark attribute tags, then in conjunction with category
Property and pedestrian's identity label, using fusion pedestrian classification and constraint contrast verification strategy, the structure on pedestrian again identification data collection
Convolutional neural networks model is built, pedestrian's identification model again is trained;Inquiry phase on line, with pedestrian again identification model respectively to looking into
Image and storehouse image zooming-out feature are ask, the Euclidean distance between query image feature and each storehouse characteristics of image is calculated, by what is obtained
Distance value is sorted from small to large, some images closest with query image in the image of storehouse, it is believed that be inquiry target, and ranking is leaned on
The corresponding storehouse image of preceding distance value and query image are that the probability of same a group traveling together is larger.
Pedestrian's attribute identifier training process of line lower training stage is specifically described below, pedestrian again trained by identification model
Journey, and the pedestrian in on-line checking stage recognize the specific implementation step of application process again.
The line lower training stage comprises the following steps:
(1) in many attribute provided from existing large-scale pedestrian's Attribute Recognition data set PETA, selection is easily determined and had
There are 6 attributes and its category division of enough discriminations;Including:1. sex:It is male, women, uncertain;2. hair length:It is long
It is hair, bob, uncertain;3. upper part of the body pattern:T-shirt, shirt, overcoat, down jackets, western-style clothes, other, it is uncertain;4. upper part of the body face
Color:It is black, white, red, yellow, blue, green, purple, brown, grey, orange, polychrome, uncertain;5. lower part of the body pattern:It is trousers, shorts, longuette, short
Skirt, other, it is uncertain;6. lower part of the body color:It is black, white, red, yellow, blue, green, purple, brown, grey, orange, polychrome, uncertain.
(2) to pedestrian's Attribute Recognition data set, the convolutional neural networks model of a variety of attribute tags is built with, one is trained
Pedestrian's attribute identifier.
Convolutional neural networks model basic structure used is VGGNet;As shown in Fig. 2 the network model has 5 groups of convolution, volume
Product core is respectively set to:1. 64 (3 × 3), 64 (3 × 3), 2. 128 (3 × 3), 128 (3 × 3), 3. 256 (3 × 3), 256 (3 ×
3), 256 (3 × 3), 256 (3 × 3), 4. 512 (3 × 3), 512 (3 × 3), 512 (3 × 3), 512 (3 × 3), 5. 512 (3 × 3),
512 (3 × 3), 512 (3 × 3), 512 (3 × 3);Full articulamentum fc6 has 4096 nodes, and full articulamentum fc7 splits into 6,
Fc7_1, fc7_2, fc7_3, fc7_4, fc7_5, fc7_6, respectively there is 2048 nodes;Classification layer after full articulamentum also has 6
It is individual, 6 attribute tags are corresponded to respectively, and the other classification layer fc8_1 of correspondence has 3 nodes, the classification layer of correspondence hair length
Fc8_2 has 3 nodes, and the classification layer fc8_3 of correspondence upper part of the body pattern has 7 nodes, the classification layer of correspondence upper part of the body color
Fc8_4 has 12 nodes, and the classification layer fc8_5 of correspondence lower part of the body pattern has 6 nodes, the classification layer of correspondence lower part of the body color
Fc8_6 has 12 nodes;Respectively there is a Classification Loss layer lost using Softmax after each classification layer.
As shown in figure 3, being as the training step of the convolutional neural networks of pedestrian's attribute identifier:
(21) 50 training samples in a training batch are pre-processed, scale the images to 256*256 sizes,
The image block of 224*224 sizes is respectively cut in the upper left corner, the upper right corner, the lower left corner, the lower right corner, center, is become with reference to horizon glass picture
The mode changed, 10 are extended for by an original image, the subtracted image average before input network;It is provided under gradient
The momentum that drop accelerates is 0.9, and weight attenuation parameter is 0.0005, and the initial learning rate of fine setting is set to 0.001, per iteration 10000
Secondary learning rate be reduced to before 0.1 times;
(22) model crossed using large data collection ImageNet training in advance inputs sample as the initial parameter of network
After network, the value of each layer of propagated forward calculating network;For jth, j=1,2 ..., 6 attributes, each attribute has K(j)It is individual
Classification, the feature input of n-th of samplePredicted valueProbability for classification k is expressed as follows:
N-th of sample be for attribute j Classification Loss function:
The N number of sample of the batch is for attribute j average classification penalty values:
(23) if making a reservation for totality iterations 50000 times not up to, step (24) is continued;If having reached, terminate instruction
Practice;
(24) each layer parameter of network is successively reversely updated using gradient descent algorithm, obtains attribute j's according to below equation
The gradient of Classification Loss backpropagation:
Minimize the Classification Loss of each attribute, repeat step (21)~(23) simultaneously according to gradient.
(3) it is that identification data collection Market-1501 uses " ballot method " to pedestrian again with the pedestrian's attribute identifier trained
Attribute tags are marked, an image is inputted after network, propagated forward calculates the sequence number of the value, wherein greatest member of each classification layer
Corresponding classification is classification of this image on the attribute, for attribute j, pedestrian image is inputted after network, forward direction is passed
That broadcasts calculating fc8_j is worth to vectorIt belongs to each classification k probabilityIt is also a vector, wherein greatest member
Sequence number corresponding to classification be attribute j classification, i.e.,For each attribute, with the every of a group traveling together
Image has one " ballot paper ", according to predicting the outcome as attribute classification " ballot " for every image, by " number of votes obtained " at most
Classification is used as the final label of the attribute;
(4) to pedestrian identification data collection and the attribute tags marked again, pedestrian's identity label and many attribute are built with
The convolutional neural networks model of label, integrated classification and constraint contrast verification, trains pedestrian identification model again;
The CaffeNet that convolutional neural networks model basic structure used is;As shown in figure 4, the convolution kernel difference of convolutional layer
It is set to:96 (11 × 11), 256 (5 × 5), 384 (3 × 3), 384 (3 × 3), 256 (3 × 3);Full articulamentum fc6 and fc7 are each
There are 4096 nodes;Classification layer has 7, and pedestrian's identity label and 6 attribute labels, the classification of correspondence pedestrian's identity are corresponded to respectively
Layer fc8_0 has 751 nodes, and the other classification layer fc8_1 of correspondence has 3 nodes, and the classification layer fc8_2 of correspondence hair length has
3 nodes, the classification layer fc8_3 of correspondence upper part of the body pattern has 7 nodes, and the classification layer fc8_4 of correspondence upper part of the body color has 12
Individual node, the classification layer fc8_5 of correspondence lower part of the body pattern has 6 nodes, and the classification layer fc8_6 of correspondence lower part of the body color has 12
Node;Respectively there is a Classification Loss layer lost using Softmax after each classification layer;Pedestrian's identity, sex, head are corresponded to respectively
Hair length, upper part of the body pattern, upper part of the body color, lower part of the body pattern, dimensionality reduction layer ip_0~ip_6 of lower part of the body color respectively have 100
Node;Respectively there is a constraint contrast verification loss layer being calculated as to distance between sample characteristics after each dimensionality reduction layer.
As shown in figure 5, being as the training step of pedestrian's convolutional neural networks of identification model again:
(41) 64 training samples in a training batch are pre-processed, pedestrian image is zoomed to 227 × 227
Size, only with the mode of horizontal mirror transformation slightly EDS extended data set, use the mode of removal mean picture brightness for
Each image subtracted image average;Training set is divided into multiple batches with checking collection during training, by one batch during each iteration
In secondary training sample input network;It is 0.9 to be provided for gradient to decline the momentum accelerated, and weight attenuation parameter is 0.0005,
The initial learning rate of fine setting is set to 0.001,0.1 times before being reduced to per 20000 learning rates of iteration;
(42) model crossed using large data collection ImageNet training in advance inputs sample as the initial parameter of network
After network, the value of each layer of propagated forward calculating network;For jth, j=0,1,2 ..., 6 attribute represents pedestrian as j=0
Identity, corresponding 6 attribute in j=1~6, each attribute has K(j)Individual classification, the feature input of n-th of samplePredicted valueProbability for classification k is expressed as follows:
The Classification Loss function of n-th of sample is:
64 samples of the batch are for attribute j average classification penalty values:
M-th of characteristics of image pairConstraint contrast verification loss function be:
Wherein,Represent m-th of characteristics of image pairWhether same a group traveling together is belonged to or with same alike result, 1
Expression is that 0 represents no;The Section 1 of formula punished with the excessive situation of distance between a group traveling together or the feature pair of same alike result,
Here the distance metric mode usedIt is L2 norms, that is, Euclidean distance;Section 2 punishes different pedestrians or different attribute
Feature pair between the too small situation of distance, be provided with a boundary threshold parameter θ here, θ=200, for prevent away from
From the excessive feature pair that peels off;Section 3 is bound term, | | | |1It is L1 norms, its value is vectorial every absolute value sum, should
Item makes the absolute value of characteristic value each single item all level off to 1 as far as possible, can normalization characteristic, can make again characteristic distance it is contemplated that
Within the scope of, β is weight shared by bound term, β=0.01.
M is the feature that can constitute of sample of a training batch to quantity, M=2016, the average constraint contrast of the batch
Checking loss function is expressed as:
The penalty values of different labels and distinct methods have respective weight, and the shared weight of pedestrian's identity loss is that each attribute is damaged
3 times lost, weight shared by Classification Loss is constrain contrast verification loss 10 times;
(43) if predetermined totality iterations, 50000 times not up to, then continue step (44);If having reached, terminate instruction
Practice;
(44) each layer parameter of network is successively reversely updated using gradient descent algorithm, obtains attribute j's according to equation below
The gradient of Classification Loss backpropagation is:
T1, t2, t3 are expressed as by three constrained in contrast verification loss function, derivation is as follows respectively:
Wherein, i value is 1 or 2,Indicator function, population gradient for this three
Item gradient sum:
Minimize Classification Loss and the constraint contrast verification loss of pedestrian's identity and each attribute simultaneously according to gradient;Repeat to walk
Suddenly (41)~(43).
S1, with the pedestrian trained identification model again, respectively to query image and the high level of fc7 layers of storehouse image zooming-out network
Feature;By in image input pedestrian again identification model, propagated forward is successively calculated, until last full articulamentum before division
Fc7, the value of this layer of each node is required feature, and each image will obtain the vector of one 4096 dimension;
Euclidean distance between S2, calculating query image feature and each storehouse characteristics of image
Wherein x1And x2It is the characteristic vector of query image and some storehouse image respectively, i is the subscript index of vector;By obtained distance
Value sorts from small to large, and the corresponding storehouse image of distance value in the top and query image are that the probability of same a group traveling together is larger, are taken
Several closest images are as a result.
Example:
In order to prove that performance and efficiency based on many attributes and many strategy fusion learning methods have advantage, the present invention passes through
Verified and analyzed below.
A. experimental data
The present invention is tested using Market-1501 data sets, and the dataset acquisition is from one in Tsinghua Campus
Individual supermarket doorway, one has the pedestrian of 1501 different identities;Data set divided training picture and candidate's picture, training
There are 751 pedestrian's identity, 12,936 pictures, in training with 9 in picture:1 ratio random division is training set and checking
Collection, has 11,642 and 1,294 pictures, Candidate Set and inquiry are concentrated with 750 pedestrian ID, there is 19,732 and 3 respectively respectively,
368 pictures;Picture format is JPEG, and image size is 64*128.
B. experiment porch
Hardware:CPU Intel Xeon E5-2650 v3, internal memory 64G DDR4 2133MHz, GPU GeForce GTX
TITAN X, video memory 12G;
Software:Operating system Ubuntu 15.04 64, experiment porch Caffe, Matlab R2014a.
C. pedestrian recognizes criteria of quality evaluation again
Cumulative matches characteristic the first accuracy rate CMC@1 calculate its feature and Candidate Set first for each query image
In the distance between all characteristics of image, these distances are sorted from low to high, the image ranked the first in Candidate Set is returned, should
Image belongs to same a group traveling together with query image and correctly matched, the percentage that statistical query collection image is correctly matched, such as formulaShown, wherein query set size is m, to query image pi, it is assumed that first is correct
The image of matching is qpi, its position in the ranking is designated as r (qpi)。
Average Accuracy average mAP is the average of all image averaging accuracy rate AP in query set, and AP computational methods areMAP computational methods areAssuming that to query graph
As piN image is returned altogether, and R is to return to all correct picture numbers, R in imagejIt is correct picture number in preceding j image,
Query set size is similarly m.
D. experimental result
Experiment shows that the inventive method is on Market-1501 data sets, and CMC@1 have reached that 70.0%, mAP reaches
45.7%, many top methods are had been over, in test, feature is extracted and takes around 4.82 seconds/100 images, calculate
Distance takes around 0.11 second/100 images with ranking, can meet the demand of practical application.
Method V-CNN, the knot of contrast verification are constrained by the method C-CNN based on pedestrian's identities, based on pedestrian's identity
Close pedestrian's identities and the constraint method CV-CNN of contrast verification, the method ATTR-CNN based on pedestrian's attributive classification, be based on
Method ATTR+C-CNN, the of the invention method ATTR that based on many attributes and many strategy merges study of pedestrian's identity with attributive classification
+ CV-CNN is compared, as a result as shown in Figure 6:Transverse axis is to calculate query image feature and all characteristics of image in Candidate Set
The distance between and the amount of images that returns after sorting from low to high, the longitudinal axis is the percentage that query set image is correctly matched.
This method more has the feature of ability to express due to that can extract compared with without the method using deep learning,
The problem of smaller visible changes such as intensity of illumination, resolution ratio, translation scaling, attitudes vibration are caused is solved to a certain extent,
Substantially increase classification accuracy;Compared with the method for employing deep learning, because attribute is compared to low-level feature more language
Justice expression performance is more preferable and insensitive to the larger cosmetic variation such as shooting angle, background change, partial occlusion so that algorithm
Recognition effect it is more excellent.
As it will be easily appreciated by one skilled in the art that the foregoing is merely illustrative of the preferred embodiments of the present invention, it is not used to
The limitation present invention, any modifications, equivalent substitutions and improvements made within the spirit and principles of the invention etc., it all should include
Within protection scope of the present invention.
Claims (6)
1. a kind of pedestrian's recognition methods again based on many attributes and many strategy fusion study, it is characterised in that methods described is divided into
Line lower training stage and on-line checking stage:
The line lower training stage specifically includes following steps:
(1) many attribute are chosen from pedestrian's attribute data collection, and is every kind of Attribute transposition classification, build attribute tags;
(2) to pedestrian's Attribute Recognition data set, the convolutional neural networks model of a variety of attribute tags is built with, a pedestrian is trained
Attribute identifier;
(3) it is that identification data collection marks attribute tags to pedestrian using " ballot method " again with the pedestrian's attribute identifier trained, will
After pedestrian's image input convolutional neural networks that identification data is concentrated again, propagated forward calculates the value of each classification layer, wherein maximum
Classification corresponding to the sequence number of element is classification of this image to the attribute;For each attribute, with every of a group traveling together
Image has one " ballot paper ", according to predicting the outcome as attribute classification " ballot " for every image, by " number of votes obtained " most class
Not as the final label of the attribute;
(4) to pedestrian identification data collection and the attribute tags marked again, be built with pedestrian's identity label, a variety of attribute tags,
The convolutional neural networks model of integrated classification and constraint contrast verification, the model that one pedestrian of training recognizes again;
The on-line checking stage comprises the following steps:
S1, with the pedestrian trained identification model again, respectively to query image and the high-level characteristic of storehouse image zooming-out network, by one
Open image to input after network, propagated forward calculates the value of last full articulamentum before division, as required feature, each image
A vector characteristics will be obtained;
Euclidean distance between S2, calculating query image feature and each storehouse characteristics of image, obtained distance value is arranged from small to large
Sequence, the corresponding storehouse image of distance value in the top and query image are that the probability of same a group traveling together is larger, are taken closest many
Open image and be used as inquiry target.
2. a kind of pedestrian's recognition methods again based on many attributes and many strategy fusion study according to claim 1, it is special
Levy and be, the step of the line lower training stage in (2) as pedestrian's attribute identifier convolutional neural networks model it is basic
Structure is VGGNet;Last full articulamentum of the convolutional neural networks model splits into multiple;Point after full articulamentum
Class layer also has multiple, and a variety of attribute tags are corresponded to respectively;Respectively there is one to be damaged using the Softmax classification lost after each classification layer
Lose layer.
3. a kind of pedestrian's recognition methods again based on many attributes and many strategy fusion study according to claim 1, it is special
Levy and be, the training of the convolutional neural networks of pedestrian's attribute identifier is specifically included in (2) the step of the line lower training stage
Following sub-step:
(21) sample in a training batch is pre-processed;
(22) model crossed using large data collection ImageNet training in advance inputs training sample as the initial parameter of network
After network, the value of each layer of propagated forward calculating network, until the Classification Loss layer of each attribute, the weight phase of each Classification Loss value
Deng;
(23) if predetermined totality iterations is not up to, step (24) is continued;If having reached, terminate training;
(24) each layer parameter of network is successively reversely updated using gradient descent algorithm, while minimizing the Classification Loss of each attribute;
Repeat step (21) to (23).
4. a kind of pedestrian's recognition methods again based on many attributes and many strategy fusion study according to claim 1, it is special
Levy and be, tied substantially as the convolutional neural networks model of pedestrian's identification model again in (4) the step of the line lower training stage
Structure is CaffeNet;Last classification split layer of the convolutional neural networks corresponded into multiple, respectively pedestrian's identity label and
A variety of attribute tags;Respectively there is a Classification Loss layer lost using Softmax after each classification layer;In addition to layer of classifying, also have
Multiple dimensionality reduction layers;Respectively there is a constraint contrast verification loss layer for calculating distance between sample characteristics pair after each dimensionality reduction layer.
5. a kind of pedestrian's recognition methods again based on many attributes and many strategy fusion study according to claim 1, it is special
Levy and be, the training of pedestrian's convolutional neural networks of identification model again is specifically included in (4) the step of the line lower training stage
Following sub-step:
(41) training sample in a training batch is pre-processed;
(42) model crossed using large data collection ImageNet training in advance as convolutional neural networks initial parameter, by sample
Input after convolutional neural networks, propagated forward calculates the value of each layer of convolutional neural networks, until point of pedestrian's identity and each attribute
Class loss layer and constraint contrast verification loss layer, the penalty values of different labels and distinct methods have respective weight, pedestrian's identity
The shared weight of loss is 3 times of each attribute loss, and weight shared by Classification Loss is constrain contrast verification loss 10 times;
(43) if predetermined totality iterations is not up to, step (44) is continued;If having reached, terminate training;
(44) each layer parameter of network is successively reversely updated using gradient descent algorithm, while minimizing pedestrian's identity and each attribute
Classification Loss and constraint contrast verification loss;Repeat step (41) to (43).
6. a kind of pedestrian's recognition methods again based on many attributes and many strategy fusion study according to claim 5, it is special
Levy and be, the expression-form of constraint contrast verification loss function is in the step (42):
<mrow>
<mtable>
<mtr>
<mtd>
<mrow>
<msubsup>
<mi>L</mi>
<mrow>
<mi>A</mi>
<mi>T</mi>
<mi>T</mi>
<mi>R</mi>
<mo>+</mo>
<mi>V</mi>
<mo>_</mo>
<mi>m</mi>
</mrow>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mo>)</mo>
</mrow>
</msubsup>
<mo>=</mo>
<mfrac>
<mn>1</mn>
<mn>2</mn>
</mfrac>
<msubsup>
<mi>s</mi>
<mi>m</mi>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mo>)</mo>
</mrow>
</msubsup>
<mo>|</mo>
<mo>|</mo>
<msubsup>
<mi>x</mi>
<mrow>
<mi>m</mi>
<mo>,</mo>
<mn>1</mn>
</mrow>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mo>)</mo>
</mrow>
</msubsup>
<mo>-</mo>
<msubsup>
<mi>x</mi>
<mrow>
<mi>m</mi>
<mo>,</mo>
<mn>2</mn>
</mrow>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mo>)</mo>
</mrow>
</msubsup>
<mo>|</mo>
<msubsup>
<mo>|</mo>
<mn>2</mn>
<mn>2</mn>
</msubsup>
<mo>+</mo>
<mfrac>
<mn>1</mn>
<mn>2</mn>
</mfrac>
<mi>max</mi>
<mrow>
<mo>(</mo>
<mi>&theta;</mi>
<mo>-</mo>
<mo>|</mo>
<mo>|</mo>
<msubsup>
<mi>x</mi>
<mrow>
<mi>m</mi>
<mo>,</mo>
<mn>1</mn>
</mrow>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mo>)</mo>
</mrow>
</msubsup>
<mo>-</mo>
<msubsup>
<mi>x</mi>
<mrow>
<mi>m</mi>
<mo>,</mo>
<mn>2</mn>
</mrow>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mo>)</mo>
</mrow>
</msubsup>
<mo>|</mo>
<msubsup>
<mo>|</mo>
<mn>2</mn>
<mn>2</mn>
</msubsup>
<mo>,</mo>
<mn>0</mn>
<mo>)</mo>
</mrow>
</mrow>
</mtd>
</mtr>
<mtr>
<mtd>
<mrow>
<mo>+</mo>
<mi>&beta;</mi>
<mrow>
<mo>(</mo>
<mo>|</mo>
<mo>|</mo>
<mo>|</mo>
<msubsup>
<mi>x</mi>
<mrow>
<mi>m</mi>
<mo>,</mo>
<mn>1</mn>
</mrow>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mo>)</mo>
</mrow>
</msubsup>
<mo>|</mo>
<mo>-</mo>
<mn>1</mn>
<mo>|</mo>
<msub>
<mo>|</mo>
<mn>1</mn>
</msub>
<mo>-</mo>
<mo>|</mo>
<mo>|</mo>
<mo>|</mo>
<msubsup>
<mi>x</mi>
<mrow>
<mi>m</mi>
<mo>,</mo>
<mn>2</mn>
</mrow>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mo>)</mo>
</mrow>
</msubsup>
<mo>|</mo>
<mo>-</mo>
<mn>1</mn>
<mo>|</mo>
<msub>
<mo>|</mo>
<mn>1</mn>
</msub>
<mo>)</mo>
</mrow>
</mrow>
</mtd>
</mtr>
</mtable>
<mo>,</mo>
</mrow>
Wherein, j=0 represents pedestrian's identity, j=1, the n attribute of 2 ..., n correspondence pedestrians;Represent m in the training batch
Individual characteristics of image pairWithWhether same a group traveling together is belonged to or with same alike result, 1 represents it is that 0 represents no;It is L2 models
Number, that is, Euclidean distance;θ represents boundary threshold parameter;||·||1It is L1 norms;β is weight shared by bound term;
The average constraint contrast verification loss function of the batch is expressed as:
<mrow>
<msubsup>
<mi>L</mi>
<mrow>
<mi>A</mi>
<mi>T</mi>
<mi>T</mi>
<mi>R</mi>
<mo>+</mo>
<mi>V</mi>
</mrow>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mo>)</mo>
</mrow>
</msubsup>
<mo>=</mo>
<mfrac>
<mn>1</mn>
<mi>M</mi>
</mfrac>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>m</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>M</mi>
</munderover>
<msubsup>
<mi>L</mi>
<mrow>
<mi>A</mi>
<mi>T</mi>
<mi>T</mi>
<mi>R</mi>
<mo>+</mo>
<mi>V</mi>
<mo>_</mo>
<mi>m</mi>
</mrow>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mo>)</mo>
</mrow>
</msubsup>
<mo>,</mo>
</mrow>
Wherein, M is the feature that can constitute of sample of a training batch to quantity.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710504662.6A CN107330396B (en) | 2017-06-28 | 2017-06-28 | Pedestrian re-identification method based on multi-attribute and multi-strategy fusion learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710504662.6A CN107330396B (en) | 2017-06-28 | 2017-06-28 | Pedestrian re-identification method based on multi-attribute and multi-strategy fusion learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN107330396A true CN107330396A (en) | 2017-11-07 |
CN107330396B CN107330396B (en) | 2020-05-19 |
Family
ID=60198291
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710504662.6A Active CN107330396B (en) | 2017-06-28 | 2017-06-28 | Pedestrian re-identification method based on multi-attribute and multi-strategy fusion learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107330396B (en) |
Cited By (60)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107944366A (en) * | 2017-11-16 | 2018-04-20 | 山东财经大学 | A kind of finger vein identification method and device based on attribute study |
CN108108674A (en) * | 2017-12-08 | 2018-06-01 | 浙江捷尚视觉科技股份有限公司 | A kind of recognition methods again of the pedestrian based on joint point analysis |
CN108108754A (en) * | 2017-12-15 | 2018-06-01 | 北京迈格威科技有限公司 | The training of identification network, again recognition methods, device and system again |
CN108229503A (en) * | 2018-01-04 | 2018-06-29 | 浙江大学 | A kind of feature extracting method for clothes photo |
CN108229543A (en) * | 2017-12-22 | 2018-06-29 | 中国科学院深圳先进技术研究院 | Image classification design methods and device |
CN108229398A (en) * | 2018-01-04 | 2018-06-29 | 中科汇通投资控股有限公司 | A kind of face verification method of self-teaching |
CN108416295A (en) * | 2018-03-08 | 2018-08-17 | 天津师范大学 | A kind of recognition methods again of the pedestrian based on locally embedding depth characteristic |
CN108537136A (en) * | 2018-03-19 | 2018-09-14 | 复旦大学 | The pedestrian's recognition methods again generated based on posture normalized image |
CN108595558A (en) * | 2018-04-12 | 2018-09-28 | 福建工程学院 | A kind of image labeling method of data balancing strategy and multiple features fusion |
CN108647577A (en) * | 2018-04-10 | 2018-10-12 | 华中科技大学 | A kind of pedestrian's weight identification model that adaptive difficult example is excavated, method and system |
CN108764065A (en) * | 2018-05-04 | 2018-11-06 | 华中科技大学 | A kind of method of pedestrian's weight identification feature fusion assisted learning |
CN108846413A (en) * | 2018-05-21 | 2018-11-20 | 复旦大学 | A kind of zero sample learning method based on global semantic congruence network |
CN108875765A (en) * | 2017-11-14 | 2018-11-23 | 北京旷视科技有限公司 | Method, apparatus, equipment and the computer storage medium of EDS extended data set |
CN108921054A (en) * | 2018-06-15 | 2018-11-30 | 华中科技大学 | A kind of more attribute recognition approaches of pedestrian based on semantic segmentation |
CN108921022A (en) * | 2018-05-30 | 2018-11-30 | 腾讯科技(深圳)有限公司 | A kind of human body attribute recognition approach, device, equipment and medium |
CN108960331A (en) * | 2018-07-10 | 2018-12-07 | 重庆邮电大学 | A kind of recognition methods again of the pedestrian based on pedestrian image feature clustering |
CN108960184A (en) * | 2018-07-20 | 2018-12-07 | 天津师范大学 | A kind of recognition methods again of the pedestrian based on heterogeneous components deep neural network |
CN109033154A (en) * | 2018-06-12 | 2018-12-18 | 佛山欧神诺陶瓷有限公司 | A kind of category management method |
CN109101866A (en) * | 2018-06-05 | 2018-12-28 | 中国科学院自动化研究所 | Pedestrian recognition methods and system again based on segmentation outline |
CN109165306A (en) * | 2018-08-09 | 2019-01-08 | 长沙理工大学 | Image search method based on the study of multitask Hash |
CN109190120A (en) * | 2018-08-31 | 2019-01-11 | 第四范式(北京)技术有限公司 | Neural network training method and device and name entity recognition method and device |
CN109190646A (en) * | 2018-06-25 | 2019-01-11 | 北京达佳互联信息技术有限公司 | A kind of data predication method neural network based, device and nerve network system |
CN109214308A (en) * | 2018-08-15 | 2019-01-15 | 武汉唯理科技有限公司 | A kind of traffic abnormity image identification method based on focal loss function |
CN109271898A (en) * | 2018-08-31 | 2019-01-25 | 电子科技大学 | Solution cavity body recognizer based on optimization convolutional neural networks |
CN109325547A (en) * | 2018-10-23 | 2019-02-12 | 苏州科达科技股份有限公司 | Non-motor vehicle image multi-tag classification method, system, equipment and storage medium |
CN109522855A (en) * | 2018-11-23 | 2019-03-26 | 广州广电银通金融电子科技有限公司 | In conjunction with low resolution pedestrian detection method, system and the storage medium of ResNet and SENet |
CN109635634A (en) * | 2018-10-29 | 2019-04-16 | 西北大学 | A kind of pedestrian based on stochastic linear interpolation identifies data enhancement methods again |
CN109711354A (en) * | 2018-12-28 | 2019-05-03 | 哈尔滨工业大学(威海) | A kind of method for tracking target indicating study based on video attribute |
CN109711281A (en) * | 2018-12-10 | 2019-05-03 | 复旦大学 | A kind of pedestrian based on deep learning identifies again identifies fusion method with feature |
CN109740480A (en) * | 2018-12-26 | 2019-05-10 | 浙江捷尚视觉科技股份有限公司 | A kind of identified again based on non-motor vehicle pedestrian retrieval method under monitoring scene |
CN109754005A (en) * | 2018-12-25 | 2019-05-14 | 任飞翔 | Dynamic increase and decrease method and device |
CN109766790A (en) * | 2018-12-24 | 2019-05-17 | 重庆邮电大学 | A kind of pedestrian detection method based on self-adaptive features channel |
CN109784288A (en) * | 2019-01-22 | 2019-05-21 | 天津师范大学 | A kind of pedestrian's recognition methods again based on differentiation perception fusion |
CN109815908A (en) * | 2019-01-25 | 2019-05-28 | 同济大学 | It is a kind of based on the discrimination method again of the pedestrian that measures between deep learning and overlapping image block |
CN109919073A (en) * | 2019-03-01 | 2019-06-21 | 中山大学 | A kind of recognition methods again of the pedestrian with illumination robustness |
CN109919166A (en) * | 2017-12-12 | 2019-06-21 | 杭州海康威视数字技术股份有限公司 | The method and apparatus for obtaining the classification information of attribute |
CN110008913A (en) * | 2019-04-08 | 2019-07-12 | 南京工业大学 | Pedestrian re-identification method based on fusion of attitude estimation and viewpoint mechanism |
CN110046553A (en) * | 2019-03-21 | 2019-07-23 | 华中科技大学 | A kind of pedestrian weight identification model, method and system merging attributive character |
CN110096346A (en) * | 2019-03-29 | 2019-08-06 | 广州思德医疗科技有限公司 | A kind of training mission processing method and processing device of more calculate nodes |
CN110163041A (en) * | 2018-04-04 | 2019-08-23 | 腾讯科技(深圳)有限公司 | Video pedestrian recognition methods, device and storage medium again |
CN110288512A (en) * | 2019-05-16 | 2019-09-27 | 成都品果科技有限公司 | Illumination for image synthesis remaps method, apparatus, storage medium and processor |
CN110516569A (en) * | 2019-08-15 | 2019-11-29 | 华侨大学 | A kind of pedestrian's attribute recognition approach of identity-based and non-identity attribute interactive learning |
CN110516512A (en) * | 2018-05-21 | 2019-11-29 | 北京中科奥森数据科技有限公司 | Training method, pedestrian's attribute recognition approach and the device of pedestrian's attributive analysis model |
CN110598654A (en) * | 2019-09-18 | 2019-12-20 | 合肥工业大学 | Multi-granularity cross modal feature fusion pedestrian re-identification method and re-identification system |
CN110728216A (en) * | 2019-09-27 | 2020-01-24 | 西北工业大学 | Unsupervised pedestrian re-identification method based on pedestrian attribute adaptive learning |
CN110765960A (en) * | 2019-10-29 | 2020-02-07 | 黄山学院 | Pedestrian re-identification method for adaptive multi-task deep learning |
CN110796079A (en) * | 2019-10-29 | 2020-02-14 | 深圳龙岗智能视听研究院 | Multi-camera visitor identification method and system based on face depth features and human body local depth features |
CN110909565A (en) * | 2018-09-14 | 2020-03-24 | 阿里巴巴集团控股有限公司 | Image recognition and pedestrian re-recognition method and apparatus, electronic and storage device |
CN111259910A (en) * | 2018-11-30 | 2020-06-09 | 阿里巴巴集团控股有限公司 | Object extraction method and device |
CN111259701A (en) * | 2018-12-03 | 2020-06-09 | 杭州海康威视数字技术股份有限公司 | Pedestrian re-identification method and device and electronic equipment |
CN111288999A (en) * | 2020-02-19 | 2020-06-16 | 深圳大学 | Pedestrian road network attribute detection method, device and equipment based on mobile terminal |
CN111339849A (en) * | 2020-02-14 | 2020-06-26 | 北京工业大学 | Pedestrian re-identification method integrating pedestrian attributes |
CN111832584A (en) * | 2019-04-16 | 2020-10-27 | 富士通株式会社 | Image processing apparatus, training apparatus and training method thereof |
CN112001353A (en) * | 2020-09-03 | 2020-11-27 | 杭州云栖智慧视通科技有限公司 | Pedestrian re-identification method based on multi-task joint supervised learning |
CN112199983A (en) * | 2020-07-08 | 2021-01-08 | 北京航空航天大学 | Multi-level screening long-time large-range pedestrian re-identification method |
CN112232241A (en) * | 2020-10-22 | 2021-01-15 | 华中科技大学 | Pedestrian re-identification method and device, electronic equipment and readable storage medium |
CN112613474A (en) * | 2020-12-30 | 2021-04-06 | 珠海大横琴科技发展有限公司 | Pedestrian re-identification method and device |
CN112712066A (en) * | 2021-01-19 | 2021-04-27 | 腾讯科技(深圳)有限公司 | Image recognition method and device, computer equipment and storage medium |
CN113762005A (en) * | 2020-11-09 | 2021-12-07 | 北京沃东天骏信息技术有限公司 | Method, device, equipment and medium for training feature selection model and classifying objects |
CN114239754A (en) * | 2022-02-24 | 2022-03-25 | 中国科学院自动化研究所 | Pedestrian attribute identification method and system based on attribute feature learning decoupling |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104915643A (en) * | 2015-05-26 | 2015-09-16 | 中山大学 | Deep-learning-based pedestrian re-identification method |
-
2017
- 2017-06-28 CN CN201710504662.6A patent/CN107330396B/en active Active
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104915643A (en) * | 2015-05-26 | 2015-09-16 | 中山大学 | Deep-learning-based pedestrian re-identification method |
Non-Patent Citations (6)
Title |
---|
CHI SU等: "Deep Attributes Driven Multi-Camera Person Re-identification", 《ARXIV:1605.03259V2》 * |
KAREN SIMONYAN等: "VERY DEEP CONVOLUTIONAL NETWORKS FOR LARGE-SCALE IMAGE RECOGNITION", 《ARXIV:1409.1556V6》 * |
PATRICK SUDOWE等: "Person Attribute Recognition with a Jointly-trained Holistic CNN Model", 《2015 IEEE INTERNATIONAL CONFERENCE ON COMPUTER VISION WORKSHOPS》 * |
YUTIAN LIN等: "Improving Person Re-identification by Attribute and Identity Learning", 《ARXIV:1703.07220V2》 * |
ZHEDONG ZHENG等: "A Discriminatively Learned CNN Embedding for Person Re-identification", 《ARXIV:1611.05666V2》 * |
刘成: "行人再识别关键技术研究", 《中国优秀硕士学位论文全文数据库-信息科技辑》 * |
Cited By (89)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108875765A (en) * | 2017-11-14 | 2018-11-23 | 北京旷视科技有限公司 | Method, apparatus, equipment and the computer storage medium of EDS extended data set |
CN107944366A (en) * | 2017-11-16 | 2018-04-20 | 山东财经大学 | A kind of finger vein identification method and device based on attribute study |
CN107944366B (en) * | 2017-11-16 | 2020-04-17 | 山东财经大学 | Finger vein identification method and device based on attribute learning |
CN108108674A (en) * | 2017-12-08 | 2018-06-01 | 浙江捷尚视觉科技股份有限公司 | A kind of recognition methods again of the pedestrian based on joint point analysis |
CN109919166A (en) * | 2017-12-12 | 2019-06-21 | 杭州海康威视数字技术股份有限公司 | The method and apparatus for obtaining the classification information of attribute |
CN109919166B (en) * | 2017-12-12 | 2021-04-09 | 杭州海康威视数字技术股份有限公司 | Method and device for acquiring classification information of attributes |
CN108108754A (en) * | 2017-12-15 | 2018-06-01 | 北京迈格威科技有限公司 | The training of identification network, again recognition methods, device and system again |
CN108229543A (en) * | 2017-12-22 | 2018-06-29 | 中国科学院深圳先进技术研究院 | Image classification design methods and device |
CN108229398A (en) * | 2018-01-04 | 2018-06-29 | 中科汇通投资控股有限公司 | A kind of face verification method of self-teaching |
CN108229503A (en) * | 2018-01-04 | 2018-06-29 | 浙江大学 | A kind of feature extracting method for clothes photo |
CN108416295A (en) * | 2018-03-08 | 2018-08-17 | 天津师范大学 | A kind of recognition methods again of the pedestrian based on locally embedding depth characteristic |
CN108416295B (en) * | 2018-03-08 | 2021-10-15 | 天津师范大学 | Pedestrian re-identification method based on local embedding depth features |
CN108537136A (en) * | 2018-03-19 | 2018-09-14 | 复旦大学 | The pedestrian's recognition methods again generated based on posture normalized image |
CN110163041A (en) * | 2018-04-04 | 2019-08-23 | 腾讯科技(深圳)有限公司 | Video pedestrian recognition methods, device and storage medium again |
CN108647577A (en) * | 2018-04-10 | 2018-10-12 | 华中科技大学 | A kind of pedestrian's weight identification model that adaptive difficult example is excavated, method and system |
CN108647577B (en) * | 2018-04-10 | 2021-04-20 | 华中科技大学 | Self-adaptive pedestrian re-identification method and system for difficult excavation |
CN108595558A (en) * | 2018-04-12 | 2018-09-28 | 福建工程学院 | A kind of image labeling method of data balancing strategy and multiple features fusion |
CN108595558B (en) * | 2018-04-12 | 2022-03-15 | 福建工程学院 | Image annotation method based on data equalization strategy and multi-feature fusion |
CN108764065A (en) * | 2018-05-04 | 2018-11-06 | 华中科技大学 | A kind of method of pedestrian's weight identification feature fusion assisted learning |
CN108764065B (en) * | 2018-05-04 | 2020-12-08 | 华中科技大学 | Pedestrian re-recognition feature fusion aided learning method |
CN108846413B (en) * | 2018-05-21 | 2021-07-23 | 复旦大学 | Zero sample learning method based on global semantic consensus network |
CN110516512B (en) * | 2018-05-21 | 2023-08-25 | 北京中科奥森数据科技有限公司 | Training method of pedestrian attribute analysis model, pedestrian attribute identification method and device |
CN110516512A (en) * | 2018-05-21 | 2019-11-29 | 北京中科奥森数据科技有限公司 | Training method, pedestrian's attribute recognition approach and the device of pedestrian's attributive analysis model |
CN108846413A (en) * | 2018-05-21 | 2018-11-20 | 复旦大学 | A kind of zero sample learning method based on global semantic congruence network |
US11275932B2 (en) | 2018-05-30 | 2022-03-15 | Tencent Technology (Shenzhen) Company Limited | Human body attribute recognition method, apparatus, and device and medium |
CN108921022A (en) * | 2018-05-30 | 2018-11-30 | 腾讯科技(深圳)有限公司 | A kind of human body attribute recognition approach, device, equipment and medium |
WO2019228089A1 (en) * | 2018-05-30 | 2019-12-05 | 腾讯科技(深圳)有限公司 | Human body attribute recognition method, apparatus, and device, and medium |
CN109101866A (en) * | 2018-06-05 | 2018-12-28 | 中国科学院自动化研究所 | Pedestrian recognition methods and system again based on segmentation outline |
CN109101866B (en) * | 2018-06-05 | 2020-12-15 | 中国科学院自动化研究所 | Pedestrian re-identification method and system based on segmentation silhouette |
CN109033154A (en) * | 2018-06-12 | 2018-12-18 | 佛山欧神诺陶瓷有限公司 | A kind of category management method |
CN108921054B (en) * | 2018-06-15 | 2021-08-03 | 华中科技大学 | Pedestrian multi-attribute identification method based on semantic segmentation |
CN108921054A (en) * | 2018-06-15 | 2018-11-30 | 华中科技大学 | A kind of more attribute recognition approaches of pedestrian based on semantic segmentation |
CN109190646A (en) * | 2018-06-25 | 2019-01-11 | 北京达佳互联信息技术有限公司 | A kind of data predication method neural network based, device and nerve network system |
CN109190646B (en) * | 2018-06-25 | 2019-08-20 | 北京达佳互联信息技术有限公司 | A kind of data predication method neural network based, device and nerve network system |
CN108960331A (en) * | 2018-07-10 | 2018-12-07 | 重庆邮电大学 | A kind of recognition methods again of the pedestrian based on pedestrian image feature clustering |
CN108960184B (en) * | 2018-07-20 | 2021-08-24 | 天津师范大学 | Pedestrian re-identification method based on heterogeneous component deep neural network |
CN108960184A (en) * | 2018-07-20 | 2018-12-07 | 天津师范大学 | A kind of recognition methods again of the pedestrian based on heterogeneous components deep neural network |
CN109165306A (en) * | 2018-08-09 | 2019-01-08 | 长沙理工大学 | Image search method based on the study of multitask Hash |
CN109165306B (en) * | 2018-08-09 | 2021-11-23 | 长沙理工大学 | Image retrieval method based on multitask Hash learning |
CN109214308A (en) * | 2018-08-15 | 2019-01-15 | 武汉唯理科技有限公司 | A kind of traffic abnormity image identification method based on focal loss function |
CN109271898A (en) * | 2018-08-31 | 2019-01-25 | 电子科技大学 | Solution cavity body recognizer based on optimization convolutional neural networks |
CN109190120A (en) * | 2018-08-31 | 2019-01-11 | 第四范式(北京)技术有限公司 | Neural network training method and device and name entity recognition method and device |
CN109190120B (en) * | 2018-08-31 | 2020-01-21 | 第四范式(北京)技术有限公司 | Neural network training method and device and named entity identification method and device |
CN110909565A (en) * | 2018-09-14 | 2020-03-24 | 阿里巴巴集团控股有限公司 | Image recognition and pedestrian re-recognition method and apparatus, electronic and storage device |
CN110909565B (en) * | 2018-09-14 | 2023-06-16 | 阿里巴巴集团控股有限公司 | Image recognition and pedestrian re-recognition method and device, electronic and storage equipment |
CN109325547A (en) * | 2018-10-23 | 2019-02-12 | 苏州科达科技股份有限公司 | Non-motor vehicle image multi-tag classification method, system, equipment and storage medium |
CN109635634A (en) * | 2018-10-29 | 2019-04-16 | 西北大学 | A kind of pedestrian based on stochastic linear interpolation identifies data enhancement methods again |
CN109635634B (en) * | 2018-10-29 | 2023-03-31 | 西北大学 | Pedestrian re-identification data enhancement method based on random linear interpolation |
CN109522855A (en) * | 2018-11-23 | 2019-03-26 | 广州广电银通金融电子科技有限公司 | In conjunction with low resolution pedestrian detection method, system and the storage medium of ResNet and SENet |
CN111259910A (en) * | 2018-11-30 | 2020-06-09 | 阿里巴巴集团控股有限公司 | Object extraction method and device |
CN111259701A (en) * | 2018-12-03 | 2020-06-09 | 杭州海康威视数字技术股份有限公司 | Pedestrian re-identification method and device and electronic equipment |
CN111259701B (en) * | 2018-12-03 | 2023-04-25 | 杭州海康威视数字技术股份有限公司 | Pedestrian re-identification method and device and electronic equipment |
CN109711281A (en) * | 2018-12-10 | 2019-05-03 | 复旦大学 | A kind of pedestrian based on deep learning identifies again identifies fusion method with feature |
CN109711281B (en) * | 2018-12-10 | 2023-05-02 | 复旦大学 | Pedestrian re-recognition and feature recognition fusion method based on deep learning |
CN109766790A (en) * | 2018-12-24 | 2019-05-17 | 重庆邮电大学 | A kind of pedestrian detection method based on self-adaptive features channel |
CN109766790B (en) * | 2018-12-24 | 2022-08-23 | 重庆邮电大学 | Pedestrian detection method based on self-adaptive characteristic channel |
CN109754005B (en) * | 2018-12-25 | 2022-05-10 | 任飞翔 | Dynamic increasing and decreasing method and device |
CN109754005A (en) * | 2018-12-25 | 2019-05-14 | 任飞翔 | Dynamic increase and decrease method and device |
CN109740480A (en) * | 2018-12-26 | 2019-05-10 | 浙江捷尚视觉科技股份有限公司 | A kind of identified again based on non-motor vehicle pedestrian retrieval method under monitoring scene |
CN109711354A (en) * | 2018-12-28 | 2019-05-03 | 哈尔滨工业大学(威海) | A kind of method for tracking target indicating study based on video attribute |
CN109784288A (en) * | 2019-01-22 | 2019-05-21 | 天津师范大学 | A kind of pedestrian's recognition methods again based on differentiation perception fusion |
CN109815908A (en) * | 2019-01-25 | 2019-05-28 | 同济大学 | It is a kind of based on the discrimination method again of the pedestrian that measures between deep learning and overlapping image block |
CN109919073A (en) * | 2019-03-01 | 2019-06-21 | 中山大学 | A kind of recognition methods again of the pedestrian with illumination robustness |
CN110046553A (en) * | 2019-03-21 | 2019-07-23 | 华中科技大学 | A kind of pedestrian weight identification model, method and system merging attributive character |
CN110096346A (en) * | 2019-03-29 | 2019-08-06 | 广州思德医疗科技有限公司 | A kind of training mission processing method and processing device of more calculate nodes |
CN110008913A (en) * | 2019-04-08 | 2019-07-12 | 南京工业大学 | Pedestrian re-identification method based on fusion of attitude estimation and viewpoint mechanism |
CN111832584A (en) * | 2019-04-16 | 2020-10-27 | 富士通株式会社 | Image processing apparatus, training apparatus and training method thereof |
CN110288512B (en) * | 2019-05-16 | 2023-04-18 | 成都品果科技有限公司 | Illumination remapping method, device, storage medium and processor for image synthesis |
CN110288512A (en) * | 2019-05-16 | 2019-09-27 | 成都品果科技有限公司 | Illumination for image synthesis remaps method, apparatus, storage medium and processor |
CN110516569A (en) * | 2019-08-15 | 2019-11-29 | 华侨大学 | A kind of pedestrian's attribute recognition approach of identity-based and non-identity attribute interactive learning |
CN110598654B (en) * | 2019-09-18 | 2022-02-11 | 合肥工业大学 | Multi-granularity cross modal feature fusion pedestrian re-identification method and re-identification system |
CN110598654A (en) * | 2019-09-18 | 2019-12-20 | 合肥工业大学 | Multi-granularity cross modal feature fusion pedestrian re-identification method and re-identification system |
CN110728216A (en) * | 2019-09-27 | 2020-01-24 | 西北工业大学 | Unsupervised pedestrian re-identification method based on pedestrian attribute adaptive learning |
CN110765960A (en) * | 2019-10-29 | 2020-02-07 | 黄山学院 | Pedestrian re-identification method for adaptive multi-task deep learning |
CN110796079A (en) * | 2019-10-29 | 2020-02-14 | 深圳龙岗智能视听研究院 | Multi-camera visitor identification method and system based on face depth features and human body local depth features |
CN110765960B (en) * | 2019-10-29 | 2022-03-04 | 黄山学院 | Pedestrian re-identification method for adaptive multi-task deep learning |
CN111339849A (en) * | 2020-02-14 | 2020-06-26 | 北京工业大学 | Pedestrian re-identification method integrating pedestrian attributes |
CN111288999A (en) * | 2020-02-19 | 2020-06-16 | 深圳大学 | Pedestrian road network attribute detection method, device and equipment based on mobile terminal |
CN111288999B (en) * | 2020-02-19 | 2021-08-31 | 深圳大学 | Pedestrian road network attribute detection method, device and equipment based on mobile terminal |
CN112199983A (en) * | 2020-07-08 | 2021-01-08 | 北京航空航天大学 | Multi-level screening long-time large-range pedestrian re-identification method |
CN112001353B (en) * | 2020-09-03 | 2023-02-17 | 杭州云栖智慧视通科技有限公司 | Pedestrian re-identification method based on multi-task joint supervised learning |
CN112001353A (en) * | 2020-09-03 | 2020-11-27 | 杭州云栖智慧视通科技有限公司 | Pedestrian re-identification method based on multi-task joint supervised learning |
CN112232241A (en) * | 2020-10-22 | 2021-01-15 | 华中科技大学 | Pedestrian re-identification method and device, electronic equipment and readable storage medium |
CN113762005A (en) * | 2020-11-09 | 2021-12-07 | 北京沃东天骏信息技术有限公司 | Method, device, equipment and medium for training feature selection model and classifying objects |
CN112613474B (en) * | 2020-12-30 | 2022-01-18 | 珠海大横琴科技发展有限公司 | Pedestrian re-identification method and device |
CN112613474A (en) * | 2020-12-30 | 2021-04-06 | 珠海大横琴科技发展有限公司 | Pedestrian re-identification method and device |
CN112712066A (en) * | 2021-01-19 | 2021-04-27 | 腾讯科技(深圳)有限公司 | Image recognition method and device, computer equipment and storage medium |
CN114239754A (en) * | 2022-02-24 | 2022-03-25 | 中国科学院自动化研究所 | Pedestrian attribute identification method and system based on attribute feature learning decoupling |
CN114239754B (en) * | 2022-02-24 | 2022-05-03 | 中国科学院自动化研究所 | Pedestrian attribute identification method and system based on attribute feature learning decoupling |
Also Published As
Publication number | Publication date |
---|---|
CN107330396B (en) | 2020-05-19 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107330396A (en) | A kind of pedestrian's recognition methods again based on many attributes and many strategy fusion study | |
CN108108657B (en) | Method for correcting locality sensitive Hash vehicle retrieval based on multitask deep learning | |
CN111666843B (en) | Pedestrian re-recognition method based on global feature and local feature splicing | |
CN105512680B (en) | A kind of more view SAR image target recognition methods based on deep neural network | |
CN106326886B (en) | Finger vein image quality appraisal procedure based on convolutional neural networks | |
CN109948425A (en) | A kind of perception of structure is from paying attention to and online example polymerize matched pedestrian's searching method and device | |
CN110084151B (en) | Video abnormal behavior discrimination method based on non-local network deep learning | |
CN108921054B (en) | Pedestrian multi-attribute identification method based on semantic segmentation | |
CN108710868A (en) | A kind of human body critical point detection system and method based under complex scene | |
CN109711281A (en) | A kind of pedestrian based on deep learning identifies again identifies fusion method with feature | |
CN110427813A (en) | Pedestrian's recognition methods again based on the twin production confrontation network that posture instructs pedestrian image to generate | |
CN107133569A (en) | The many granularity mask methods of monitor video based on extensive Multi-label learning | |
CN110598543B (en) | Model training method based on attribute mining and reasoning and pedestrian re-identification method | |
CN112131967A (en) | Remote sensing scene classification method based on multi-classifier anti-transfer learning | |
CN106529499A (en) | Fourier descriptor and gait energy image fusion feature-based gait identification method | |
CN104992142A (en) | Pedestrian recognition method based on combination of depth learning and property learning | |
CN106845373A (en) | Towards pedestrian's attribute forecast method of monitor video | |
CN108960184A (en) | A kind of recognition methods again of the pedestrian based on heterogeneous components deep neural network | |
CN104915643A (en) | Deep-learning-based pedestrian re-identification method | |
CN103778441B (en) | A kind of sequence Aircraft Target Recognition based on DSmT and HMM | |
He et al. | Exemplar-driven top-down saliency detection via deep association | |
CN109034035A (en) | Pedestrian's recognition methods again based on conspicuousness detection and Fusion Features | |
CN111950372A (en) | Unsupervised pedestrian re-identification method based on graph convolution network | |
CN113221625A (en) | Method for re-identifying pedestrians by utilizing local features of deep learning | |
CN110097029A (en) | Identity identifying method based on Highway network multi-angle of view Gait Recognition |
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 |